Named Entity Recognition with Conditional Random Fields
H
{
"cells": [
{
"cell_type": "markdown",
"metadata": {},
"source": [
"# Named Entity Recognition with Conditional Random Fields"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"One of the classic challenges of Natural Language Processing is sequence labelling. In sequence labelling, the goal is to label each word in a text with a word class. In part-of-speech tagging, these word classes are parts of speech, such as noun or verb. In named entity recognition (NER), they're types of generic named entities, such as locations, people or organizations, or more specialized entities, such as diseases or symptoms in the healthcare domain. In this way, sequence labelling can help us extract the most important information from a text and improve the performance of analytics, search or matching applications. "
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"In this notebook we'll explore Conditional Random Fields, the most popular approach to sequence labelling before Deep Learning arrived. Deep Learning may get all the attention right now, but Conditional Random Fields are still a powerful tool to build a simple sequence labeller. \n",
"\n",
"The tool we're going to use is `sklearn-crfsuite`. This is a wrapper around `python-crfsuite`, which itself is a Python binding of [CRFSuite](http://www.chokkan.org/software/crfsuite/). The reason we're using `sklearn-crfsuite` is that it provides a number of handy utility functions, for example for evaluating the output of the model. You can install it with `pip install sklearn-crfsuite`."
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"## Data"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"First we get some data. A well-known data set for training and testing NER models is the CoNLL-2002 data, which has Spanish and Dutch texts labelled with four types of entities: locations (LOC), persons (PER), organizations (ORG) and miscellaneous entities (MISC). Both corpora are split up in three portions: a training portion and two smaller test portions, one of which we'll use as development data. It's easy to collect the data from NLTK. "
]
},
{
"cell_type": "code",
"execution_count": 1,
"metadata": {},
"outputs": [],
"source": [
"import nltk\n",
"import sklearn\n",
"from sklearn.metrics import classification_report, confusion_matrix\n",
"from sklearn.preprocessing import LabelBinarizer\n",
"import sklearn_crfsuite as crfsuite\n",
"from sklearn_crfsuite import metrics"
]
},
{
"cell_type": "code",
"execution_count": 2,
"metadata": {},
"outputs": [],
"source": [
"train_sents = list(nltk.corpus.conll2002.iob_sents('ned.train'))\n",
"dev_sents = list(nltk.corpus.conll2002.iob_sents('ned.testa'))\n",
"test_sents = list(nltk.corpus.conll2002.iob_sents('ned.testb'))"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"The data consists of a list of tokenized sentences. For each of the tokens we have the string itself, its part-of-speech tag and its entity tag, which follows the BIO convention. In the deep learning world we live in today, it's common to ignore the part-of-speech tags. However, since CRFs rely on good feature extraction, we'll gladly make use of this information. After all, the part of speech of a word tells us a lot about its possible status as a named entity: nouns will more often be entities than verbs, for example."
]
},
{
"cell_type": "code",
"execution_count": 3,
"metadata": {},
"outputs": [
{
"data": {
"text/plain": [
"[('De', 'Art', 'O'),\n",
" ('tekst', 'N', 'O'),\n",
" ('van', 'Prep', 'O'),\n",
" ('het', 'Art', 'O'),\n",
" ('arrest', 'N', 'O'),\n",
" ('is', 'V', 'O'),\n",
" ('nog', 'Adv', 'O'),\n",
" ('niet', 'Adv', 'O'),\n",
" ('schriftelijk', 'Adj', 'O'),\n",
" ('beschikbaar', 'Adj', 'O'),\n",
" ('maar', 'Conj', 'O'),\n",
" ('het', 'Art', 'O'),\n",
" ('bericht', 'N', 'O'),\n",
" ('werd', 'V', 'O'),\n",
" ('alvast', 'Adv', 'O'),\n",
" ('bekendgemaakt', 'V', 'O'),\n",
" ('door', 'Prep', 'O'),\n",
" ('een', 'Art', 'O'),\n",
" ('communicatiebureau', 'N', 'O'),\n",
" ('dat', 'Conj', 'O'),\n",
" ('Floralux', 'N', 'B-ORG'),\n",
" ('inhuurde', 'V', 'O'),\n",
" ('.', 'Punc', 'O')]"
]
},
"execution_count": 3,
"metadata": {},
"output_type": "execute_result"
}
],
"source": [
"train_sents[0]"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"## Feature Extraction"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"Whereas today neural networks are expected to learn the relevant features of the input texts themselves, this is very different with Conditional Random Fields. CRFs learn the relationship between the features we give them and the label of a token in a given context. They're not going to earn these features themselves. Instead, the quality of the model will depend highly on the relevance of the features we show it. \n",
"\n",
"The most important method in this tutorial is therefore the one that collects the features for every token. What information could be useful? The word itself, of course, together with its part of speech tag. It can also be interesting to know whether the word is completely uppercase, whether it starts with a capital or is a digit. In addition, we also take a look at the character bigram and trigram the word ends with. We also give every token a `bias` feature, which always has the same value. This bias feature helps the CRF learn the relative frequency of each label type in the training data.\n",
"\n",
"To give the CRF more information about the meaning of a word, we also introduce information from word embeddings. In our [Word Embedding notebook](https://github.com/nlptown/nlp-notebooks/blob/master/An%20Introduction%20to%20Word%20Embeddings.ipynb), we trained word embeddings on Dutch Wikipedia and clustered them in 500 clusters. Here we'll read these 500 clusters from a file, and map each word to the id of the cluster it is in. This is really useful for Named Entity Recognition, as most entity types cluster together. This allows CRFs to generalize above the word level. For example, when the CRF encounters a word it has never seen (say, *Albania*), it can base its decision on the cluster the word is in. If this cluster contains many other entities the CRF has met in its training data (say, *Italy*, *Germany* and *France*), it will have learnt a string link between this cluster and a specific entity type. As a result, it can still assign that entity type to the unknown word. In our experiments, this feature alone boosts the performance with around 3%. \n",
"\n",
"Finally, apart from the token itself, we also want the CRF to look at its context. More specifically, we're going to give it some extra information about the two words to the left and the right of the targt word. We'll tell the CRF what these words are, whether they start with a capital or are completely uppercase, and give it their part-of-speech tag. If there is no left or right context, we'll inform the CRF that the token is at the beginning or end of the sentence (`BOS` or `EOS`). "
]
},
{
"cell_type": "code",
"execution_count": 4,
"metadata": {},
"outputs": [],
"source": [
"def read_clusters(cluster_file):\n",
" word2cluster = {}\n",
" with open(cluster_file) as i:\n",
" for line in i:\n",
" word, cluster = line.strip().split('\\t')\n",
" word2cluster[word] = cluster\n",
" return word2cluster\n",
"\n",
"\n",
"def word2features(sent, i, word2cluster):\n",
" word = sent[i][0]\n",
" postag = sent[i][1]\n",
" features = [\n",
" 'bias',\n",
" 'word.lower=' + word.lower(),\n",
" 'word[-3:]=' + word[-3:],\n",
" 'word[-2:]=' + word[-2:],\n",
" 'word.isupper=%s' % word.isupper(),\n",
" 'word.istitle=%s' % word.istitle(),\n",
" 'word.isdigit=%s' % word.isdigit(),\n",
" 'word.cluster=%s' % word2cluster[word.lower()] if word.lower() in word2cluster else \"0\",\n",
" 'postag=' + postag\n",
" ]\n",
" if i > 0:\n",
" word1 = sent[i-1][0]\n",
" postag1 = sent[i-1][1]\n",
" features.extend([\n",
" '-1:word.lower=' + word1.lower(),\n",
" '-1:word.istitle=%s' % word1.istitle(),\n",
" '-1:word.isupper=%s' % word1.isupper(),\n",
" '-1:postag=' + postag1\n",
" ])\n",
" else:\n",
" features.append('BOS')\n",
"\n",
" if i > 1: \n",
" word2 = sent[i-2][0]\n",
" postag2 = sent[i-2][1]\n",
" features.extend([\n",
" '-2:word.lower=' + word2.lower(),\n",
" '-2:word.istitle=%s' % word2.istitle(),\n",
" '-2:word.isupper=%s' % word2.isupper(),\n",
" '-2:postag=' + postag2\n",
" ]) \n",
"\n",
" \n",
" if i < len(sent)-1:\n",
" word1 = sent[i+1][0]\n",
" postag1 = sent[i+1][1]\n",
" features.extend([\n",
" '+1:word.lower=' + word1.lower(),\n",
" '+1:word.istitle=%s' % word1.istitle(),\n",
" '+1:word.isupper=%s' % word1.isupper(),\n",
" '+1:postag=' + postag1\n",
" ])\n",
" else:\n",
" features.append('EOS')\n",
"\n",
" if i < len(sent)-2:\n",
" word2 = sent[i+2][0]\n",
" postag2 = sent[i+2][1]\n",
" features.extend([\n",
" '+2:word.lower=' + word2.lower(),\n",
" '+2:word.istitle=%s' % word2.istitle(),\n",
" '+2:word.isupper=%s' % word2.isupper(),\n",
" '+2:postag=' + postag2\n",
" ])\n",
"\n",
" \n",
" return features\n",
"\n",
"\n",
"def sent2features(sent, word2cluster):\n",
" return [word2features(sent, i, word2cluster) for i in range(len(sent))]\n",
"\n",
"def sent2labels(sent):\n",
" return [label for token, postag, label in sent]\n",
"\n",
"def sent2tokens(sent):\n",
" return [token for token, postag, label in sent]\n",
"\n",
"word2cluster = read_clusters(\"data/embeddings/clusters_nl.tsv\")"
]
},
{
"cell_type": "code",
"execution_count": 5,
"metadata": {},
"outputs": [
{
"data": {
"text/plain": [
"['bias',\n",
" 'word.lower=de',\n",
" 'word[-3:]=De',\n",
" 'word[-2:]=De',\n",
" 'word.isupper=False',\n",
" 'word.istitle=True',\n",
" 'word.isdigit=False',\n",
" 'word.cluster=38',\n",
" 'postag=Art',\n",
" 'BOS',\n",
" '+1:word.lower=tekst',\n",
" '+1:word.istitle=False',\n",
" '+1:word.isupper=False',\n",
" '+1:postag=N',\n",
" '+2:word.lower=van',\n",
" '+2:word.istitle=False',\n",
" '+2:word.isupper=False',\n",
" '+2:postag=Prep']"
]
},
"execution_count": 5,
"metadata": {},
"output_type": "execute_result"
}
],
"source": [
"sent2features(train_sents[0], word2cluster)[0]"
]
},
{
"cell_type": "code",
"execution_count": 6,
"metadata": {},
"outputs": [],
"source": [
"X_train = [sent2features(s, word2cluster) for s in train_sents]\n",
"y_train = [sent2labels(s) for s in train_sents]\n",
"\n",
"X_dev = [sent2features(s, word2cluster) for s in dev_sents]\n",
"y_dev = [sent2labels(s) for s in dev_sents]\n",
"\n",
"X_test = [sent2features(s, word2cluster) for s in test_sents]\n",
"y_test = [sent2labels(s) for s in test_sents]"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"## Training"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"We now create a CRF model and train it. We'll use the standard [L-BFGS](https://en.wikipedia.org/wiki/Limited-memory_BFGS) algorithm for our parameter estimation and run it for 100 iterations. When we're done, we save the model with `joblib`."
]
},
{
"cell_type": "code",
"execution_count": 7,
"metadata": {},
"outputs": [
{
"name": "stderr",
"output_type": "stream",
"text": [
"loading training data to CRFsuite: 100%|██████████| 15806/15806 [00:02<00:00, 7623.17it/s]\n",
"loading dev data to CRFsuite: 27%|██▋ | 769/2895 [00:00<00:00, 7689.13it/s]"
]
},
{
"name": "stdout",
"output_type": "stream",
"text": [
"\n"
]
},
{
"name": "stderr",
"output_type": "stream",
"text": [
"loading dev data to CRFsuite: 100%|██████████| 2895/2895 [00:00<00:00, 7186.08it/s]\n"
]
},
{
"name": "stdout",
"output_type": "stream",
"text": [
"\n",
"Holdout group: 2\n",
"\n",
"Feature generation\n",
"type: CRF1d\n",
"feature.minfreq: 0.000000\n",
"feature.possible_states: 0\n",
"feature.possible_transitions: 0\n",
"0....1....2....3....4....5....6....7....8....9....10\n",
"Number of features: 152117\n",
"Seconds required: 0.424\n",
"\n",
"L-BFGS optimization\n",
"c1: 0.000000\n",
"c2: 1.000000\n",
"num_memories: 6\n",
"max_iterations: 100\n",
"epsilon: 0.000010\n",
"stop: 10\n",
"delta: 0.000010\n",
"linesearch: MoreThuente\n",
"linesearch.max_iterations: 20\n",
"\n",
"Iter 1 time=0.37 loss=104214.83 active=152117 precision=0.100 recall=0.111 F1=0.105 Acc(item/seq)=0.901 0.496 feature_norm=1.00\n",
"Iter 2 time=0.21 loss=96997.81 active=152117 precision=0.100 recall=0.111 F1=0.105 Acc(item/seq)=0.901 0.496 feature_norm=1.13\n",
"Iter 3 time=0.21 loss=92085.38 active=152117 precision=0.100 recall=0.111 F1=0.105 Acc(item/seq)=0.901 0.496 feature_norm=1.26\n",
"Iter 4 time=0.21 loss=84277.67 active=152117 precision=0.100 recall=0.111 F1=0.105 Acc(item/seq)=0.901 0.496 feature_norm=1.51\n",
"Iter 5 time=0.21 loss=67577.53 active=152117 precision=0.169 recall=0.113 F1=0.109 Acc(item/seq)=0.902 0.496 feature_norm=2.32\n",
"Iter 6 time=0.21 loss=47854.26 active=152117 precision=0.326 recall=0.347 F1=0.320 Acc(item/seq)=0.930 0.580 feature_norm=4.34\n",
"Iter 7 time=0.21 loss=43326.19 active=152117 precision=0.340 recall=0.365 F1=0.333 Acc(item/seq)=0.933 0.592 feature_norm=5.06\n",
"Iter 8 time=0.21 loss=38617.07 active=152117 precision=0.372 recall=0.399 F1=0.362 Acc(item/seq)=0.938 0.618 feature_norm=6.43\n",
"Iter 9 time=0.21 loss=35511.85 active=152117 precision=0.491 recall=0.442 F1=0.421 Acc(item/seq)=0.942 0.631 feature_norm=8.62\n",
"Iter 10 time=0.21 loss=32735.31 active=152117 precision=0.535 recall=0.456 F1=0.445 Acc(item/seq)=0.944 0.654 feature_norm=9.50\n",
"Iter 11 time=0.21 loss=31687.17 active=152117 precision=0.530 recall=0.491 F1=0.502 Acc(item/seq)=0.947 0.663 feature_norm=10.94\n",
"Iter 12 time=0.21 loss=29323.19 active=152117 precision=0.532 recall=0.482 F1=0.477 Acc(item/seq)=0.946 0.666 feature_norm=11.18\n",
"Iter 13 time=0.21 loss=28733.55 active=152117 precision=0.596 recall=0.489 F1=0.489 Acc(item/seq)=0.947 0.668 feature_norm=11.58\n",
"Iter 14 time=0.21 loss=27120.69 active=152117 precision=0.640 recall=0.507 F1=0.512 Acc(item/seq)=0.948 0.677 feature_norm=12.36\n",
"Iter 15 time=0.21 loss=24849.05 active=152117 precision=0.640 recall=0.547 F1=0.558 Acc(item/seq)=0.952 0.697 feature_norm=13.86\n",
"Iter 16 time=0.40 loss=24033.40 active=152117 precision=0.654 recall=0.580 F1=0.586 Acc(item/seq)=0.954 0.706 feature_norm=14.53\n",
