21.6.3 A Log-Linear model for Pronominal Anaphora Resolution

As our final model of pronominal anaphora resolution, we present a simple supervised machine learning approach, in which we train a log-linear classifier on a corpus in which the antecedents are marked for each pronoun. Any supervised classifier can be used for this purpose; log-linear models are popular, but Naive Bayes and other classifiers have been used as well.

For training, the system relies on a hand-labeled corpus in which each pronoun has been linked by hand with the correct antecedent. The system needs to extract positive and negative examples of anaphoric relations. Positive examples occur directly in the

training set. Negative examples can be found by pairing each pronoun with some other noun phrase. Features (discussed in the next section) are extracted for each training observation, and a classifier is trained to predict 1 for the true pronoun-antecedent pairs, and 0 for the incorrect pronoun-antecedent pairs.

For testing, just as we saw with as with the Hobbs and Centering classifiers, the log-linear classifier takes as input a pronoun (he, him, his, she, her, it, they, them, their), together with the current and preceding sentences.

In order to deal with non-referential pronouns, we first filter out pleonastic pronouns (like the pleonastic it is raining), using hand-written rules based on frequent lexical patterns.

The classifier then extracts all potential antecedents by doing a parse of the current and previous sentences, either using a full parser or a simple chunker. Next, each NP in the parse is considered a potential antecedent for each following pronoun. Each pronoun-potential antecedent pair is then presented to the classifier.