📄 Discriminative Approach to Predicate-Argument Structure Analysis with Zero-Anaphora Resolution

Summary

Japanese sentences often leave arguments unexpressed as zero pronouns. Purely syntactic, sentence-level analysis therefore cannot recover full predicate-argument structures. The authors combine predicate-argument structure analysis and zero-anaphora resolution into a single supervised selection task. For each predicate, the candidate pool contains noun phrases from the predicate's own sentence, selected phrases from earlier sentences, and a special NULL option for exophoric or absent arguments. Separate maximum entropy models for the nominative, accusative and dative cases each choose the best candidate. Because the models are discriminative, new features are easy to add. The authors add three kinds:

• dependency-based language model scores trained on newspaper text,
• centering-inspired contextual cues,
• a context-based filter that shrinks the candidate set.

Experiments use the NAIST Text Corpus and the Kyoto Text Corpus with gold parses. Arguments that directly depend on the predicate are identified well. Zero-anaphoric arguments, especially those in other sentences, are rarely recovered because the model tends to prefer NULL. The work shows that a unified, feature-extensible framework can handle both tasks together.

Key points

• Data: NAIST Text Corpus v1.4b aligned with Kyoto Text Corpus 4.0 (newspaper articles). The training set has 1,751 articles and 67,145 predicates; the test set has 695 articles and 25,500 predicates.
• Language model features: Generation probabilities P(n|c,v) were estimated from Mainichi Shinbun articles (1991–2002, about 5.5M sentences), with Good-Turing discounting and backoff smoothing.
• Contextual features and filtering:
  • The 'Used' feature marks phrases that served as arguments of earlier predicates.
  • The 'SRLOrder' feature takes a phrase's rank in the Salience Reference List (Nariyama, 2002).
  • The candidate filter cuts the average pool from 102.2 to 3.2 phrases while still covering 62.5% of zero-pronoun referents.
• Overall results: Predicate-argument accuracy was 59.4% (15,140/25,500). Case-level F-scores were 73.2% (nominative), 89.2% (accusative) and 76.6% (dative). Directly dependent arguments exceeded 80% F in every case.
• Main weakness: Recall on inter-sentential nominative zero anaphora was only 7.6%. The system produced 488 such arguments against 3,063 in the gold standard, reflecting a strong bias toward NULL.