📄 Comparison of Classification and Ranking Approaches to Pronominal Anaphora Resolution in Czech

Summary

The paper addresses resolving third-person personal and possessive pronouns in Czech, including zero (pro-dropped) pronouns that are reconstructed in the deep-syntactic layer. Using the manually annotated textual coreference in the Prague Dependency Treebank 2.0, the authors build two supervised systems that share the same preprocessing and a rich feature set. That feature set draws on all annotation levels, on ideas from Lappin and Leass, on Mitkov's approach, on topic-focus articulation, and on a collocation feature computed from the Czech National Corpus.

The first system is a C5.0 decision-tree classifier that judges each anaphor–candidate pair independently. The second is a ranker whose scoring weights are learned with a Collins-style perceptron and which compares all candidates for an anaphor jointly. On the evaluation test set, the ranker reaches an F-score of 79.43%. This beats the classifier, a rule-based baseline and all previously published Czech systems evaluated on PDT data.

The results add evidence from a new language and dataset that ranking suits anaphora resolution better than pairwise classification. The comparison with English work is only approximate, however, because this system relies on gold-standard annotation.

Key points

• Data: PDT 2.0 contains 3,168 newspaper texts (49,431 sentences) and 45,174 coreference links. Personal (including zero) and possessive pronouns account for 37.4% of these links (16,888).
• Candidate selection: candidates are semantic nouns that either precede the anaphor in the same sentence or appear in the previous sentence. This restriction loses the correct antecedent for only about 6% of anaphors.
• F-scores:
  • rule baseline BASE 3+2+1: 62.79%
  • CLASS classifier: 70.17%
  • CLASS+3+2+1 (classifier with rule fallback): 76.30%
  • RANK perceptron ranker: 79.43%
• Earlier Czech systems: a rule-based filter system by Kučová & Žabokrtský (2005) reached 60.4%, and the rule-based system of Nguy & Žabokrtský (2007) reached 74.2%.
• English comparison: the ranker comes close to the 82.2% F-score that Denis & Baldridge (2008) report for English third-person pronouns, though the tasks and input annotations differ.