哪些特征可以帮助对句尾进行分类？ 序列分类

Question

Problem:

I have pairs of sentences that lack a period and a capitalized letter in between them. Need to segment them from each other. I'm looking for some help in picking the good features to improve the model.

Background:

I'm using pycrfsuite to perform sequence classification and find the end of the first sentence, like so:

From brown corpus, I join every two sentences together and get their pos tags. Then, I label every token in the sentence with 'S' if the space follows it and 'P' if the period follows it in the sentence. Then I delete a period between the sentences, and lower the following token. I get something like this:

Input:

data = ['I love Harry Potter.', 'It is my favorite book.']

Output:

sent = [('I', 'PRP'), ('love', 'VBP'), ('Harry', 'NNP'), ('Potter', 'NNP'), ('it', 'PRP'), ('is', 'VBZ'), ('my', 'PRP$'), ('favorite', 'JJ'), ('book', 'NN')]
labels = ['S', 'S', 'S', 'P', 'S', 'S', 'S', 'S', 'S']

At the moment, I extract these general features:

def word2features2(sent, i):
    word = sent[i][0]
    postag = sent[i][1]

    # Common features for all words
    features = [
        'bias',
        'word.lower=' + word.lower(),
        'word[-3:]=' + word[-3:],
        'word[-2:]=' + word[-2:],
        'word.isupper=%s' % word.isupper(),
        'word.isdigit=%s' % word.isdigit(),
        'postag=' + postag
    ]

    # Features for words that are not
    # at the beginning of a document
    if i > 0:
        word1 = sent[i-1][0]
        postag1 = sent[i-1][1]
        features.extend([
            '-1:word.lower=' + word1.lower(),
            '-1:word.isupper=%s' % word1.isupper(),
            '-1:word.isdigit=%s' % word1.isdigit(),
            '-1:postag=' + postag1
        ])
    else:
        # Indicate that it is the 'beginning of a sentence'
        features.append('BOS')

    # Features for words that are not
    # at the end of a document
    if i < len(sent)-1:
        word1 = sent[i+1][0]
        postag1 = sent[i+1][1]
        features.extend([
            '+1:word.lower=' + word1.lower(),
            '+1:word.isupper=%s' % word1.isupper(),
            '+1:word.isdigit=%s' % word1.isdigit(),
            '+1:postag=' + postag1
        ])
    else:
        # Indicate that it is the 'end of a sentence'
        features.append('EOS')

And train crf with these parameters:

    trainer = pycrfsuite.Trainer(verbose=True)

    # Submit training data to the trainer
    for xseq, yseq in zip(X_train, y_train):
        trainer.append(xseq, yseq)

    # Set the parameters of the model
    trainer.set_params({
        # coefficient for L1 penalty
        'c1': 0.1,

        # coefficient for L2 penalty
        'c2': 0.01,

        # maximum number of iterations
        'max_iterations': 200,

        # whether to include transitions that
        # are possible, but not observed
        'feature.possible_transitions': True
    })

    trainer.train('crf.model')

Results:

Accuracy report shows:

              precision    recall  f1-score   support

           S       0.99      1.00      0.99    214627
           P       0.81      0.57      0.67      5734

   micro avg       0.99      0.99      0.99    220361
   macro avg       0.90      0.79      0.83    220361
weighted avg       0.98      0.99      0.98    220361

What are some ways I could edit word2features2() in order to improve the model? (or any other part)

Here is the link to the full code as it is today.

Also, I am just a beginner in nlp so I would greatly appreciate any feedback overall, links to relevant or helpful sources, and rather simple explanations. Thank you very-very much!

Answer 1

Since your classes are very imbalanced due to the nature of the problem, I would suggest using weighted loss, where the loss for the P tag is given a higher value than those of the S class. I think the problem might be that due to the equivalent weight of both classes, the classifier not give enough attention to those P tags since their effect on the loss is very small.

Another thing you could try is hyper-parameter tuning, make sure to optimize for the macro f1-score then, since it will give equal weights to both classes regardless of the number of support instances.