Data-driven methods for anaphora resolution of Russian texts
The paper considers two data-driven methods for anaphora resolution of Russian texts. These methods are based on machine learning with annotated corpora and using no additional information except linguistic features. The first method uses Support Vector Machine as learning and classifying algorithms, the second method uses Decision Tree inducer. We evaluate the performance of the methods with several feature sets and corpora. Feature sets included morphological, syntactic and semantic features. In this paper we also evaluate how semantic features, namely semantic roles, impact the performance of anaphora resolution in Russian. We used our manually annotated corpus as well as a corpus provided by the organizing committee of the forum for the evaluation of linguistic text analysis systems, an event of Dialogue 2014. Experiments showed that precision of SVM is higher on experimental data for almost all cases. It was shown that semantic features enhance the performance of the methods for anaphora resolution of Russian texts. We have also calculated the optimal distance between the anaphor and the hypothetic antecedent and used it in our methods.