"Iter 17 time=0.21 loss=22935.94 active=152117 precision=0.669 recall=0.578 F1=0.598 Acc(item/seq)=0.955 0.712 feature_norm=15.15\n",
"Iter 18 time=0.21 loss=21803.53 active=152117 precision=0.682 recall=0.584 F1=0.605 Acc(item/seq)=0.956 0.713 feature_norm=15.67\n",
"Iter 19 time=0.21 loss=21046.75 active=152117 precision=0.725 recall=0.565 F1=0.586 Acc(item/seq)=0.956 0.717 feature_norm=16.04\n",
"Iter 20 time=0.21 loss=20465.96 active=152117 precision=0.701 recall=0.566 F1=0.594 Acc(item/seq)=0.956 0.711 feature_norm=15.81\n",
"Iter 21 time=0.21 loss=19991.29 active=152117 precision=0.706 recall=0.586 F1=0.611 Acc(item/seq)=0.957 0.717 feature_norm=15.63\n",
"Iter 22 time=0.21 loss=19560.23 active=152117 precision=0.684 recall=0.597 F1=0.616 Acc(item/seq)=0.957 0.722 feature_norm=15.58\n",
"Iter 23 time=0.21 loss=19241.14 active=152117 precision=0.672 recall=0.602 F1=0.615 Acc(item/seq)=0.957 0.726 feature_norm=15.65\n",
"Iter 24 time=0.21 loss=18787.87 active=152117 precision=0.678 recall=0.627 F1=0.637 Acc(item/seq)=0.958 0.731 feature_norm=16.31\n",
"Iter 25 time=0.21 loss=18145.13 active=152117 precision=0.690 recall=0.625 F1=0.640 Acc(item/seq)=0.959 0.734 feature_norm=16.94\n",
"Iter 26 time=0.21 loss=17786.38 active=152117 precision=0.710 recall=0.621 F1=0.642 Acc(item/seq)=0.959 0.738 feature_norm=17.48\n",
"Iter 27 time=0.21 loss=17247.02 active=152117 precision=0.711 recall=0.625 F1=0.649 Acc(item/seq)=0.959 0.736 feature_norm=18.46\n",
"Iter 28 time=0.21 loss=16876.01 active=152117 precision=0.737 recall=0.627 F1=0.655 Acc(item/seq)=0.960 0.748 feature_norm=19.83\n",
"Iter 29 time=0.21 loss=16543.87 active=152117 precision=0.732 recall=0.631 F1=0.663 Acc(item/seq)=0.961 0.751 feature_norm=20.12\n",
"Iter 30 time=0.21 loss=16263.21 active=152117 precision=0.725 recall=0.644 F1=0.671 Acc(item/seq)=0.962 0.753 feature_norm=20.42\n",
"Iter 31 time=0.21 loss=15665.78 active=152117 precision=0.715 recall=0.661 F1=0.676 Acc(item/seq)=0.963 0.758 feature_norm=21.50\n",
"Iter 32 time=0.21 loss=15247.34 active=152117 precision=0.700 recall=0.641 F1=0.650 Acc(item/seq)=0.961 0.748 feature_norm=23.18\n",
"Iter 33 time=0.21 loss=14866.51 active=152117 precision=0.702 recall=0.658 F1=0.670 Acc(item/seq)=0.963 0.754 feature_norm=24.67\n",
"Iter 34 time=0.21 loss=14650.96 active=152117 precision=0.704 recall=0.659 F1=0.675 Acc(item/seq)=0.963 0.755 feature_norm=25.38\n",
"Iter 35 time=0.21 loss=14386.98 active=152117 precision=0.730 recall=0.677 F1=0.695 Acc(item/seq)=0.964 0.761 feature_norm=26.47\n",
"Iter 36 time=0.21 loss=14158.49 active=152117 precision=0.742 recall=0.681 F1=0.704 Acc(item/seq)=0.965 0.763 feature_norm=28.58\n",
"Iter 37 time=0.21 loss=13895.30 active=152117 precision=0.736 recall=0.684 F1=0.701 Acc(item/seq)=0.965 0.765 feature_norm=28.72\n",
"Iter 38 time=0.21 loss=13656.45 active=152117 precision=0.730 recall=0.683 F1=0.695 Acc(item/seq)=0.965 0.768 feature_norm=29.07\n",
"Iter 39 time=0.21 loss=13499.28 active=152117 precision=0.727 recall=0.680 F1=0.691 Acc(item/seq)=0.965 0.769 feature_norm=29.61\n",
"Iter 40 time=0.21 loss=13174.95 active=152117 precision=0.726 recall=0.677 F1=0.689 Acc(item/seq)=0.965 0.766 feature_norm=31.03\n",
"Iter 41 time=0.21 loss=13104.00 active=152117 precision=0.736 recall=0.662 F1=0.678 Acc(item/seq)=0.964 0.760 feature_norm=33.06\n",
"Iter 42 time=0.21 loss=12750.79 active=152117 precision=0.731 recall=0.685 F1=0.701 Acc(item/seq)=0.966 0.764 feature_norm=34.35\n",
"Iter 43 time=0.21 loss=12637.02 active=152117 precision=0.740 recall=0.690 F1=0.708 Acc(item/seq)=0.966 0.766 feature_norm=34.63\n",
"Iter 44 time=0.21 loss=12534.20 active=152117 precision=0.745 recall=0.692 F1=0.712 Acc(item/seq)=0.966 0.766 feature_norm=35.30\n",
"Iter 45 time=0.21 loss=12390.89 active=152117 precision=0.739 recall=0.682 F1=0.702 Acc(item/seq)=0.966 0.761 feature_norm=37.15\n",
"Iter 46 time=0.21 loss=12277.42 active=152117 precision=0.733 recall=0.689 F1=0.704 Acc(item/seq)=0.966 0.763 feature_norm=37.58\n",
"Iter 47 time=0.21 loss=12219.03 active=152117 precision=0.739 recall=0.690 F1=0.706 Acc(item/seq)=0.966 0.767 feature_norm=37.20\n",
"Iter 48 time=0.21 loss=12125.64 active=152117 precision=0.743 recall=0.695 F1=0.711 Acc(item/seq)=0.967 0.770 feature_norm=36.99\n",
"Iter 49 time=0.21 loss=11970.65 active=152117 precision=0.745 recall=0.697 F1=0.712 Acc(item/seq)=0.967 0.775 feature_norm=36.84\n",
"Iter 50 time=0.21 loss=11780.73 active=152117 precision=0.754 recall=0.696 F1=0.718 Acc(item/seq)=0.968 0.777 feature_norm=37.57\n",
"Iter 51 time=0.21 loss=11623.83 active=152117 precision=0.740 recall=0.691 F1=0.710 Acc(item/seq)=0.967 0.774 feature_norm=38.21\n",
"Iter 52 time=0.21 loss=11549.38 active=152117 precision=0.740 recall=0.690 F1=0.709 Acc(item/seq)=0.967 0.773 feature_norm=38.94\n",
"Iter 53 time=0.21 loss=11497.85 active=152117 precision=0.739 recall=0.696 F1=0.713 Acc(item/seq)=0.967 0.772 feature_norm=39.54\n",
"Iter 54 time=0.21 loss=11419.64 active=152117 precision=0.733 recall=0.692 F1=0.707 Acc(item/seq)=0.966 0.773 feature_norm=40.40\n",
"Iter 55 time=0.21 loss=11280.29 active=152117 precision=0.744 recall=0.707 F1=0.719 Acc(item/seq)=0.967 0.773 feature_norm=41.75\n",
"Iter 56 time=0.21 loss=11131.39 active=152117 precision=0.746 recall=0.710 F1=0.722 Acc(item/seq)=0.968 0.773 feature_norm=42.72\n",
"Iter 57 time=0.21 loss=11043.40 active=152117 precision=0.751 recall=0.713 F1=0.726 Acc(item/seq)=0.968 0.774 feature_norm=42.92\n",
"Iter 58 time=0.21 loss=10954.38 active=152117 precision=0.769 recall=0.713 F1=0.736 Acc(item/seq)=0.969 0.781 feature_norm=43.18\n",
"Iter 59 time=0.21 loss=10836.31 active=152117 precision=0.773 recall=0.713 F1=0.736 Acc(item/seq)=0.969 0.781 feature_norm=43.74\n",
"Iter 60 time=0.21 loss=10712.24 active=152117 precision=0.779 recall=0.719 F1=0.744 Acc(item/seq)=0.970 0.788 feature_norm=44.41\n",
"Iter 61 time=0.21 loss=10602.81 active=152117 precision=0.789 recall=0.709 F1=0.740 Acc(item/seq)=0.970 0.789 feature_norm=44.68\n",
"Iter 62 time=0.21 loss=10508.84 active=152117 precision=0.782 recall=0.711 F1=0.739 Acc(item/seq)=0.970 0.787 feature_norm=45.37\n"
]
},
{
"name": "stdout",
"output_type": "stream",
"text": [
"Iter 63 time=0.21 loss=10458.88 active=152117 precision=0.783 recall=0.717 F1=0.744 Acc(item/seq)=0.970 0.788 feature_norm=45.51\n",
"Iter 64 time=0.21 loss=10420.78 active=152117 precision=0.763 recall=0.711 F1=0.730 Acc(item/seq)=0.969 0.786 feature_norm=45.67\n",
"Iter 65 time=0.21 loss=10315.28 active=152117 precision=0.766 recall=0.721 F1=0.735 Acc(item/seq)=0.969 0.786 feature_norm=46.30\n",
"Iter 66 time=0.21 loss=10204.10 active=152117 precision=0.769 recall=0.728 F1=0.740 Acc(item/seq)=0.970 0.786 feature_norm=47.10\n",
"Iter 67 time=0.21 loss=10134.54 active=152117 precision=0.769 recall=0.716 F1=0.737 Acc(item/seq)=0.970 0.787 feature_norm=47.87\n",
"Iter 68 time=0.21 loss=10095.10 active=152117 precision=0.773 recall=0.718 F1=0.741 Acc(item/seq)=0.970 0.787 feature_norm=47.85\n",
"Iter 69 time=0.21 loss=10059.52 active=152117 precision=0.773 recall=0.715 F1=0.738 Acc(item/seq)=0.970 0.790 feature_norm=47.81\n",
"Iter 70 time=0.21 loss=10012.53 active=152117 precision=0.767 recall=0.712 F1=0.733 Acc(item/seq)=0.970 0.790 feature_norm=47.95\n",
"Iter 71 time=0.21 loss=9931.41 active=152117 precision=0.765 recall=0.710 F1=0.731 Acc(item/seq)=0.970 0.791 feature_norm=48.49\n",
"Iter 72 time=0.21 loss=9861.45 active=152117 precision=0.763 recall=0.709 F1=0.730 Acc(item/seq)=0.970 0.793 feature_norm=49.07\n",
"Iter 73 time=0.21 loss=9803.75 active=152117 precision=0.769 recall=0.723 F1=0.741 Acc(item/seq)=0.970 0.796 feature_norm=49.32\n",
"Iter 74 time=0.21 loss=9762.61 active=152117 precision=0.758 recall=0.720 F1=0.733 Acc(item/seq)=0.970 0.793 feature_norm=49.58\n",
"Iter 75 time=0.21 loss=9726.47 active=152117 precision=0.761 recall=0.722 F1=0.736 Acc(item/seq)=0.970 0.790 feature_norm=49.80\n",
"Iter 76 time=0.21 loss=9628.97 active=152117 precision=0.764 recall=0.722 F1=0.736 Acc(item/seq)=0.970 0.789 feature_norm=50.47\n",
"Iter 77 time=0.40 loss=9586.61 active=152117 precision=0.763 recall=0.725 F1=0.740 Acc(item/seq)=0.971 0.791 feature_norm=50.96\n",
"Iter 78 time=0.21 loss=9522.00 active=152117 precision=0.767 recall=0.723 F1=0.741 Acc(item/seq)=0.970 0.788 feature_norm=51.34\n",
"Iter 79 time=0.21 loss=9479.87 active=152117 precision=0.765 recall=0.720 F1=0.736 Acc(item/seq)=0.970 0.788 feature_norm=51.77\n",
"Iter 80 time=0.21 loss=9448.33 active=152117 precision=0.771 recall=0.720 F1=0.739 Acc(item/seq)=0.970 0.789 feature_norm=51.80\n",
"Iter 81 time=0.21 loss=9423.55 active=152117 precision=0.769 recall=0.723 F1=0.740 Acc(item/seq)=0.970 0.789 feature_norm=51.83\n",
"Iter 82 time=0.21 loss=9367.51 active=152117 precision=0.763 recall=0.720 F1=0.736 Acc(item/seq)=0.970 0.790 feature_norm=52.25\n",
"Iter 83 time=0.21 loss=9322.98 active=152117 precision=0.762 recall=0.722 F1=0.737 Acc(item/seq)=0.970 0.793 feature_norm=52.34\n",
"Iter 84 time=0.21 loss=9277.51 active=152117 precision=0.765 recall=0.725 F1=0.740 Acc(item/seq)=0.971 0.798 feature_norm=52.47\n",
"Iter 85 time=0.21 loss=9229.07 active=152117 precision=0.772 recall=0.726 F1=0.743 Acc(item/seq)=0.971 0.798 feature_norm=52.79\n",
"Iter 86 time=0.21 loss=9214.50 active=152117 precision=0.782 recall=0.729 F1=0.751 Acc(item/seq)=0.971 0.799 feature_norm=52.88\n",
"Iter 87 time=0.21 loss=9194.55 active=152117 precision=0.785 recall=0.731 F1=0.753 Acc(item/seq)=0.971 0.800 feature_norm=52.82\n",
"Iter 88 time=0.21 loss=9182.23 active=152117 precision=0.783 recall=0.728 F1=0.750 Acc(item/seq)=0.971 0.798 feature_norm=52.86\n",
"Iter 89 time=0.21 loss=9161.54 active=152117 precision=0.786 recall=0.726 F1=0.750 Acc(item/seq)=0.971 0.795 feature_norm=52.98\n",
"Iter 90 time=0.21 loss=9130.06 active=152117 precision=0.788 recall=0.728 F1=0.752 Acc(item/seq)=0.971 0.797 feature_norm=53.17\n",
"Iter 91 time=0.21 loss=9063.09 active=152117 precision=0.794 recall=0.738 F1=0.760 Acc(item/seq)=0.972 0.800 feature_norm=53.64\n",
"Iter 92 time=0.40 loss=9041.98 active=152117 precision=0.791 recall=0.739 F1=0.759 Acc(item/seq)=0.972 0.801 feature_norm=53.77\n",
"Iter 93 time=0.21 loss=9010.27 active=152117 precision=0.783 recall=0.736 F1=0.754 Acc(item/seq)=0.972 0.801 feature_norm=53.97\n",
"Iter 94 time=0.21 loss=8984.80 active=152117 precision=0.786 recall=0.740 F1=0.759 Acc(item/seq)=0.972 0.803 feature_norm=54.09\n",
"Iter 95 time=0.21 loss=8965.14 active=152117 precision=0.788 recall=0.740 F1=0.760 Acc(item/seq)=0.972 0.803 feature_norm=54.14\n",
"Iter 96 time=0.21 loss=8931.00 active=152117 precision=0.786 recall=0.737 F1=0.758 Acc(item/seq)=0.972 0.804 feature_norm=54.26\n",
"Iter 97 time=0.21 loss=8923.90 active=152117 precision=0.791 recall=0.737 F1=0.757 Acc(item/seq)=0.971 0.799 feature_norm=54.68\n",
"Iter 98 time=0.21 loss=8860.78 active=152117 precision=0.784 recall=0.735 F1=0.755 Acc(item/seq)=0.971 0.802 feature_norm=54.64\n",
"Iter 99 time=0.21 loss=8846.13 active=152117 precision=0.788 recall=0.737 F1=0.758 Acc(item/seq)=0.972 0.803 feature_norm=54.68\n",
"Iter 100 time=0.21 loss=8832.34 active=152117 precision=0.789 recall=0.737 F1=0.758 Acc(item/seq)=0.972 0.803 feature_norm=54.77\n",
"================================================\n",
"Label Precision Recall F1 Support\n",
"------- ----------- -------- ----- ---------\n",
"B-LOC 0.823 0.823 0.823 479\n",
"B-MISC 0.806 0.651 0.720 748\n",
"B-ORG 0.853 0.620 0.718 686\n",
"B-PER 0.752 0.855 0.800 703\n",
"I-LOC 0.583 0.547 0.565 64\n",
"I-MISC 0.645 0.507 0.568 215\n",
"I-ORG 0.806 0.684 0.740 396\n",
"I-PER 0.844 0.946 0.892 423\n",
"O 0.990 0.998 0.994 33973\n",
"------------------------------------------------\n",
"L-BFGS terminated with the maximum number of iterations\n",
"Total seconds required for training: 21.768\n",
"\n",
"Storing the model\n",
"Number of active features: 152117 (152117)\n",
"Number of active attributes: 130306 (145602)\n",
"Number of active labels: 9 (9)\n",
"Writing labels\n",
"Writing attributes\n",
"Writing feature references for transitions\n",
"Writing feature references for attributes\n",
"Seconds required: 0.058\n",
"\n"
]
},
{
"data": {
"text/plain": [
"CRF(algorithm='lbfgs', all_possible_states=None,\n",
" all_possible_transitions=None, averaging=None, c=None, c1=None, c2=None,\n",
" calibration_candidates=None, calibration_eta=None,\n",
" calibration_max_trials=None, calibration_rate=None,\n",
" calibration_samples=None, delta=None, epsilon=None, error_sensitive=None,\n",
" gamma=None, keep_tempfiles=None, linesearch=None, max_iterations=100,\n",
" max_linesearch=None, min_freq=None, model_filename=None,\n",
" num_memories=None, pa_type=None, period=None, trainer_cls=None,\n",
" variance=None, verbose='true')"
]
},
"execution_count": 7,
"metadata": {},
"output_type": "execute_result"
}
],
"source": [
"crf = crfsuite.CRF(\n",
" verbose='true',\n",
" algorithm='lbfgs',\n",
" max_iterations=100\n",
")\n",
"\n",
"crf.fit(X_train, y_train, X_dev=X_dev, y_dev=y_dev)"
]
},
{
"cell_type": "code",
"execution_count": 8,
"metadata": {},
"outputs": [
{
"data": {
"text/plain": [
"['models/ner/crf_model']"
]
},
"execution_count": 8,
"metadata": {},
"output_type": "execute_result"
}
],
"source": [
"import joblib\n",
"import os\n",
"\n",
"OUTPUT_PATH = \"models/ner/\"\n",
"OUTPUT_FILE = \"crf_model\"\n",
"\n",
"if not os.path.exists(OUTPUT_PATH):\n",
" os.mkdir(OUTPUT_PATH)\n",
"\n",
"joblib.dump(crf, os.path.join(OUTPUT_PATH, OUTPUT_FILE))"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"## Evaluation"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"Let's evaluate the output of our CRF. We'll load the model from the output file above and have it predict labels for the full test set.\n",
"\n",
"As a sanity check, let's take a look at its predictions for the first test sentence. This output looks pretty good: the CRF is able to predict all four locations in the sentence correctly. It only misses the person entity, which is a strange case anyway, because it is not actually a person name."
]
},
{
"cell_type": "code",
"execution_count": 9,
"metadata": {},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"Sentence: Dat is in Italië , Spanje of Engeland misschien geen probleem , maar volgens ' Der Kaiser ' in Duitsland wel .\n",
"Predicted: O O O B-LOC O B-LOC O B-LOC O O O O O O O B-MISC I-MISC O O B-LOC O O\n",
"Correct: O O O B-LOC O B-LOC O B-LOC O O O O O O O B-PER I-PER O O B-LOC O O\n"
]
}
],
"source": [
"crf = joblib.load(os.path.join(OUTPUT_PATH, OUTPUT_FILE))\n",
"y_pred = crf.predict(X_test)\n",
"\n",
"example_sent = test_sents[0]\n",
"\n",
"print(\"Sentence:\", ' '.join(sent2tokens(example_sent)))\n",
"print(\"Predicted:\", ' '.join(crf.predict([sent2features(example_sent, word2cluster)])[0]))\n",
"print(\"Correct: \", ' '.join(sent2labels(example_sent)))"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"Now we evaluate on the full test set. We'll print out a classification report for all labels except `O`. If we were to include `O`, which far outnumbers the entity labels in our data, the average scores would be inflated artificially, simply because there's an inherently high probability that the `O` labels from our CRF are correct. We obtain an average F-score of 77% (micro average) across all entity types, with particularly good results for `B-LOC`and `B-PER`. "
]
},
{
"cell_type": "code",
"execution_count": 10,
"metadata": {},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
" precision recall f1-score support\n",
"\n",
" B-LOC 0.83 0.83 0.83 774\n",
" I-LOC 0.29 0.41 0.34 49\n",
" B-MISC 0.84 0.61 0.71 1187\n",
" I-MISC 0.59 0.42 0.49 410\n",
" B-ORG 0.80 0.69 0.74 882\n",
" I-ORG 0.74 0.66 0.70 551\n",
" B-PER 0.80 0.90 0.85 1098\n",
" I-PER 0.87 0.95 0.91 807\n",
"\n",
" micro avg 0.80 0.74 0.77 5758\n",
" macro avg 0.72 0.68 0.70 5758\n",
"weighted avg 0.80 0.74 0.76 5758\n",
"\n"
]
}
],
"source": [
"labels = list(crf.classes_)\n",
"labels.remove(\"O\")\n",
"y_pred = crf.predict(X_test)\n",
"sorted_labels = sorted(\n",
" labels,\n",
" key=lambda name: (name[1:], name[0])\n",
")\n",
"\n",
"print(metrics.flat_classification_report(y_test, y_pred, labels=sorted_labels))"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"Now we can also look at the most likely transitions the CRF has identified, and at the top features for every label. We'll do this with the `eli5` library, which helps us explain the predictions of machine learning models.\n",
"\n",
"The top transitions are quite intuitive: the most likely transitions are those within the same entity type (from a B-label to an O-label), and those where a B-label follows an O-label. \n",
"\n",
"The features, too, make sense. For example, if a word does not start with an uppercase letter, it is unlikely to be an entity. By contrast, a word is very likely to be a location if it ends in `ië`, which is indeed a very common suffix for locations in Dutch. Notice also how informative the embedding clusters are: for all entity types, the word clusters form some of the most informative features for the CRF. "
]
},
{
"cell_type": "code",
"execution_count": 11,
"metadata": {},
"outputs": [
{
"data": {
"text/html": [
"\n",
" <style>\n",
" table.eli5-weights tr:hover {\n",
" filter: brightness(85%);\n",
" }\n",
"</style>\n",
"\n",
"\n",
"\n",
" \n",
"\n",
" \n",
"\n",
" \n",
" \n",
"\n",
"\n",
"<table class=\"eli5-transition-features\" style=\"margin-bottom: 0.5em;\">\n",
" <thead>\n",
" <tr>\n",
" <td>From \\ To</td>\n",
" \n",
" <th>O</th>\n",
" \n",
" <th>B-LOC</th>\n",
" \n",
" <th>I-LOC</th>\n",
" \n",
" <th>B-MISC</th>\n",
" \n",
" <th>I-MISC</th>\n",
" \n",
" <th>B-ORG</th>\n",
" \n",
" <th>I-ORG</th>\n",
" \n",
" <th>B-PER</th>\n",
" \n",
" <th>I-PER</th>\n",
" \n",
" </tr>\n",
" </thead>\n",
" <tbody>\n",
" \n",
" \n",
" <tr>\n",
" <th>O</th>\n",
" \n",
" <td style=\"background-color: hsl(120, 100.00%, 88.09%)\" title=\"O ⇒ O\">\n",
" 4.141\n",
" </td>\n",
" \n",
" <td style=\"background-color: hsl(120, 100.00%, 87.21%)\" title=\"O ⇒ B-LOC\">\n",
" 4.583\n",
" </td>\n",
" \n",
" <td style=\"background-color: hsl(0, 100.00%, 100.00%)\" title=\"O ⇒ I-LOC\">\n",
" 0.0\n",
" </td>\n",
" \n",
" <td style=\"background-color: hsl(120, 100.00%, 88.09%)\" title=\"O ⇒ B-MISC\">\n",
" 4.141\n",
" </td>\n",
" \n",
" <td style=\"background-color: hsl(0, 100.00%, 100.00%)\" title=\"O ⇒ I-MISC\">\n",
" 0.0\n",
" </td>\n",
" \n",
" <td style=\"background-color: hsl(120, 100.00%, 87.64%)\" title=\"O ⇒ B-ORG\">\n",
" 4.366\n",
" </td>\n",
" \n",
" <td style=\"background-color: hsl(0, 100.00%, 100.00%)\" title=\"O ⇒ I-ORG\">\n",
" 0.0\n",
" </td>\n",
" \n",
" <td style=\"background-color: hsl(120, 100.00%, 88.74%)\" title=\"O ⇒ B-PER\">\n",
" 3.819\n",
" </td>\n",
" \n",
" <td style=\"background-color: hsl(0, 100.00%, 100.00%)\" title=\"O ⇒ I-PER\">\n",
" 0.0\n",
" </td>\n",
" \n",
" </tr>\n",
" \n",
" <tr>\n",
" <th>B-LOC</th>\n",
" \n",
" <td style=\"background-color: hsl(0, 100.00%, 98.34%)\" title=\"B-LOC ⇒ O\">\n",
" -0.248\n",
" </td>\n",
" \n",
" <td style=\"background-color: hsl(0, 100.00%, 98.20%)\" title=\"B-LOC ⇒ B-LOC\">\n",
" -0.279\n",
" </td>\n",
" \n",
" <td style=\"background-color: hsl(120, 100.00%, 82.62%)\" title=\"B-LOC ⇒ I-LOC\">\n",
" 7.101\n",
" </td>\n",
" \n",
" <td style=\"background-color: hsl(0, 100.00%, 100.00%)\" title=\"B-LOC ⇒ B-MISC\">\n",
" 0.0\n",
" </td>\n",
" \n",
" <td style=\"background-color: hsl(0, 100.00%, 100.00%)\" title=\"B-LOC ⇒ I-MISC\">\n",
" 0.0\n",
" </td>\n",
" \n",
" <td style=\"background-color: hsl(0, 100.00%, 100.00%)\" title=\"B-LOC ⇒ B-ORG\">\n",
" 0.0\n",
" </td>\n",
" \n",
" <td style=\"background-color: hsl(0, 100.00%, 100.00%)\" title=\"B-LOC ⇒ I-ORG\">\n",
" 0.0\n",
" </td>\n",
" \n",
" <td style=\"background-color: hsl(0, 100.00%, 96.70%)\" title=\"B-LOC ⇒ B-PER\">\n",
" -0.661\n",
" </td>\n",
" \n",
" <td style=\"background-color: hsl(0, 100.00%, 100.00%)\" title=\"B-LOC ⇒ I-PER\">\n",
" 0.0\n",
" </td>\n",
" \n",
" </tr>\n",
" \n",
" <tr>\n",
" <th>I-LOC</th>\n",
" \n",
" <td style=\"background-color: hsl(0, 100.00%, 95.40%)\" title=\"I-LOC ⇒ O\">\n",
" -1.062\n",
" </td>\n",
" \n",
" <td style=\"background-color: hsl(0, 100.00%, 98.40%)\" title=\"I-LOC ⇒ B-LOC\">\n",
" -0.235\n",
" </td>\n",
" \n",
" <td style=\"background-color: hsl(120, 100.00%, 84.62%)\" title=\"I-LOC ⇒ I-LOC\">\n",
" 5.967\n",
" </td>\n",
" \n",
" <td style=\"background-color: hsl(0, 100.00%, 100.00%)\" title=\"I-LOC ⇒ B-MISC\">\n",
" 0.0\n",
" </td>\n",
" \n",
" <td style=\"background-color: hsl(0, 100.00%, 100.00%)\" title=\"I-LOC ⇒ I-MISC\">\n",
" 0.0\n",
" </td>\n",
" \n",
" <td style=\"background-color: hsl(0, 100.00%, 100.00%)\" title=\"I-LOC ⇒ B-ORG\">\n",
" 0.0\n",
" </td>\n",
" \n",
" <td style=\"background-color: hsl(0, 100.00%, 100.00%)\" title=\"I-LOC ⇒ I-ORG\">\n",
" 0.0\n",
" </td>\n",
" \n",
" <td style=\"background-color: hsl(0, 100.00%, 100.00%)\" title=\"I-LOC ⇒ B-PER\">\n",
" 0.0\n",
" </td>\n",
" \n",
" <td style=\"background-color: hsl(0, 100.00%, 100.00%)\" title=\"I-LOC ⇒ I-PER\">\n",
" 0.0\n",
" </td>\n",
" \n",
" </tr>\n",
" \n",
" <tr>\n",
" <th>B-MISC</th>\n",
" \n",
" <td style=\"background-color: hsl(0, 100.00%, 95.64%)\" title=\"B-MISC ⇒ O\">\n",
" -0.985\n",
" </td>\n",
" \n",
" <td style=\"background-color: hsl(120, 100.00%, 96.73%)\" title=\"B-MISC ⇒ B-LOC\">\n",
" 0.655\n",
" </td>\n",
" \n",
" <td style=\"background-color: hsl(0, 100.00%, 100.00%)\" title=\"B-MISC ⇒ I-LOC\">\n",
" 0.0\n",
" </td>\n",
" \n",
" <td style=\"background-color: hsl(0, 100.00%, 98.03%)\" title=\"B-MISC ⇒ B-MISC\">\n",
" -0.316\n",
" </td>\n",
" \n",
" <td style=\"background-color: hsl(120, 100.00%, 81.56%)\" title=\"B-MISC ⇒ I-MISC\">\n",
" 7.73\n",
" </td>\n",
" \n",
" <td style=\"background-color: hsl(120, 100.00%, 97.10%)\" title=\"B-MISC ⇒ B-ORG\">\n",
" 0.551\n",
" </td>\n",
" \n",
" <td style=\"background-color: hsl(0, 100.00%, 100.00%)\" title=\"B-MISC ⇒ I-ORG\">\n",
" 0.0\n",
" </td>\n",
" \n",
" <td style=\"background-color: hsl(120, 100.00%, 97.44%)\" title=\"B-MISC ⇒ B-PER\">\n",
" 0.46\n",
" </td>\n",
" \n",
" <td style=\"background-color: hsl(0, 100.00%, 100.00%)\" title=\"B-MISC ⇒ I-PER\">\n",
" 0.0\n",
" </td>\n",
" \n",
" </tr>\n",
" \n",
" <tr>\n",
" <th>I-MISC</th>\n",
" \n",
" <td style=\"background-color: hsl(0, 100.00%, 93.40%)\" title=\"I-MISC ⇒ O\">\n",
" -1.781\n",
" </td>\n",
" \n",
" <td style=\"background-color: hsl(0, 100.00%, 100.00%)\" title=\"I-MISC ⇒ B-LOC\">\n",
" 0.0\n",
" </td>\n",
" \n",
" <td style=\"background-color: hsl(0, 100.00%, 100.00%)\" title=\"I-MISC ⇒ I-LOC\">\n",
" 0.0\n",
" </td>\n",
" \n",
" <td style=\"background-color: hsl(0, 100.00%, 97.75%)\" title=\"I-MISC ⇒ B-MISC\">\n",
" -0.382\n",
" </td>\n",
" \n",
" <td style=\"background-color: hsl(120, 100.00%, 81.49%)\" title=\"I-MISC ⇒ I-MISC\">\n",
" 7.769\n",
" </td>\n",
" \n",
" <td style=\"background-color: hsl(120, 100.00%, 95.16%)\" title=\"I-MISC ⇒ B-ORG\">\n",
" 1.145\n",
" </td>\n",
" \n",
" <td style=\"background-color: hsl(0, 100.00%, 100.00%)\" title=\"I-MISC ⇒ I-ORG\">\n",
" 0.0\n",
" </td>\n",
" \n",
" <td style=\"background-color: hsl(0, 100.00%, 96.50%)\" title=\"I-MISC ⇒ B-PER\">\n",
" -0.719\n",
" </td>\n",
" \n",
" <td style=\"background-color: hsl(0, 100.00%, 100.00%)\" title=\"I-MISC ⇒ I-PER\">\n",
" 0.0\n",
" </td>\n",
" \n",
" </tr>\n",
" \n",
" <tr>\n",
" <th>B-ORG</th>\n",
" \n",
" <td style=\"background-color: hsl(0, 100.00%, 98.28%)\" title=\"B-ORG ⇒ O\">\n",
" -0.261\n",
" </td>\n",
" \n",
" <td style=\"background-color: hsl(0, 100.00%, 100.00%)\" title=\"B-ORG ⇒ B-LOC\">\n",
" 0.0\n",
" </td>\n",
" \n",
" <td style=\"background-color: hsl(0, 100.00%, 100.00%)\" title=\"B-ORG ⇒ I-LOC\">\n",
" 0.0\n",
" </td>\n",
" \n",
" <td style=\"background-color: hsl(0, 100.00%, 96.20%)\" title=\"B-ORG ⇒ B-MISC\">\n",
" -0.809\n",
" </td>\n",
" \n",
" <td style=\"background-color: hsl(0, 100.00%, 100.00%)\" title=\"B-ORG ⇒ I-MISC\">\n",
" 0.0\n",
" </td>\n",
" \n",
" <td style=\"background-color: hsl(0, 100.00%, 100.00%)\" title=\"B-ORG ⇒ B-ORG\">\n",
" 0.0\n",
" </td>\n",
" \n",
" <td style=\"background-color: hsl(120, 100.00%, 81.44%)\" title=\"B-ORG ⇒ I-ORG\">\n",
" 7.803\n",
" </td>\n",
" \n",
" <td style=\"background-color: hsl(120, 100.00%, 99.09%)\" title=\"B-ORG ⇒ B-PER\">\n",
" 0.106\n",
" </td>\n",
" \n",
" <td style=\"background-color: hsl(0, 100.00%, 100.00%)\" title=\"B-ORG ⇒ I-PER\">\n",
" 0.0\n",
" </td>\n",
" \n",
" </tr>\n",
" \n",
" <tr>\n",
" <th>I-ORG</th>\n",
" \n",
" <td style=\"background-color: hsl(0, 100.00%, 96.25%)\" title=\"I-ORG ⇒ O\">\n",
" -0.794\n",
" </td>\n",
" \n",
" <td style=\"background-color: hsl(0, 100.00%, 100.00%)\" title=\"I-ORG ⇒ B-LOC\">\n",
" 0.0\n",
" </td>\n",
" \n",
" <td style=\"background-color: hsl(0, 100.00%, 100.00%)\" title=\"I-ORG ⇒ I-LOC\">\n",
" 0.0\n",
" </td>\n",
" \n",
" <td style=\"background-color: hsl(0, 100.00%, 100.00%)\" title=\"I-ORG ⇒ B-MISC\">\n",
" 0.0\n",
" </td>\n",
" \n",
" <td style=\"background-color: hsl(0, 100.00%, 100.00%)\" title=\"I-ORG ⇒ I-MISC\">\n",
" 0.0\n",
" </td>\n",
" \n",
" <td style=\"background-color: hsl(0, 100.00%, 100.00%)\" title=\"I-ORG ⇒ B-ORG\">\n",
" 0.0\n",
" </td>\n",
" \n",
" <td style=\"background-color: hsl(120, 100.00%, 82.50%)\" title=\"I-ORG ⇒ I-ORG\">\n",
" 7.174\n",
" </td>\n",
" \n",
" <td style=\"background-color: hsl(120, 100.00%, 99.22%)\" title=\"I-ORG ⇒ B-PER\">\n",
" 0.084\n",
" </td>\n",
" \n",
" <td style=\"background-color: hsl(0, 100.00%, 100.00%)\" title=\"I-ORG ⇒ I-PER\">\n",
" 0.0\n",
" </td>\n",
" \n",
" </tr>\n",
" \n",
" <tr>\n",
" <th>B-PER</th>\n",
" \n",
" <td style=\"background-color: hsl(120, 100.00%, 98.06%)\" title=\"B-PER ⇒ O\">\n",
" 0.31\n",
" </td>\n",
" \n",
" <td style=\"background-color: hsl(0, 100.00%, 97.90%)\" title=\"B-PER ⇒ B-LOC\">\n",
" -0.346\n",
" </td>\n",
" \n",
" <td style=\"background-color: hsl(0, 100.00%, 100.00%)\" title=\"B-PER ⇒ I-LOC\">\n",
" 0.0\n",
" </td>\n",
" \n",
" <td style=\"background-color: hsl(0, 100.00%, 96.88%)\" title=\"B-PER ⇒ B-MISC\">\n",
" -0.611\n",
" </td>\n",
" \n",
" <td style=\"background-color: hsl(0, 100.00%, 100.00%)\" title=\"B-PER ⇒ I-MISC\">\n",
" 0.0\n",
" </td>\n",
" \n",
" <td style=\"background-color: hsl(0, 100.00%, 100.00%)\" title=\"B-PER ⇒ B-ORG\">\n",
" 0.0\n",
" </td>\n",
" \n",
" <td style=\"background-color: hsl(0, 100.00%, 100.00%)\" title=\"B-PER ⇒ I-ORG\">\n",
" 0.0\n",
" </td>\n",
" \n",
" <td style=\"background-color: hsl(0, 100.00%, 94.40%)\" title=\"B-PER ⇒ B-PER\">\n",
" -1.408\n",
" </td>\n",
" \n",
" <td style=\"background-color: hsl(120, 100.00%, 80.00%)\" title=\"B-PER ⇒ I-PER\">\n",
" 8.68\n",
" </td>\n",
" \n",
" </tr>\n",
" \n",
" <tr>\n",
" <th>I-PER</th>\n",
" \n",
" <td style=\"background-color: hsl(120, 100.00%, 99.10%)\" title=\"I-PER ⇒ O\">\n",
" 0.104\n",
" </td>\n",
" \n",
" <td style=\"background-color: hsl(0, 100.00%, 100.00%)\" title=\"I-PER ⇒ B-LOC\">\n",
" 0.0\n",
" </td>\n",
" \n",
" <td style=\"background-color: hsl(0, 100.00%, 100.00%)\" title=\"I-PER ⇒ I-LOC\">\n",
" 0.0\n",
" </td>\n",
" \n",
" <td style=\"background-color: hsl(0, 100.00%, 100.00%)\" title=\"I-PER ⇒ B-MISC\">\n",
" 0.0\n",
" </td>\n",
" \n",
" <td style=\"background-color: hsl(0, 100.00%, 100.00%)\" title=\"I-PER ⇒ I-MISC\">\n",
" 0.0\n",
" </td>\n",
" \n",
" <td style=\"background-color: hsl(0, 100.00%, 100.00%)\" title=\"I-PER ⇒ B-ORG\">\n",
" 0.0\n",
" </td>\n",
" \n",
" <td style=\"background-color: hsl(0, 100.00%, 100.00%)\" title=\"I-PER ⇒ I-ORG\">\n",
" 0.0\n",
" </td>\n",
" \n",
" <td style=\"background-color: hsl(0, 100.00%, 100.00%)\" title=\"I-PER ⇒ B-PER\">\n",
" 0.0\n",
" </td>\n",
" \n",
" <td style=\"background-color: hsl(120, 100.00%, 83.13%)\" title=\"I-PER ⇒ I-PER\">\n",
" 6.804\n",
" </td>\n",
" \n",
" </tr>\n",
" \n",
" \n",
" </tbody>\n",
"</table>\n",
" \n",
"\n",
" \n",
"\n",
" \n",
"\n",
" \n",
"\n",
"\n",
" \n",
"\n",
" \n",
"\n",
" \n",
"\n",
" \n",
" \n",
"\n",
" \n",
" <table class=\"eli5-weights-wrapper\" style=\"border-collapse: collapse; border: none; margin-bottom: 1.5em;\">\n",
" <tr>\n",
" \n",
" <td style=\"padding: 0.5em; border: 1px solid black; text-align: center;\">\n",
" <b>\n",
" \n",
" y=O\n",
" \n",
"</b>\n",
"\n",
"top features\n",
" </td>\n",
" \n",
" <td style=\"padding: 0.5em; border: 1px solid black; text-align: center;\">\n",
" <b>\n",
" \n",
" y=B-LOC\n",
" \n",
"</b>\n",
"\n",
"top features\n",
" </td>\n",
" \n",
" <td style=\"padding: 0.5em; border: 1px solid black; text-align: center;\">\n",
" <b>\n",
" \n",
" y=I-LOC\n",
" \n",
"</b>\n",
"\n",
"top features\n",
" </td>\n",
" \n",
" <td style=\"padding: 0.5em; border: 1px solid black; text-align: center;\">\n",
" <b>\n",
" \n",
" y=B-MISC\n",
" \n",
"</b>\n",
"\n",
"top features\n",
" </td>\n",
" \n",
" <td style=\"padding: 0.5em; border: 1px solid black; text-align: center;\">\n",
" <b>\n",
" \n",
" y=I-MISC\n",
" \n",
"</b>\n",
"\n",
"top features\n",
" </td>\n",
" \n",
" <td style=\"padding: 0.5em; border: 1px solid black; text-align: center;\">\n",
" <b>\n",
" \n",
" y=B-ORG\n",
" \n",
"</b>\n",
"\n",
"top features\n",
" </td>\n",
" \n",
" <td style=\"padding: 0.5em; border: 1px solid black; text-align: center;\">\n",
" <b>\n",
" \n",
" y=I-ORG\n",
" \n",
"</b>\n",
"\n",
"top features\n",
" </td>\n",
" \n",
" <td style=\"padding: 0.5em; border: 1px solid black; text-align: center;\">\n",
" <b>\n",
" \n",
" y=B-PER\n",
" \n",
"</b>\n",
"\n",
"top features\n",
" </td>\n",
" \n",
" <td style=\"padding: 0.5em; border: 1px solid black; text-align: center;\">\n",
" <b>\n",
" \n",
" y=I-PER\n",
" \n",
"</b>\n",
"\n",
"top features\n",
" </td>\n",
" \n",
" </tr>\n",
" <tr>\n",
" \n",
" \n",
" <td style=\"padding: 0px; border: 1px solid black; vertical-align: top;\">\n",
" \n",
" \n",
" \n",
" \n",
" \n",
" <table class=\"eli5-weights\"\n",
" style=\"border-collapse: collapse; border: none; margin-top: 0em; table-layout: auto; width: 100%;\">\n",
" <thead>\n",
" <tr style=\"border: none;\">\n",
" \n",
" <th style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\" title=\"Feature weights. Note that weights do not account for feature value scales, so if feature values have different scales, features with highest weights might not be the most important.\">\n",
" Weight<sup>?</sup>\n",
" </th>\n",
" \n",
" <th style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">Feature</th>\n",
" \n",
" </tr>\n",
" </thead>\n",
" <tbody>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 80.16%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +3.692\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word.istitle=False\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 81.61%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +3.312\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word.isupper=False\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 86.14%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +2.211\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" -1:word.lower="\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 87.08%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +1.999\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" -1:word.lower=+\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 87.84%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +1.835\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" EOS\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 87.90%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +1.820\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" BOS\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 88.28%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +1.740\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word.cluster=158\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 88.54%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +1.684\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word.cluster=415\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 88.99%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +1.590\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word[-2:]=ag\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 89.01%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +1.588\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word.cluster=195\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 89.16%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +1.556\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word[-3:]=dag\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 89.34%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +1.520\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word.cluster=185\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 89.34%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +1.520\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word.cluster=178\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 89.53%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +1.481\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word.cluster=24\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 89.55%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +1.477\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word.cluster=177\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 89.56%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +1.475\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word.lower=u\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 89.56%); border: none;\">\n",
" <td colspan=\"2\" style=\"padding: 0 0.5em 0 0.5em; text-align: center; border: none; white-space: nowrap;\">\n",
" <i>… 108173 more positive …</i>\n",
" </td>\n",
" </tr>\n",
" \n",
"\n",
" \n",
" <tr style=\"background-color: hsl(0, 100.00%, 89.36%); border: none;\">\n",
" <td colspan=\"2\" style=\"padding: 0 0.5em 0 0.5em; text-align: center; border: none; white-space: nowrap;\">\n",
" <i>… 6671 more negative …</i>\n",
" </td>\n",
" </tr>\n",
" \n",
" \n",
" <tr style=\"background-color: hsl(0, 100.00%, 89.36%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" -1.515\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word.cluster=375\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(0, 100.00%, 89.25%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" -1.538\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" -1:word.lower=ronde\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(0, 100.00%, 89.14%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" -1.561\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" +1:word.lower=(\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(0, 100.00%, 88.86%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" -1.618\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" 0\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(0, 100.00%, 88.50%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" -1.693\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" -1:word.lower=de\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(0, 100.00%, 88.28%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" -1.740\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" postag=N\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(0, 100.00%, 87.96%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" -1.808\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word[-2:]='s\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(0, 100.00%, 87.80%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" -1.843\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word[-3:]=ple\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(0, 100.00%, 87.75%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" -1.853\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" postag=Misc\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(0, 100.00%, 87.47%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" -1.914\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word.cluster=6\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(0, 100.00%, 87.33%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" -1.945\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" +2:word.lower=1\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(0, 100.00%, 87.13%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" -1.990\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word.cluster=111\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(0, 100.00%, 86.78%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" -2.068\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word.isupper=True\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(0, 100.00%, 85.12%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" -2.447\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word.istitle=True\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
"\n",
" </tbody>\n",
" </table>\n",
"\n",
" \n",
" \n",
" </td>\n",
" \n",
" <td style=\"padding: 0px; border: 1px solid black; vertical-align: top;\">\n",
" \n",
" \n",
" \n",
" \n",
" \n",
" <table class=\"eli5-weights\"\n",
" style=\"border-collapse: collapse; border: none; margin-top: 0em; table-layout: auto; width: 100%;\">\n",
" <thead>\n",
" <tr style=\"border: none;\">\n",
" \n",
" <th style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\" title=\"Feature weights. Note that weights do not account for feature value scales, so if feature values have different scales, features with highest weights might not be the most important.\">\n",
" Weight<sup>?</sup>\n",
" </th>\n",
" \n",
" <th style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">Feature</th>\n",
" \n",
" </tr>\n",
" </thead>\n",
" <tbody>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 80.00%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +3.734\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word.cluster=325\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 80.31%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +3.650\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word.cluster=375\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 81.01%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +3.466\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word.cluster=68\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 82.17%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +3.169\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word.cluster=139\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 86.99%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +2.020\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" -1:word.lower=in\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 87.10%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +1.995\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word.cluster=143\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 87.20%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +1.973\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word.cluster=476\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 87.87%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +1.828\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word.cluster=102\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 88.87%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +1.617\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word[-2:]=ië\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 89.25%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +1.538\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" -1:word.lower=(\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 89.36%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +1.516\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" -1:word.lower=uit\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 89.46%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +1.495\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word.cluster=154\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 89.73%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +1.441\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" +1:word.lower=/\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 89.85%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +1.417\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" +2:word.lower=-\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 89.91%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +1.404\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" +2:word.lower=tel.\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 90.04%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +1.379\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" -1:word.lower=het\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 90.07%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +1.374\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word.cluster=440\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 90.09%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +1.369\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word.lower=sint-michiels\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 90.25%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +1.337\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word[-3:]=urs\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 90.34%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +1.320\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word.lower=futuroscope\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 90.44%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +1.301\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word[-3:]=ope\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 90.44%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +1.301\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word.cluster=363\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 90.56%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +1.278\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word.lower=vst\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 90.56%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +1.278\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word[-3:]=VSt\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 90.67%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +1.257\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word[-3:]=den\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 90.73%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +1.245\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word.cluster=146\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 90.73%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +1.244\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word.cluster=116\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 90.90%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +1.212\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word[-2:]=St\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 90.98%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +1.198\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word.cluster=492\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 90.98%); border: none;\">\n",
" <td colspan=\"2\" style=\"padding: 0 0.5em 0 0.5em; text-align: center; border: none; white-space: nowrap;\">\n",
" <i>… 4441 more positive …</i>\n",
" </td>\n",
" </tr>\n",
" \n",
"\n",
" \n",
" <tr style=\"background-color: hsl(0, 100.00%, 90.33%); border: none;\">\n",
" <td colspan=\"2\" style=\"padding: 0 0.5em 0 0.5em; text-align: center; border: none; white-space: nowrap;\">\n",
" <i>… 668 more negative …</i>\n",
" </td>\n",
" </tr>\n",
" \n",
" \n",
" <tr style=\"background-color: hsl(0, 100.00%, 90.33%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" -1.322\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" -2:word.lower=ronde\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
"\n",
" </tbody>\n",
" </table>\n",
"\n",
" \n",
" \n",
" </td>\n",
" \n",
" <td style=\"padding: 0px; border: 1px solid black; vertical-align: top;\">\n",
" \n",
" \n",
" \n",
" \n",
" \n",
" <table class=\"eli5-weights\"\n",
" style=\"border-collapse: collapse; border: none; margin-top: 0em; table-layout: auto; width: 100%;\">\n",
" <thead>\n",
" <tr style=\"border: none;\">\n",
" \n",
" <th style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\" title=\"Feature weights. Note that weights do not account for feature value scales, so if feature values have different scales, features with highest weights might not be the most important.\">\n",
" Weight<sup>?</sup>\n",
" </th>\n",
" \n",
" <th style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">Feature</th>\n",
" \n",
" </tr>\n",
" </thead>\n",
" <tbody>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 88.10%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +1.778\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word.cluster=238\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 89.50%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +1.488\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" +2:word.lower=m\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 90.10%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +1.367\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word.cluster=161\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 92.07%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +0.996\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" -1:word.lower=col\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 92.17%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +0.977\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word[-2:]=rk\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 92.89%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +0.852\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" -1:word.istitle=False\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 93.07%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +0.821\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word[-2:]=al\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 93.14%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +0.810\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word.lower=york\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 93.14%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +0.810\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word[-3:]=ork\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 93.14%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +0.809\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word[-3:]=eum\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 93.31%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +0.781\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" -1:word.lower=san\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 93.79%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +0.703\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word.lower=abeba\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 93.79%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +0.703\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" -1:word.lower=addis\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 93.79%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +0.702\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word[-3:]=eba\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 93.79%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +0.701\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" -1:word.isupper=True\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 93.91%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +0.683\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" -2:word.lower='\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 93.92%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +0.682\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" +2:word.lower=,\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 93.95%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +0.677\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word[-3:]=den\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 93.95%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +0.676\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word[-3:]=aan\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 93.97%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +0.674\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word[-2:]=um\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 93.99%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +0.670\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word.lower=staten\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 94.00%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +0.669\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word.cluster=38\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 94.05%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +0.661\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" postag=Prep\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 94.12%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +0.650\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" -2:word.lower=col\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 94.12%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +0.650\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" -1:postag=Art\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 94.13%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +0.648\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" -1:word.lower=museum\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 94.15%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +0.645\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word[-2:]=ba\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 94.16%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +0.643\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word.lower=hotel\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 94.16%); border: none;\">\n",
" <td colspan=\"2\" style=\"padding: 0 0.5em 0 0.5em; text-align: center; border: none; white-space: nowrap;\">\n",
" <i>… 1294 more positive …</i>\n",
" </td>\n",
" </tr>\n",
" \n",
"\n",
" \n",
" <tr style=\"background-color: hsl(0, 100.00%, 93.97%); border: none;\">\n",
" <td colspan=\"2\" style=\"padding: 0 0.5em 0 0.5em; text-align: center; border: none; white-space: nowrap;\">\n",
" <i>… 131 more negative …</i>\n",
" </td>\n",
" </tr>\n",
" \n",
" \n",
" <tr style=\"background-color: hsl(0, 100.00%, 93.97%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" -0.674\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" EOS\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(0, 100.00%, 93.31%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" -0.781\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" postag=Adj\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
"\n",
" </tbody>\n",
" </table>\n",
"\n",
" \n",
" \n",
" </td>\n",
" \n",
" <td style=\"padding: 0px; border: 1px solid black; vertical-align: top;\">\n",
" \n",
" \n",
" \n",
" \n",
" \n",
" <table class=\"eli5-weights\"\n",
" style=\"border-collapse: collapse; border: none; margin-top: 0em; table-layout: auto; width: 100%;\">\n",
" <thead>\n",
" <tr style=\"border: none;\">\n",
" \n",
" <th style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\" title=\"Feature weights. Note that weights do not account for feature value scales, so if feature values have different scales, features with highest weights might not be the most important.\">\n",
" Weight<sup>?</sup>\n",
" </th>\n",
" \n",
" <th style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">Feature</th>\n",
" \n",
" </tr>\n",
" </thead>\n",
" <tbody>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 81.59%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +3.318\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word.cluster=23\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 84.92%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +2.494\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word.cluster=100\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 85.24%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +2.419\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word.cluster=39\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 86.65%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +2.097\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" +2:word.lower=1\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 86.91%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +2.039\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word.cluster=294\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 86.92%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +2.036\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word.cluster=338\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 88.06%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +1.786\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word[-2:]='s\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 88.13%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +1.772\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word.cluster=11\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 88.47%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +1.700\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word.lower=sport\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 88.73%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +1.646\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word.lower=buitenland\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 88.81%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +1.628\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word[-2:]=se\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 88.86%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +1.618\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word.cluster=222\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 89.27%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +1.535\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word.cluster=427\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 89.41%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +1.505\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word.lower=journaal\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 89.54%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +1.478\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word.cluster=40\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 89.85%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +1.417\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word[-3:]=nse\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 89.88%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +1.411\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" postag=Adj\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 90.06%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +1.374\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word.cluster=111\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 90.19%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +1.350\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word.lower=aula\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 90.20%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +1.347\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word.lower=tobin-taks\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 90.30%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +1.329\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word.cluster=218\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 90.44%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +1.301\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word.cluster=128\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 90.58%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +1.274\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word[-3:]=ula\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 90.84%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +1.223\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" BOS\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 90.90%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +1.213\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word.lower=tobin-heffing\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 90.94%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +1.204\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word.lower=ronde\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 90.94%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +1.204\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word[-2:]=ks\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 90.98%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +1.197\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word.lower=v-plan\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 91.05%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +1.184\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word[-2:]=ex\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 91.05%); border: none;\">\n",
" <td colspan=\"2\" style=\"padding: 0 0.5em 0 0.5em; text-align: center; border: none; white-space: nowrap;\">\n",
" <i>… 5588 more positive …</i>\n",
" </td>\n",
" </tr>\n",
" \n",
"\n",
" \n",
" <tr style=\"background-color: hsl(0, 100.00%, 90.67%); border: none;\">\n",
" <td colspan=\"2\" style=\"padding: 0 0.5em 0 0.5em; text-align: center; border: none; white-space: nowrap;\">\n",
" <i>… 853 more negative …</i>\n",
" </td>\n",
" </tr>\n",
" \n",
" \n",
" <tr style=\"background-color: hsl(0, 100.00%, 90.67%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" -1.256\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" -1:word.isupper=True\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
"\n",
" </tbody>\n",
" </table>\n",
"\n",
" \n",
" \n",
" </td>\n",
" \n",
" <td style=\"padding: 0px; border: 1px solid black; vertical-align: top;\">\n",
" \n",
" \n",
" \n",
" \n",
" \n",
" <table class=\"eli5-weights\"\n",
" style=\"border-collapse: collapse; border: none; margin-top: 0em; table-layout: auto; width: 100%;\">\n",
" <thead>\n",
" <tr style=\"border: none;\">\n",
" \n",
" <th style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\" title=\"Feature weights. Note that weights do not account for feature value scales, so if feature values have different scales, features with highest weights might not be the most important.\">\n",
" Weight<sup>?</sup>\n",
" </th>\n",
" \n",
" <th style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">Feature</th>\n",
" \n",
" </tr>\n",
" </thead>\n",
" <tbody>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 88.04%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +1.792\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" -2:word.lower=ronde\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 88.54%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +1.684\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" -1:word.isupper=True\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 89.11%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +1.567\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" -1:word.lower=ronde\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 90.28%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +1.332\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word.cluster=37\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 90.32%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +1.323\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" +1:word.lower=ned\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 90.36%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +1.316\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word.cluster=325\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 90.45%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +1.298\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word.cluster=1\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 90.58%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +1.274\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word.lower=leven\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 90.89%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +1.215\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" -1:word.istitle=True\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 90.96%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +1.201\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" -1:postag=Num\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 91.06%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +1.182\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" -1:word.lower=euro\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 91.17%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +1.161\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word.cluster=86\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 91.24%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +1.147\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word.lower=bijsluiter\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 91.48%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +1.103\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" -1:word.lower=financiële\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 91.50%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +1.100\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" +2:word.lower=ned\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 91.52%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +1.096\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word[-2:]=ap\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 91.67%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +1.068\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" postag=Num\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 91.82%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +1.042\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word.cluster=413\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 91.89%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +1.029\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word[-2:]=00\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 91.98%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +1.012\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" -1:word.lower=de\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 91.99%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +1.010\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" -1:postag=Prep\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 92.16%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +0.979\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word.cluster=279\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 92.31%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +0.954\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" -1:word.lower=travel\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 92.33%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +0.950\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" -1:word.lower=formule\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 92.34%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +0.947\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word.lower=financiële\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 92.36%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +0.945\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" -1:word.lower=brussel\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 92.52%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +0.916\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" +1:word.lower='\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 92.52%); border: none;\">\n",
" <td colspan=\"2\" style=\"padding: 0 0.5em 0 0.5em; text-align: center; border: none; white-space: nowrap;\">\n",
" <i>… 3172 more positive …</i>\n",
" </td>\n",
" </tr>\n",
" \n",
"\n",
" \n",
" <tr style=\"background-color: hsl(0, 100.00%, 91.42%); border: none;\">\n",
" <td colspan=\"2\" style=\"padding: 0 0.5em 0 0.5em; text-align: center; border: none; white-space: nowrap;\">\n",
" <i>… 465 more negative …</i>\n",
" </td>\n",
" </tr>\n",
" \n",
" \n",
" <tr style=\"background-color: hsl(0, 100.00%, 91.42%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" -1.115\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" postag=V\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(0, 100.00%, 89.60%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" -1.467\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" +1:postag=Num\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(0, 100.00%, 83.03%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" -2.953\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" -1:postag=Adj\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
"\n",
" </tbody>\n",
" </table>\n",
"\n",
" \n",
" \n",
" </td>\n",
" \n",
" <td style=\"padding: 0px; border: 1px solid black; vertical-align: top;\">\n",
" \n",
" \n",
" \n",
" \n",
" \n",
" <table class=\"eli5-weights\"\n",
" style=\"border-collapse: collapse; border: none; margin-top: 0em; table-layout: auto; width: 100%;\">\n",
" <thead>\n",
" <tr style=\"border: none;\">\n",
" \n",
" <th style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\" title=\"Feature weights. Note that weights do not account for feature value scales, so if feature values have different scales, features with highest weights might not be the most important.\">\n",
" Weight<sup>?</sup>\n",
" </th>\n",
" \n",
" <th style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">Feature</th>\n",
" \n",
" </tr>\n",
" </thead>\n",
" <tbody>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 83.66%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +2.798\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word.cluster=424\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 84.33%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +2.635\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word.cluster=228\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 86.54%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +2.121\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word[-3:]=com\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 87.12%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +1.991\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word.cluster=187\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 87.20%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +1.974\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word.lower=quizpeople\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 87.44%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +1.922\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word[-3:]=ple\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 87.78%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +1.848\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" +1:word.lower=morgen\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 88.01%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +1.798\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word.cluster=250\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 88.55%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +1.683\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word.cluster=83\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 89.14%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +1.560\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word.cluster=29\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 89.25%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +1.538\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word[-2:]=ga\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 89.31%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +1.525\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" BOS\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 89.44%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +1.500\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word[-3:]=bel\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 89.55%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +1.476\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" -2:word.lower=minister\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 89.82%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +1.422\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word.cluster=207\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 89.85%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +1.418\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word.cluster=249\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 89.87%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +1.412\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word.lower=waterleau\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 89.88%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +1.411\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word.isupper=True\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 90.02%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +1.384\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word.lower=belga\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 90.02%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +1.384\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word[-3:]=lga\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 90.30%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +1.328\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word[-3:]=our\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 90.34%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +1.320\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word[-2:]=om\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 90.35%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +1.319\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word[-2:]=to\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 90.45%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +1.299\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word.lower=freenet\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 90.56%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +1.277\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word.lower=baan\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 90.76%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +1.240\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" -1:word.lower=bij\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 90.82%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +1.227\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word[-3:]=lux\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 90.82%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +1.227\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word[-2:]=co\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 90.85%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +1.222\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word.cluster=21\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 90.85%); border: none;\">\n",
" <td colspan=\"2\" style=\"padding: 0 0.5em 0 0.5em; text-align: center; border: none; white-space: nowrap;\">\n",
" <i>… 3591 more positive …</i>\n",
" </td>\n",
" </tr>\n",
" \n",
"\n",
" \n",
" <tr style=\"background-color: hsl(0, 100.00%, 91.20%); border: none;\">\n",
" <td colspan=\"2\" style=\"padding: 0 0.5em 0 0.5em; text-align: center; border: none; white-space: nowrap;\">\n",
" <i>… 469 more negative …</i>\n",
" </td>\n",
" </tr>\n",
" \n",
" \n",
" <tr style=\"background-color: hsl(0, 100.00%, 91.20%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" -1.156\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word.isupper=False\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
"\n",
" </tbody>\n",
" </table>\n",
"\n",
" \n",
" \n",
" </td>\n",
" \n",
" <td style=\"padding: 0px; border: 1px solid black; vertical-align: top;\">\n",
" \n",
" \n",
" \n",
" \n",
" \n",
" <table class=\"eli5-weights\"\n",
" style=\"border-collapse: collapse; border: none; margin-top: 0em; table-layout: auto; width: 100%;\">\n",
" <thead>\n",
" <tr style=\"border: none;\">\n",
" \n",
" <th style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\" title=\"Feature weights. Note that weights do not account for feature value scales, so if feature values have different scales, features with highest weights might not be the most important.\">\n",
" Weight<sup>?</sup>\n",
" </th>\n",
" \n",
" <th style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">Feature</th>\n",
" \n",
" </tr>\n",
" </thead>\n",
" <tbody>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 90.25%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +1.338\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word.lower=morgen\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 90.42%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +1.304\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word.cluster=403\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 90.74%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +1.243\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word.cluster=413\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 90.96%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +1.200\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" -1:word.lower=vlaams\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 91.28%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +1.141\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word.cluster=187\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 91.39%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +1.120\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word[-3:]=gen\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 91.49%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +1.101\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word[-3:]=ion\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 91.73%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +1.057\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" -1:word.lower=radio\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 91.89%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +1.028\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word.cluster=321\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 92.22%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +0.970\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" -1:word.lower=ned\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 92.91%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +0.849\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word[-2:]=nt\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 92.95%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +0.841\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word.cluster=143\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 93.17%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +0.805\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" -2:word.lower=voor\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 93.39%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +0.767\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word.cluster=375\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 93.40%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +0.767\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" -1:postag=Misc\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 93.42%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +0.763\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word[-2:]=es\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 93.44%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +0.760\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word[-2:]=3\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 93.44%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +0.760\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word.lower=3\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 93.44%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +0.760\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word[-3:]=3\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 93.53%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +0.745\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word.cluster=478\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 93.60%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +0.733\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word.cluster=411\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 93.69%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +0.719\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word[-2:]=ey\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 93.77%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +0.705\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word[-3:]=ola\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 93.79%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +0.702\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word[-2:]=le\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 93.79%); border: none;\">\n",
" <td colspan=\"2\" style=\"padding: 0 0.5em 0 0.5em; text-align: center; border: none; white-space: nowrap;\">\n",
" <i>… 2209 more positive …</i>\n",
" </td>\n",
" </tr>\n",
" \n",
"\n",
" \n",
" <tr style=\"background-color: hsl(0, 100.00%, 93.33%); border: none;\">\n",
" <td colspan=\"2\" style=\"padding: 0 0.5em 0 0.5em; text-align: center; border: none; white-space: nowrap;\">\n",
" <i>… 250 more negative …</i>\n",
" </td>\n",
" </tr>\n",
" \n",
" \n",
" <tr style=\"background-color: hsl(0, 100.00%, 93.33%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" -0.777\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" -2:postag=V\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(0, 100.00%, 93.24%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" -0.793\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word[-2:]=in\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(0, 100.00%, 93.06%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" -0.824\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" -1:word.lower=financiële\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(0, 100.00%, 92.95%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" -0.842\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" -1:postag=Num\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(0, 100.00%, 91.55%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" -1.090\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" 0\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(0, 100.00%, 91.16%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" -1.162\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" postag=V\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
"\n",
" </tbody>\n",
" </table>\n",
"\n",
" \n",
" \n",
" </td>\n",
" \n",
" <td style=\"padding: 0px; border: 1px solid black; vertical-align: top;\">\n",
" \n",
" \n",
" \n",
" \n",
" \n",
" <table class=\"eli5-weights\"\n",
" style=\"border-collapse: collapse; border: none; margin-top: 0em; table-layout: auto; width: 100%;\">\n",
" <thead>\n",
" <tr style=\"border: none;\">\n",
" \n",
" <th style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\" title=\"Feature weights. Note that weights do not account for feature value scales, so if feature values have different scales, features with highest weights might not be the most important.\">\n",
" Weight<sup>?</sup>\n",
" </th>\n",
" \n",
" <th style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">Feature</th>\n",
" \n",
" </tr>\n",
" </thead>\n",
" <tbody>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 80.80%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +3.523\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word.cluster=489\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 83.29%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +2.888\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word.cluster=204\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 83.58%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +2.818\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word.cluster=301\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 83.63%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +2.804\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word.cluster=3\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 83.79%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +2.765\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word.cluster=246\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 85.13%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +2.444\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word.cluster=337\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 85.24%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +2.419\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word.cluster=6\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 85.49%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +2.361\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word.cluster=326\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 86.19%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +2.199\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word.cluster=296\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 86.77%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +2.069\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word.cluster=87\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 87.20%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +1.975\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word.cluster=349\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 88.34%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +1.727\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word.cluster=12\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 89.07%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +1.575\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word.cluster=105\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 89.10%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +1.569\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word.cluster=350\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 89.69%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +1.448\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word.lower=bode\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 89.74%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +1.439\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word.cluster=85\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 90.20%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +1.347\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word[-2:]=ov\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 90.24%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +1.340\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" +1:word.lower=(\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 90.27%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +1.333\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" -1:word.lower=-\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 90.34%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +1.320\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word.cluster=111\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 90.41%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +1.307\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" +2:word.lower=--\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 90.44%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +1.301\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word[-2:]=ff\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 90.44%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +1.300\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" -1:postag=V\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 90.71%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +1.248\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word[-2:]=án\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 90.72%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +1.247\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" -1:word.lower=volgens\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 90.80%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +1.231\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word[-3:]=par\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 90.80%); border: none;\">\n",
" <td colspan=\"2\" style=\"padding: 0 0.5em 0 0.5em; text-align: center; border: none; white-space: nowrap;\">\n",
" <i>… 7464 more positive …</i>\n",
" </td>\n",
" </tr>\n",
" \n",
"\n",
" \n",
" <tr style=\"background-color: hsl(0, 100.00%, 90.19%); border: none;\">\n",
" <td colspan=\"2\" style=\"padding: 0 0.5em 0 0.5em; text-align: center; border: none; white-space: nowrap;\">\n",
" <i>… 851 more negative …</i>\n",
" </td>\n",
" </tr>\n",
" \n",
" \n",
" <tr style=\"background-color: hsl(0, 100.00%, 90.19%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" -1.349\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word[-3:]=ing\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(0, 100.00%, 90.14%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" -1.360\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word[-2:]=ur\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(0, 100.00%, 88.90%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" -1.610\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" -1:word.lower=in\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(0, 100.00%, 87.37%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" -1.937\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" -1:postag=Art\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
"\n",
" </tbody>\n",
" </table>\n",
"\n",
" \n",
" \n",
" </td>\n",
" \n",
" <td style=\"padding: 0px; border: 1px solid black; vertical-align: top;\">\n",
" \n",
" \n",
" \n",
" \n",
" \n",
" <table class=\"eli5-weights\"\n",
" style=\"border-collapse: collapse; border: none; margin-top: 0em; table-layout: auto; width: 100%;\">\n",
" <thead>\n",
" <tr style=\"border: none;\">\n",
" \n",
" <th style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\" title=\"Feature weights. Note that weights do not account for feature value scales, so if feature values have different scales, features with highest weights might not be the most important.\">\n",
" Weight<sup>?</sup>\n",
" </th>\n",
" \n",
" <th style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">Feature</th>\n",
" \n",
" </tr>\n",
" </thead>\n",
" <tbody>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 88.24%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +1.748\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" -1:word.lower=van\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 89.81%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +1.425\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word.cluster=3\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 90.38%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +1.313\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word.cluster=388\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 90.51%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +1.287\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word.cluster=450\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 90.70%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +1.250\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word.cluster=6\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 90.80%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +1.231\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word.cluster=249\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 91.22%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +1.152\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" +1:word.lower=(\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 91.35%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +1.127\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" +2:word.lower=die\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 91.80%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +1.044\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word.lower=gucht\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 92.21%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +0.970\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word.cluster=337\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 92.46%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +0.927\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" -1:word.lower=pu\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 92.46%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +0.927\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word[-3:]=nfu\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 92.46%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +0.927\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word.lower=shenfu\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 92.47%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +0.925\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word[-2:]=fu\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 92.49%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +0.921\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word.cluster=296\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 92.64%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +0.895\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word.lower=grauwe\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 92.76%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +0.875\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word[-3:]=uwe\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 92.79%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +0.869\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word[-2:]=we\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 92.80%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +0.867\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" -1:word.lower=de\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 92.95%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +0.841\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word[-2:]=ck\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 93.11%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +0.815\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" -1:postag=Pron\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 93.11%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" +0.814\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word.lower=den\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" \n",
" <tr style=\"background-color: hsl(120, 100.00%, 93.11%); border: none;\">\n",
" <td colspan=\"2\" style=\"padding: 0 0.5em 0 0.5em; text-align: center; border: none; white-space: nowrap;\">\n",
" <i>… 4971 more positive …</i>\n",
" </td>\n",
" </tr>\n",
" \n",
"\n",
" \n",
" <tr style=\"background-color: hsl(0, 100.00%, 93.11%); border: none;\">\n",
" <td colspan=\"2\" style=\"padding: 0 0.5em 0 0.5em; text-align: center; border: none; white-space: nowrap;\">\n",
" <i>… 548 more negative …</i>\n",
" </td>\n",
" </tr>\n",
" \n",
" \n",
" <tr style=\"background-color: hsl(0, 100.00%, 93.11%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" -0.814\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word[-2:]=ma\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(0, 100.00%, 93.05%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" -0.824\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word[-3:]=aan\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(0, 100.00%, 92.90%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" -0.851\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word[-2:]=al\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(0, 100.00%, 92.26%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" -0.961\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word.cluster=238\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(0, 100.00%, 91.87%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" -1.032\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word.istitle=False\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(0, 100.00%, 91.74%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" -1.055\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" postag=V\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(0, 100.00%, 91.65%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" -1.071\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word[-3:]=ter\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
" <tr style=\"background-color: hsl(0, 100.00%, 91.61%); border: none;\">\n",
" <td style=\"padding: 0 1em 0 0.5em; text-align: right; border: none;\">\n",
" -1.079\n",
" </td>\n",
" <td style=\"padding: 0 0.5em 0 0.5em; text-align: left; border: none;\">\n",
" word[-2:]=um\n",
" </td>\n",
" \n",
"</tr>\n",
" \n",
"\n",
" </tbody>\n",
" </table>\n",
"\n",
" \n",
" \n",
" </td>\n",
" \n",
" \n",
" </tr>\n",
" </table>\n",
" \n",
"\n",
" \n",
" \n",
"\n",
"\n",
" \n",
" \n",
"\n",
"\n",
" \n",
" \n",
"\n",
"\n",
" \n",
" \n",
"\n",
"\n",
" \n",
" \n",
"\n",
"\n",
" \n",
" \n",
"\n",
"\n",
" \n",
" \n",
"\n",
"\n",
" \n",
" \n",
"\n",
"\n",
" \n",
" \n",
"\n",
"\n",
" \n",
"\n",
"\n",
" \n",
"\n",
" \n",
"\n",
" \n",
"\n",
"\n",
" \n",
"\n",
" \n",
"\n",
" \n",
"\n",
" \n",
"\n",
" \n",
"\n",
" \n",
"\n",
"\n",
" \n",
"\n",
" \n",
"\n",
" \n",
"\n",
" \n",
"\n",
" \n",
"\n",
" \n",
"\n",
"\n",
"\n"
],
"text/plain": [
"<IPython.core.display.HTML object>"
]
},
"execution_count": 11,
"metadata": {},
"output_type": "execute_result"
}
],
"source": [
"import eli5\n",
"\n",
"eli5.show_weights(crf, top=30)"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"## Finding the optimal hyperparameters"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"So far we've trained a model with the default parameters. It's unlikely that these will give us the best performance possible. Therefore we're going to search automatically for the best hyperparameter settings by iteratively training different models and evaluating them. Eventually we'll pick the best one.\n",
"\n",
"Here we'll focus on two parameters: `c1` and `c2`. These are the parameters for L1 and L2 regularization, respectively. Regularization prevents overfitting on the training data by adding a penalty to the loss function. In L1 regularization, this penalty is the sum of the absolute values of the weights; in L2 regularization, it is the sum of the squared weights. L1 regularization performs a type of feature selection, as it assigns 0 weight to irrelevant features. L2 regularization, by contrast, makes the weight of irrelevant features small, but not necessarily zero. L1 regularization is often called the Lasso method, L2 is called the Ridge method, and the linear combination of both is called Elastic Net regularization.\n",
"\n",
"We define the parameter space for c1 and c2 and use the flat F1-score to compare the individual models. We'll rely on three-fold cross validation to score each of the 50 candidates. We use a randomized search, which means we're not going to try out all specified parameter settings, but instead, we'll let the process sample randomly from the distributions we've specified in the parameter space. It will do this 50 (`n_iter`) times. This process takes a while, but it's worth the wait."
]
},
{
"cell_type": "code",
"execution_count": 12,
"metadata": {},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"Fitting 3 folds for each of 50 candidates, totalling 150 fits\n"
]
},
{
"name": "stderr",
"output_type": "stream",
"text": [
"[Parallel(n_jobs=-1)]: Using backend LokyBackend with 16 concurrent workers.\n",
"[Parallel(n_jobs=-1)]: Done 18 tasks | elapsed: 3.2min\n",
"[Parallel(n_jobs=-1)]: Done 150 out of 150 | elapsed: 24.3min finished\n"
]
},
{
"data": {
"text/plain": [
"RandomizedSearchCV(cv=3, error_score='raise-deprecating',\n",
" estimator=CRF(algorithm='lbfgs', all_possible_states=None,\n",
" all_possible_transitions=True, averaging=None, c=None, c1=None, c2=None,\n",
" calibration_candidates=None, calibration_eta=None,\n",
" calibration_max_trials=None, calibration_rate=None,\n",
" calibration_samples=None, delta=None, epsilon=None, error...e,\n",
" num_memories=None, pa_type=None, period=None, trainer_cls=None,\n",
" variance=None, verbose=False),\n",
" fit_params=None, iid='warn', n_iter=50, n_jobs=-1,\n",
" param_distributions={'c1': <scipy.stats._distn_infrastructure.rv_frozen object at 0x7f9947f04e10>, 'c2': <scipy.stats._distn_infrastructure.rv_frozen object at 0x7f9947f04c88>},\n",
" pre_dispatch='2*n_jobs', random_state=None, refit=True,\n",
" return_train_score='warn',\n",
" scoring=make_scorer(flat_f1_score, average=weighted, labels=['B-ORG', 'B-MISC', 'B-PER', 'I-PER', 'B-LOC', 'I-MISC', 'I-ORG', 'I-LOC']),\n",
" verbose=1)"
]
},
"execution_count": 12,
"metadata": {},
"output_type": "execute_result"
}
],
"source": [
"import scipy\n",
"from sklearn.metrics import make_scorer\n",
"from sklearn.model_selection import RandomizedSearchCV\n",
"\n",
"crf = crfsuite.CRF(\n",
" algorithm='lbfgs',\n",
" max_iterations=100,\n",
" all_possible_transitions=True\n",
")\n",
"\n",
"params_space = {\n",
" 'c1': scipy.stats.expon(scale=0.5),\n",
" 'c2': scipy.stats.expon(scale=0.05),\n",
"}\n",
"\n",
"f1_scorer = make_scorer(metrics.flat_f1_score,\n",
" average='weighted', labels=labels)\n",
"\n",
"rs = RandomizedSearchCV(crf, params_space,\n",
" cv=3,\n",
" verbose=1,\n",
" n_jobs=-1,\n",
" n_iter=50,\n",
" scoring=f1_scorer)\n",
"rs.fit(X_train, y_train)"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"Let's take a look at the best hyperparameter settings. Our random search suggests a combination of L1 and L2 normalization."
]
},
{
"cell_type": "code",
"execution_count": 13,
"metadata": {},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"best params: {'c1': 0.08869645933566639, 'c2': 0.005642379370340676}\n",
"best CV score: 0.7608794798691931\n",
"model size: 1.06M\n"
]
}
],
"source": [
"print('best params:', rs.best_params_)\n",
"print('best CV score:', rs.best_score_)\n",
"print('model size: {:0.2f}M'.format(rs.best_estimator_.size_ / 1000000))"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"To find out what precision, recall and F1-score this translates to, we take the best estimator from our random search and evaluate it on the test set. This indeed shows a nice improvement from our initial model. We've gone from an average F1-score of 77% to 79.1%. Both precision and recall have improved, and we see a positive result for all four entity types."
]
},
{
"cell_type": "code",
"execution_count": 14,
"metadata": {},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
" precision recall f1-score support\n",
"\n",
" B-LOC 0.849 0.863 0.856 774\n",
" I-LOC 0.359 0.571 0.441 49\n",
" B-MISC 0.847 0.622 0.717 1187\n",
" I-MISC 0.664 0.415 0.511 410\n",
" B-ORG 0.806 0.727 0.764 882\n",
" I-ORG 0.772 0.677 0.721 551\n",
" B-PER 0.834 0.903 0.867 1098\n",
" I-PER 0.892 0.958 0.924 807\n",
"\n",
" micro avg 0.823 0.761 0.791 5758\n",
" macro avg 0.753 0.717 0.725 5758\n",
"weighted avg 0.820 0.761 0.784 5758\n",
"\n"
]
}
],
"source": [
"best_crf = rs.best_estimator_\n",
"y_pred = best_crf.predict(X_test)\n",
"print(metrics.flat_classification_report(\n",
" y_test, y_pred, labels=sorted_labels, digits=3\n",
"))"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"## Conclusions"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"Conditional Random Fields have lost some of their popularity since the advent of neural-network models. Still, they can be very effective for named entity recognition, particularly when word embedding information is taken into account. "
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": []
}
],
"metadata": {
"kernelspec": {
"display_name": "Python 3",
"language": "python",
"name": "python3"
},
"language_info": {
"codemirror_mode": {
"name": "ipython",
"version": 3
},
"file_extension": ".py",
"mimetype": "text/x-python",
"name": "python",
"nbconvert_exporter": "python",
"pygments_lexer": "ipython3",
"version": "3.7.1"
}
},
"nbformat": 4,
"nbformat_minor": 2
}
이 알고리즘에 대해
Named Entity Recognition with Conditional Random Fields
One of the classic challenges of Natural Language Processing is sequence labelling. In sequence labelling, the goal is to label each word in a text with a word class. In part-of-speech tagging, these word classes are parts of speech, such as noun or verb. In named entity recognition (NER), they're types of generic named entities, such as locations, people or organizations, or more specialized entities, such as diseases or symptoms in the healthcare domain. In this way, sequence labelling can help us extract the most important information from a text and improve the performance of analytics, search or matching applications.
In this notebook we'll explore Conditional Random Fields, the most popular approach to sequence labelling before Deep Learning arrived. Deep Learning may get all the attention right now, but Conditional Random Fields are still a powerful tool to build a simple sequence labeller.
The tool we're going to use is sklearn-crfsuite. This is a wrapper around python-crfsuite, which itself is a Python binding of CRFSuite. The reason we're using sklearn-crfsuite is that it provides a number of handy utility functions, for example for evaluating the output of the model. You can install it with pip install sklearn-crfsuite.
Data
First we get some data. A well-known data set for training and testing NER models is the CoNLL-2002 data, which has Spanish and Dutch texts labelled with four types of entities: locations (LOC), persons (PER), organizations (ORG) and miscellaneous entities (MISC). Both corpora are split up in three portions: a training portion and two smaller test portions, one of which we'll use as development data. It's easy to collect the data from NLTK.
import nltk
import sklearn
from sklearn.metrics import classification_report, confusion_matrix
from sklearn.preprocessing import LabelBinarizer
import sklearn_crfsuite as crfsuite
from sklearn_crfsuite import metricstrain_sents = list(nltk.corpus.conll2002.iob_sents('ned.train'))
dev_sents = list(nltk.corpus.conll2002.iob_sents('ned.testa'))
test_sents = list(nltk.corpus.conll2002.iob_sents('ned.testb'))The data consists of a list of tokenized sentences. For each of the tokens we have the string itself, its part-of-speech tag and its entity tag, which follows the BIO convention. In the deep learning world we live in today, it's common to ignore the part-of-speech tags. However, since CRFs rely on good feature extraction, we'll gladly make use of this information. After all, the part of speech of a word tells us a lot about its possible status as a named entity: nouns will more often be entities than verbs, for example.
train_sents[0][('De', 'Art', 'O'),
('tekst', 'N', 'O'),
('van', 'Prep', 'O'),
('het', 'Art', 'O'),
('arrest', 'N', 'O'),
('is', 'V', 'O'),
('nog', 'Adv', 'O'),
('niet', 'Adv', 'O'),
('schriftelijk', 'Adj', 'O'),
('beschikbaar', 'Adj', 'O'),
('maar', 'Conj', 'O'),
('het', 'Art', 'O'),
('bericht', 'N', 'O'),
('werd', 'V', 'O'),
('alvast', 'Adv', 'O'),
('bekendgemaakt', 'V', 'O'),
('door', 'Prep', 'O'),
('een', 'Art', 'O'),
('communicatiebureau', 'N', 'O'),
('dat', 'Conj', 'O'),
('Floralux', 'N', 'B-ORG'),
('inhuurde', 'V', 'O'),
('.', 'Punc', 'O')]Feature Extraction
Whereas today neural networks are expected to learn the relevant features of the input texts themselves, this is very different with Conditional Random Fields. CRFs learn the relationship between the features we give them and the label of a token in a given context. They're not going to earn these features themselves. Instead, the quality of the model will depend highly on the relevance of the features we show it.
The most important method in this tutorial is therefore the one that collects the features for every token. What information could be useful? The word itself, of course, together with its part of speech tag. It can also be interesting to know whether the word is completely uppercase, whether it starts with a capital or is a digit. In addition, we also take a look at the character bigram and trigram the word ends with. We also give every token a bias feature, which always has the same value. This bias feature helps the CRF learn the relative frequency of each label type in the training data.
To give the CRF more information about the meaning of a word, we also introduce information from word embeddings. In our Word Embedding notebook, we trained word embeddings on Dutch Wikipedia and clustered them in 500 clusters. Here we'll read these 500 clusters from a file, and map each word to the id of the cluster it is in. This is really useful for Named Entity Recognition, as most entity types cluster together. This allows CRFs to generalize above the word level. For example, when the CRF encounters a word it has never seen (say, Albania), it can base its decision on the cluster the word is in. If this cluster contains many other entities the CRF has met in its training data (say, Italy, Germany and France), it will have learnt a string link between this cluster and a specific entity type. As a result, it can still assign that entity type to the unknown word. In our experiments, this feature alone boosts the performance with around 3%.
Finally, apart from the token itself, we also want the CRF to look at its context. More specifically, we're going to give it some extra information about the two words to the left and the right of the targt word. We'll tell the CRF what these words are, whether they start with a capital or are completely uppercase, and give it their part-of-speech tag. If there is no left or right context, we'll inform the CRF that the token is at the beginning or end of the sentence (BOS or EOS).
def read_clusters(cluster_file):
word2cluster = {}
with open(cluster_file) as i:
for line in i:
word, cluster = line.strip().split('\t')
word2cluster[word] = cluster
return word2cluster
def word2features(sent, i, word2cluster):
word = sent[i][0]
postag = sent[i][1]
features = [
'bias',
'word.lower=' + word.lower(),
'word[-3:]=' + word[-3:],
'word[-2:]=' + word[-2:],
'word.isupper=%s' % word.isupper(),
'word.istitle=%s' % word.istitle(),
'word.isdigit=%s' % word.isdigit(),
'word.cluster=%s' % word2cluster[word.lower()] if word.lower() in word2cluster else "0",
'postag=' + postag
]
if i > 0:
word1 = sent[i-1][0]
postag1 = sent[i-1][1]
features.extend([
'-1:word.lower=' + word1.lower(),
'-1:word.istitle=%s' % word1.istitle(),
'-1:word.isupper=%s' % word1.isupper(),
'-1:postag=' + postag1
])
else:
features.append('BOS')
if i > 1:
word2 = sent[i-2][0]
postag2 = sent[i-2][1]
features.extend([
'-2:word.lower=' + word2.lower(),
'-2:word.istitle=%s' % word2.istitle(),
'-2:word.isupper=%s' % word2.isupper(),
'-2:postag=' + postag2
])
if i < len(sent)-1:
word1 = sent[i+1][0]
postag1 = sent[i+1][1]
features.extend([
'+1:word.lower=' + word1.lower(),
'+1:word.istitle=%s' % word1.istitle(),
'+1:word.isupper=%s' % word1.isupper(),
'+1:postag=' + postag1
])
else:
features.append('EOS')
if i < len(sent)-2:
word2 = sent[i+2][0]
postag2 = sent[i+2][1]
features.extend([
'+2:word.lower=' + word2.lower(),
'+2:word.istitle=%s' % word2.istitle(),
'+2:word.isupper=%s' % word2.isupper(),
'+2:postag=' + postag2
])
return features
def sent2features(sent, word2cluster):
return [word2features(sent, i, word2cluster) for i in range(len(sent))]
def sent2labels(sent):
return [label for token, postag, label in sent]
def sent2tokens(sent):
return [token for token, postag, label in sent]
word2cluster = read_clusters("data/embeddings/clusters_nl.tsv")sent2features(train_sents[0], word2cluster)[0]['bias',
'word.lower=de',
'word[-3:]=De',
'word[-2:]=De',
'word.isupper=False',
'word.istitle=True',
'word.isdigit=False',
'word.cluster=38',
'postag=Art',
'BOS',
'+1:word.lower=tekst',
'+1:word.istitle=False',
'+1:word.isupper=False',
'+1:postag=N',
'+2:word.lower=van',
'+2:word.istitle=False',
'+2:word.isupper=False',
'+2:postag=Prep']X_train = [sent2features(s, word2cluster) for s in train_sents]
y_train = [sent2labels(s) for s in train_sents]
X_dev = [sent2features(s, word2cluster) for s in dev_sents]
y_dev = [sent2labels(s) for s in dev_sents]
X_test = [sent2features(s, word2cluster) for s in test_sents]
y_test = [sent2labels(s) for s in test_sents]Training
We now create a CRF model and train it. We'll use the standard L-BFGS algorithm for our parameter estimation and run it for 100 iterations. When we're done, we save the model with joblib.
crf = crfsuite.CRF(
verbose='true',
algorithm='lbfgs',
max_iterations=100
)
crf.fit(X_train, y_train, X_dev=X_dev, y_dev=y_dev)loading training data to CRFsuite: 100%|██████████| 15806/15806 [00:02&lt;00:00, 7623.17it/s]
loading dev data to CRFsuite: 27%|██▋ | 769/2895 [00:00&lt;00:00, 7689.13it/s]loading dev data to CRFsuite: 100%|██████████| 2895/2895 [00:00&lt;00:00, 7186.08it/s]
Holdout group: 2
Feature generation
type: CRF1d
feature.minfreq: 0.000000
feature.possible_states: 0
feature.possible_transitions: 0
0....1....2....3....4....5....6....7....8....9....10
Number of features: 152117
Seconds required: 0.424
L-BFGS optimization
c1: 0.000000
c2: 1.000000
num_memories: 6
max_iterations: 100
epsilon: 0.000010
stop: 10
delta: 0.000010
linesearch: MoreThuente
linesearch.max_iterations: 20
Iter 1 time=0.37 loss=104214.83 active=152117 precision=0.100 recall=0.111 F1=0.105 Acc(item/seq)=0.901 0.496 feature_norm=1.00
Iter 2 time=0.21 loss=96997.81 active=152117 precision=0.100 recall=0.111 F1=0.105 Acc(item/seq)=0.901 0.496 feature_norm=1.13
Iter 3 time=0.21 loss=92085.38 active=152117 precision=0.100 recall=0.111 F1=0.105 Acc(item/seq)=0.901 0.496 feature_norm=1.26
Iter 4 time=0.21 loss=84277.67 active=152117 precision=0.100 recall=0.111 F1=0.105 Acc(item/seq)=0.901 0.496 feature_norm=1.51
Iter 5 time=0.21 loss=67577.53 active=152117 precision=0.169 recall=0.113 F1=0.109 Acc(item/seq)=0.902 0.496 feature_norm=2.32
Iter 6 time=0.21 loss=47854.26 active=152117 precision=0.326 recall=0.347 F1=0.320 Acc(item/seq)=0.930 0.580 feature_norm=4.34
Iter 7 time=0.21 loss=43326.19 active=152117 precision=0.340 recall=0.365 F1=0.333 Acc(item/seq)=0.933 0.592 feature_norm=5.06
Iter 8 time=0.21 loss=38617.07 active=152117 precision=0.372 recall=0.399 F1=0.362 Acc(item/seq)=0.938 0.618 feature_norm=6.43
Iter 9 time=0.21 loss=35511.85 active=152117 precision=0.491 recall=0.442 F1=0.421 Acc(item/seq)=0.942 0.631 feature_norm=8.62
Iter 10 time=0.21 loss=32735.31 active=152117 precision=0.535 recall=0.456 F1=0.445 Acc(item/seq)=0.944 0.654 feature_norm=9.50
Iter 11 time=0.21 loss=31687.17 active=152117 precision=0.530 recall=0.491 F1=0.502 Acc(item/seq)=0.947 0.663 feature_norm=10.94
Iter 12 time=0.21 loss=29323.19 active=152117 precision=0.532 recall=0.482 F1=0.477 Acc(item/seq)=0.946 0.666 feature_norm=11.18
Iter 13 time=0.21 loss=28733.55 active=152117 precision=0.596 recall=0.489 F1=0.489 Acc(item/seq)=0.947 0.668 feature_norm=11.58
Iter 14 time=0.21 loss=27120.69 active=152117 precision=0.640 recall=0.507 F1=0.512 Acc(item/seq)=0.948 0.677 feature_norm=12.36
Iter 15 time=0.21 loss=24849.05 active=152117 precision=0.640 recall=0.547 F1=0.558 Acc(item/seq)=0.952 0.697 feature_norm=13.86
Iter 16 time=0.40 loss=24033.40 active=152117 precision=0.654 recall=0.580 F1=0.586 Acc(item/seq)=0.954 0.706 feature_norm=14.53
Iter 17 time=0.21 loss=22935.94 active=152117 precision=0.669 recall=0.578 F1=0.598 Acc(item/seq)=0.955 0.712 feature_norm=15.15
Iter 18 time=0.21 loss=21803.53 active=152117 precision=0.682 recall=0.584 F1=0.605 Acc(item/seq)=0.956 0.713 feature_norm=15.67
Iter 19 time=0.21 loss=21046.75 active=152117 precision=0.725 recall=0.565 F1=0.586 Acc(item/seq)=0.956 0.717 feature_norm=16.04
Iter 20 time=0.21 loss=20465.96 active=152117 precision=0.701 recall=0.566 F1=0.594 Acc(item/seq)=0.956 0.711 feature_norm=15.81
Iter 21 time=0.21 loss=19991.29 active=152117 precision=0.706 recall=0.586 F1=0.611 Acc(item/seq)=0.957 0.717 feature_norm=15.63
Iter 22 time=0.21 loss=19560.23 active=152117 precision=0.684 recall=0.597 F1=0.616 Acc(item/seq)=0.957 0.722 feature_norm=15.58
Iter 23 time=0.21 loss=19241.14 active=152117 precision=0.672 recall=0.602 F1=0.615 Acc(item/seq)=0.957 0.726 feature_norm=15.65
Iter 24 time=0.21 loss=18787.87 active=152117 precision=0.678 recall=0.627 F1=0.637 Acc(item/seq)=0.958 0.731 feature_norm=16.31
Iter 25 time=0.21 loss=18145.13 active=152117 precision=0.690 recall=0.625 F1=0.640 Acc(item/seq)=0.959 0.734 feature_norm=16.94
Iter 26 time=0.21 loss=17786.38 active=152117 precision=0.710 recall=0.621 F1=0.642 Acc(item/seq)=0.959 0.738 feature_norm=17.48
Iter 27 time=0.21 loss=17247.02 active=152117 precision=0.711 recall=0.625 F1=0.649 Acc(item/seq)=0.959 0.736 feature_norm=18.46
Iter 28 time=0.21 loss=16876.01 active=152117 precision=0.737 recall=0.627 F1=0.655 Acc(item/seq)=0.960 0.748 feature_norm=19.83
Iter 29 time=0.21 loss=16543.87 active=152117 precision=0.732 recall=0.631 F1=0.663 Acc(item/seq)=0.961 0.751 feature_norm=20.12
Iter 30 time=0.21 loss=16263.21 active=152117 precision=0.725 recall=0.644 F1=0.671 Acc(item/seq)=0.962 0.753 feature_norm=20.42
Iter 31 time=0.21 loss=15665.78 active=152117 precision=0.715 recall=0.661 F1=0.676 Acc(item/seq)=0.963 0.758 feature_norm=21.50
Iter 32 time=0.21 loss=15247.34 active=152117 precision=0.700 recall=0.641 F1=0.650 Acc(item/seq)=0.961 0.748 feature_norm=23.18
Iter 33 time=0.21 loss=14866.51 active=152117 precision=0.702 recall=0.658 F1=0.670 Acc(item/seq)=0.963 0.754 feature_norm=24.67
Iter 34 time=0.21 loss=14650.96 active=152117 precision=0.704 recall=0.659 F1=0.675 Acc(item/seq)=0.963 0.755 feature_norm=25.38
Iter 35 time=0.21 loss=14386.98 active=152117 precision=0.730 recall=0.677 F1=0.695 Acc(item/seq)=0.964 0.761 feature_norm=26.47
Iter 36 time=0.21 loss=14158.49 active=152117 precision=0.742 recall=0.681 F1=0.704 Acc(item/seq)=0.965 0.763 feature_norm=28.58
Iter 37 time=0.21 loss=13895.30 active=152117 precision=0.736 recall=0.684 F1=0.701 Acc(item/seq)=0.965 0.765 feature_norm=28.72
Iter 38 time=0.21 loss=13656.45 active=152117 precision=0.730 recall=0.683 F1=0.695 Acc(item/seq)=0.965 0.768 feature_norm=29.07
Iter 39 time=0.21 loss=13499.28 active=152117 precision=0.727 recall=0.680 F1=0.691 Acc(item/seq)=0.965 0.769 feature_norm=29.61
Iter 40 time=0.21 loss=13174.95 active=152117 precision=0.726 recall=0.677 F1=0.689 Acc(item/seq)=0.965 0.766 feature_norm=31.03
Iter 41 time=0.21 loss=13104.00 active=152117 precision=0.736 recall=0.662 F1=0.678 Acc(item/seq)=0.964 0.760 feature_norm=33.06
Iter 42 time=0.21 loss=12750.79 active=152117 precision=0.731 recall=0.685 F1=0.701 Acc(item/seq)=0.966 0.764 feature_norm=34.35
Iter 43 time=0.21 loss=12637.02 active=152117 precision=0.740 recall=0.690 F1=0.708 Acc(item/seq)=0.966 0.766 feature_norm=34.63
Iter 44 time=0.21 loss=12534.20 active=152117 precision=0.745 recall=0.692 F1=0.712 Acc(item/seq)=0.966 0.766 feature_norm=35.30
Iter 45 time=0.21 loss=12390.89 active=152117 precision=0.739 recall=0.682 F1=0.702 Acc(item/seq)=0.966 0.761 feature_norm=37.15
Iter 46 time=0.21 loss=12277.42 active=152117 precision=0.733 recall=0.689 F1=0.704 Acc(item/seq)=0.966 0.763 feature_norm=37.58
Iter 47 time=0.21 loss=12219.03 active=152117 precision=0.739 recall=0.690 F1=0.706 Acc(item/seq)=0.966 0.767 feature_norm=37.20
Iter 48 time=0.21 loss=12125.64 active=152117 precision=0.743 recall=0.695 F1=0.711 Acc(item/seq)=0.967 0.770 feature_norm=36.99
Iter 49 time=0.21 loss=11970.65 active=152117 precision=0.745 recall=0.697 F1=0.712 Acc(item/seq)=0.967 0.775 feature_norm=36.84
Iter 50 time=0.21 loss=11780.73 active=152117 precision=0.754 recall=0.696 F1=0.718 Acc(item/seq)=0.968 0.777 feature_norm=37.57
Iter 51 time=0.21 loss=11623.83 active=152117 precision=0.740 recall=0.691 F1=0.710 Acc(item/seq)=0.967 0.774 feature_norm=38.21
Iter 52 time=0.21 loss=11549.38 active=152117 precision=0.740 recall=0.690 F1=0.709 Acc(item/seq)=0.967 0.773 feature_norm=38.94
Iter 53 time=0.21 loss=11497.85 active=152117 precision=0.739 recall=0.696 F1=0.713 Acc(item/seq)=0.967 0.772 feature_norm=39.54
Iter 54 time=0.21 loss=11419.64 active=152117 precision=0.733 recall=0.692 F1=0.707 Acc(item/seq)=0.966 0.773 feature_norm=40.40
Iter 55 time=0.21 loss=11280.29 active=152117 precision=0.744 recall=0.707 F1=0.719 Acc(item/seq)=0.967 0.773 feature_norm=41.75
Iter 56 time=0.21 loss=11131.39 active=152117 precision=0.746 recall=0.710 F1=0.722 Acc(item/seq)=0.968 0.773 feature_norm=42.72
Iter 57 time=0.21 loss=11043.40 active=152117 precision=0.751 recall=0.713 F1=0.726 Acc(item/seq)=0.968 0.774 feature_norm=42.92
Iter 58 time=0.21 loss=10954.38 active=152117 precision=0.769 recall=0.713 F1=0.736 Acc(item/seq)=0.969 0.781 feature_norm=43.18
Iter 59 time=0.21 loss=10836.31 active=152117 precision=0.773 recall=0.713 F1=0.736 Acc(item/seq)=0.969 0.781 feature_norm=43.74
Iter 60 time=0.21 loss=10712.24 active=152117 precision=0.779 recall=0.719 F1=0.744 Acc(item/seq)=0.970 0.788 feature_norm=44.41
Iter 61 time=0.21 loss=10602.81 active=152117 precision=0.789 recall=0.709 F1=0.740 Acc(item/seq)=0.970 0.789 feature_norm=44.68
Iter 62 time=0.21 loss=10508.84 active=152117 precision=0.782 recall=0.711 F1=0.739 Acc(item/seq)=0.970 0.787 feature_norm=45.37
Iter 63 time=0.21 loss=10458.88 active=152117 precision=0.783 recall=0.717 F1=0.744 Acc(item/seq)=0.970 0.788 feature_norm=45.51
Iter 64 time=0.21 loss=10420.78 active=152117 precision=0.763 recall=0.711 F1=0.730 Acc(item/seq)=0.969 0.786 feature_norm=45.67
Iter 65 time=0.21 loss=10315.28 active=152117 precision=0.766 recall=0.721 F1=0.735 Acc(item/seq)=0.969 0.786 feature_norm=46.30
Iter 66 time=0.21 loss=10204.10 active=152117 precision=0.769 recall=0.728 F1=0.740 Acc(item/seq)=0.970 0.786 feature_norm=47.10
Iter 67 time=0.21 loss=10134.54 active=152117 precision=0.769 recall=0.716 F1=0.737 Acc(item/seq)=0.970 0.787 feature_norm=47.87
Iter 68 time=0.21 loss=10095.10 active=152117 precision=0.773 recall=0.718 F1=0.741 Acc(item/seq)=0.970 0.787 feature_norm=47.85
Iter 69 time=0.21 loss=10059.52 active=152117 precision=0.773 recall=0.715 F1=0.738 Acc(item/seq)=0.970 0.790 feature_norm=47.81
Iter 70 time=0.21 loss=10012.53 active=152117 precision=0.767 recall=0.712 F1=0.733 Acc(item/seq)=0.970 0.790 feature_norm=47.95
Iter 71 time=0.21 loss=9931.41 active=152117 precision=0.765 recall=0.710 F1=0.731 Acc(item/seq)=0.970 0.791 feature_norm=48.49
Iter 72 time=0.21 loss=9861.45 active=152117 precision=0.763 recall=0.709 F1=0.730 Acc(item/seq)=0.970 0.793 feature_norm=49.07
Iter 73 time=0.21 loss=9803.75 active=152117 precision=0.769 recall=0.723 F1=0.741 Acc(item/seq)=0.970 0.796 feature_norm=49.32
Iter 74 time=0.21 loss=9762.61 active=152117 precision=0.758 recall=0.720 F1=0.733 Acc(item/seq)=0.970 0.793 feature_norm=49.58
Iter 75 time=0.21 loss=9726.47 active=152117 precision=0.761 recall=0.722 F1=0.736 Acc(item/seq)=0.970 0.790 feature_norm=49.80
Iter 76 time=0.21 loss=9628.97 active=152117 precision=0.764 recall=0.722 F1=0.736 Acc(item/seq)=0.970 0.789 feature_norm=50.47
Iter 77 time=0.40 loss=9586.61 active=152117 precision=0.763 recall=0.725 F1=0.740 Acc(item/seq)=0.971 0.791 feature_norm=50.96
Iter 78 time=0.21 loss=9522.00 active=152117 precision=0.767 recall=0.723 F1=0.741 Acc(item/seq)=0.970 0.788 feature_norm=51.34
Iter 79 time=0.21 loss=9479.87 active=152117 precision=0.765 recall=0.720 F1=0.736 Acc(item/seq)=0.970 0.788 feature_norm=51.77
Iter 80 time=0.21 loss=9448.33 active=152117 precision=0.771 recall=0.720 F1=0.739 Acc(item/seq)=0.970 0.789 feature_norm=51.80
Iter 81 time=0.21 loss=9423.55 active=152117 precision=0.769 recall=0.723 F1=0.740 Acc(item/seq)=0.970 0.789 feature_norm=51.83
Iter 82 time=0.21 loss=9367.51 active=152117 precision=0.763 recall=0.720 F1=0.736 Acc(item/seq)=0.970 0.790 feature_norm=52.25
Iter 83 time=0.21 loss=9322.98 active=152117 precision=0.762 recall=0.722 F1=0.737 Acc(item/seq)=0.970 0.793 feature_norm=52.34
Iter 84 time=0.21 loss=9277.51 active=152117 precision=0.765 recall=0.725 F1=0.740 Acc(item/seq)=0.971 0.798 feature_norm=52.47
Iter 85 time=0.21 loss=9229.07 active=152117 precision=0.772 recall=0.726 F1=0.743 Acc(item/seq)=0.971 0.798 feature_norm=52.79
Iter 86 time=0.21 loss=9214.50 active=152117 precision=0.782 recall=0.729 F1=0.751 Acc(item/seq)=0.971 0.799 feature_norm=52.88
Iter 87 time=0.21 loss=9194.55 active=152117 precision=0.785 recall=0.731 F1=0.753 Acc(item/seq)=0.971 0.800 feature_norm=52.82
Iter 88 time=0.21 loss=9182.23 active=152117 precision=0.783 recall=0.728 F1=0.750 Acc(item/seq)=0.971 0.798 feature_norm=52.86
Iter 89 time=0.21 loss=9161.54 active=152117 precision=0.786 recall=0.726 F1=0.750 Acc(item/seq)=0.971 0.795 feature_norm=52.98
Iter 90 time=0.21 loss=9130.06 active=152117 precision=0.788 recall=0.728 F1=0.752 Acc(item/seq)=0.971 0.797 feature_norm=53.17
Iter 91 time=0.21 loss=9063.09 active=152117 precision=0.794 recall=0.738 F1=0.760 Acc(item/seq)=0.972 0.800 feature_norm=53.64
Iter 92 time=0.40 loss=9041.98 active=152117 precision=0.791 recall=0.739 F1=0.759 Acc(item/seq)=0.972 0.801 feature_norm=53.77
Iter 93 time=0.21 loss=9010.27 active=152117 precision=0.783 recall=0.736 F1=0.754 Acc(item/seq)=0.972 0.801 feature_norm=53.97
Iter 94 time=0.21 loss=8984.80 active=152117 precision=0.786 recall=0.740 F1=0.759 Acc(item/seq)=0.972 0.803 feature_norm=54.09
Iter 95 time=0.21 loss=8965.14 active=152117 precision=0.788 recall=0.740 F1=0.760 Acc(item/seq)=0.972 0.803 feature_norm=54.14
Iter 96 time=0.21 loss=8931.00 active=152117 precision=0.786 recall=0.737 F1=0.758 Acc(item/seq)=0.972 0.804 feature_norm=54.26
Iter 97 time=0.21 loss=8923.90 active=152117 precision=0.791 recall=0.737 F1=0.757 Acc(item/seq)=0.971 0.799 feature_norm=54.68
Iter 98 time=0.21 loss=8860.78 active=152117 precision=0.784 recall=0.735 F1=0.755 Acc(item/seq)=0.971 0.802 feature_norm=54.64
Iter 99 time=0.21 loss=8846.13 active=152117 precision=0.788 recall=0.737 F1=0.758 Acc(item/seq)=0.972 0.803 feature_norm=54.68
Iter 100 time=0.21 loss=8832.34 active=152117 precision=0.789 recall=0.737 F1=0.758 Acc(item/seq)=0.972 0.803 feature_norm=54.77
================================================
Label Precision Recall F1 Support
------- ----------- -------- ----- ---------
B-LOC 0.823 0.823 0.823 479
B-MISC 0.806 0.651 0.720 748
B-ORG 0.853 0.620 0.718 686
B-PER 0.752 0.855 0.800 703
I-LOC 0.583 0.547 0.565 64
I-MISC 0.645 0.507 0.568 215
I-ORG 0.806 0.684 0.740 396
I-PER 0.844 0.946 0.892 423
O 0.990 0.998 0.994 33973
------------------------------------------------
L-BFGS terminated with the maximum number of iterations
Total seconds required for training: 21.768
Storing the model
Number of active features: 152117 (152117)
Number of active attributes: 130306 (145602)
Number of active labels: 9 (9)
Writing labels
Writing attributes
Writing feature references for transitions
Writing feature references for attributes
Seconds required: 0.058
CRF(algorithm='lbfgs', all_possible_states=None,
all_possible_transitions=None, averaging=None, c=None, c1=None, c2=None,
calibration_candidates=None, calibration_eta=None,
calibration_max_trials=None, calibration_rate=None,
calibration_samples=None, delta=None, epsilon=None, error_sensitive=None,
gamma=None, keep_tempfiles=None, linesearch=None, max_iterations=100,
max_linesearch=None, min_freq=None, model_filename=None,
num_memories=None, pa_type=None, period=None, trainer_cls=None,
variance=None, verbose='true')import joblib
import os
OUTPUT_PATH = "models/ner/"
OUTPUT_FILE = "crf_model"
if not os.path.exists(OUTPUT_PATH):
os.mkdir(OUTPUT_PATH)
joblib.dump(crf, os.path.join(OUTPUT_PATH, OUTPUT_FILE))['models/ner/crf_model']Evaluation
Let's evaluate the output of our CRF. We'll load the model from the output file above and have it predict labels for the full test set.
As a sanity check, let's take a look at its predictions for the first test sentence. This output looks pretty good: the CRF is able to predict all four locations in the sentence correctly. It only misses the person entity, which is a strange case anyway, because it is not actually a person name.
crf = joblib.load(os.path.join(OUTPUT_PATH, OUTPUT_FILE))
y_pred = crf.predict(X_test)
example_sent = test_sents[0]
print("Sentence:", ' '.join(sent2tokens(example_sent)))
print("Predicted:", ' '.join(crf.predict([sent2features(example_sent, word2cluster)])[0]))
print("Correct: ", ' '.join(sent2labels(example_sent)))Sentence: Dat is in Italië , Spanje of Engeland misschien geen probleem , maar volgens ' Der Kaiser ' in Duitsland wel .
Predicted: O O O B-LOC O B-LOC O B-LOC O O O O O O O B-MISC I-MISC O O B-LOC O O
Correct: O O O B-LOC O B-LOC O B-LOC O O O O O O O B-PER I-PER O O B-LOC O O
Now we evaluate on the full test set. We'll print out a classification report for all labels except O. If we were to include O, which far outnumbers the entity labels in our data, the average scores would be inflated artificially, simply because there's an inherently high probability that the O labels from our CRF are correct. We obtain an average F-score of 77% (micro average) across all entity types, with particularly good results for B-LOCand B-PER.
labels = list(crf.classes_)
labels.remove("O")
y_pred = crf.predict(X_test)
sorted_labels = sorted(
labels,
key=lambda name: (name[1:], name[0])
)
print(metrics.flat_classification_report(y_test, y_pred, labels=sorted_labels)) precision recall f1-score support
B-LOC 0.83 0.83 0.83 774
I-LOC 0.29 0.41 0.34 49
B-MISC 0.84 0.61 0.71 1187
I-MISC 0.59 0.42 0.49 410
B-ORG 0.80 0.69 0.74 882
I-ORG 0.74 0.66 0.70 551
B-PER 0.80 0.90 0.85 1098
I-PER 0.87 0.95 0.91 807
micro avg 0.80 0.74 0.77 5758
macro avg 0.72 0.68 0.70 5758
weighted avg 0.80 0.74 0.76 5758
Now we can also look at the most likely transitions the CRF has identified, and at the top features for every label. We'll do this with the eli5 library, which helps us explain the predictions of machine learning models.
The top transitions are quite intuitive: the most likely transitions are those within the same entity type (from a B-label to an O-label), and those where a B-label follows an O-label.
The features, too, make sense. For example, if a word does not start with an uppercase letter, it is unlikely to be an entity. By contrast, a word is very likely to be a location if it ends in ië, which is indeed a very common suffix for locations in Dutch. Notice also how informative the embedding clusters are: for all entity types, the word clusters form some of the most informative features for the CRF.
import eli5
eli5.show_weights(crf, top=30)| From \ To | O | B-LOC | I-LOC | B-MISC | I-MISC | B-ORG | I-ORG | B-PER | I-PER |
|---|---|---|---|---|---|---|---|---|---|
| O | 4.141 | 4.583 | 0.0 | 4.141 | 0.0 | 4.366 | 0.0 | 3.819 | 0.0 |
| B-LOC | -0.248 | -0.279 | 7.101 | 0.0 | 0.0 | 0.0 | 0.0 | -0.661 | 0.0 |
| I-LOC | -1.062 | -0.235 | 5.967 | 0.0 | 0.0 | 0.0 | 0.0 | 0.0 | 0.0 |
| B-MISC | -0.985 | 0.655 | 0.0 | -0.316 | 7.73 | 0.551 | 0.0 | 0.46 | 0.0 |
| I-MISC | -1.781 | 0.0 | 0.0 | -0.382 | 7.769 | 1.145 | 0.0 | -0.719 | 0.0 |
| B-ORG | -0.261 | 0.0 | 0.0 | -0.809 | 0.0 | 0.0 | 7.803 | 0.106 | 0.0 |
| I-ORG | -0.794 | 0.0 | 0.0 | 0.0 | 0.0 | 0.0 | 7.174 | 0.084 | 0.0 |
| B-PER | 0.31 | -0.346 | 0.0 | -0.611 | 0.0 | 0.0 | 0.0 | -1.408 | 8.68 |
| I-PER | 0.104 | 0.0 | 0.0 | 0.0 | 0.0 | 0.0 | 0.0 | 0.0 | 6.804 |
| y=O top features | y=B-LOC top features | y=I-LOC top features | y=B-MISC top features | y=I-MISC top features | y=B-ORG top features | y=I-ORG top features | y=B-PER top features | y=I-PER top features | ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
|
|
|
|
|
|
|
|
|
||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
Finding the optimal hyperparameters
So far we've trained a model with the default parameters. It's unlikely that these will give us the best performance possible. Therefore we're going to search automatically for the best hyperparameter settings by iteratively training different models and evaluating them. Eventually we'll pick the best one.
Here we'll focus on two parameters: c1 and c2. These are the parameters for L1 and L2 regularization, respectively. Regularization prevents overfitting on the training data by adding a penalty to the loss function. In L1 regularization, this penalty is the sum of the absolute values of the weights; in L2 regularization, it is the sum of the squared weights. L1 regularization performs a type of feature selection, as it assigns 0 weight to irrelevant features. L2 regularization, by contrast, makes the weight of irrelevant features small, but not necessarily zero. L1 regularization is often called the Lasso method, L2 is called the Ridge method, and the linear combination of both is called Elastic Net regularization.
We define the parameter space for c1 and c2 and use the flat F1-score to compare the individual models. We'll rely on three-fold cross validation to score each of the 50 candidates. We use a randomized search, which means we're not going to try out all specified parameter settings, but instead, we'll let the process sample randomly from the distributions we've specified in the parameter space. It will do this 50 (n_iter) times. This process takes a while, but it's worth the wait.
import scipy
from sklearn.metrics import make_scorer
from sklearn.model_selection import RandomizedSearchCV
crf = crfsuite.CRF(
algorithm='lbfgs',
max_iterations=100,
all_possible_transitions=True
)
params_space = {
'c1': scipy.stats.expon(scale=0.5),
'c2': scipy.stats.expon(scale=0.05),
}
f1_scorer = make_scorer(metrics.flat_f1_score,
average='weighted', labels=labels)
rs = RandomizedSearchCV(crf, params_space,
cv=3,
verbose=1,
n_jobs=-1,
n_iter=50,
scoring=f1_scorer)
rs.fit(X_train, y_train)Fitting 3 folds for each of 50 candidates, totalling 150 fits
[Parallel(n_jobs=-1)]: Using backend LokyBackend with 16 concurrent workers.
[Parallel(n_jobs=-1)]: Done 18 tasks | elapsed: 3.2min
[Parallel(n_jobs=-1)]: Done 150 out of 150 | elapsed: 24.3min finished
RandomizedSearchCV(cv=3, error_score='raise-deprecating',
estimator=CRF(algorithm='lbfgs', all_possible_states=None,
all_possible_transitions=True, averaging=None, c=None, c1=None, c2=None,
calibration_candidates=None, calibration_eta=None,
calibration_max_trials=None, calibration_rate=None,
calibration_samples=None, delta=None, epsilon=None, error...e,
num_memories=None, pa_type=None, period=None, trainer_cls=None,
variance=None, verbose=False),
fit_params=None, iid='warn', n_iter=50, n_jobs=-1,
param_distributions={'c1': <scipy.stats._distn_infrastructure.rv_frozen object at 0x7f9947f04e10>, 'c2': <scipy.stats._distn_infrastructure.rv_frozen object at 0x7f9947f04c88>},
pre_dispatch='2*n_jobs', random_state=None, refit=True,
return_train_score='warn',
scoring=make_scorer(flat_f1_score, average=weighted, labels=['B-ORG', 'B-MISC', 'B-PER', 'I-PER', 'B-LOC', 'I-MISC', 'I-ORG', 'I-LOC']),
verbose=1)Let's take a look at the best hyperparameter settings. Our random search suggests a combination of L1 and L2 normalization.
print('best params:', rs.best_params_)
print('best CV score:', rs.best_score_)
print('model size: {:0.2f}M'.format(rs.best_estimator_.size_ / 1000000))best params: {'c1': 0.08869645933566639, 'c2': 0.005642379370340676}
best CV score: 0.7608794798691931
model size: 1.06M
To find out what precision, recall and F1-score this translates to, we take the best estimator from our random search and evaluate it on the test set. This indeed shows a nice improvement from our initial model. We've gone from an average F1-score of 77% to 79.1%. Both precision and recall have improved, and we see a positive result for all four entity types.
best_crf = rs.best_estimator_
y_pred = best_crf.predict(X_test)
print(metrics.flat_classification_report(
y_test, y_pred, labels=sorted_labels, digits=3
)) precision recall f1-score support
B-LOC 0.849 0.863 0.856 774
I-LOC 0.359 0.571 0.441 49
B-MISC 0.847 0.622 0.717 1187
I-MISC 0.664 0.415 0.511 410
B-ORG 0.806 0.727 0.764 882
I-ORG 0.772 0.677 0.721 551
B-PER 0.834 0.903 0.867 1098
I-PER 0.892 0.958 0.924 807
micro avg 0.823 0.761 0.791 5758
macro avg 0.753 0.717 0.725 5758
weighted avg 0.820 0.761 0.784 5758
Conclusions
Conditional Random Fields have lost some of their popularity since the advent of neural-network models. Still, they can be very effective for named entity recognition, particularly when word embedding information is taken into account.
