📄 Knowledge-aware Pronoun Coreference Resolution

Summary

Pronouns carry little meaning on their own, so linking them to the right antecedent often depends on background knowledge. This is especially true in specialised domains such as clinical text. Earlier systems either hand-coded knowledge as rules or features, or relied on neural end-to-end models that learn only from context and tend to overfit their training domain.

The authors propose an end-to-end neural model that takes knowledge directly as triplets, the standard knowledge-graph format. Spans and pronouns are encoded with a BiLSTM plus attention. Triplets whose head string matches a span are retrieved and represented by averaging the embeddings of their tail words. A knowledge attention module, conditioned on each span–pronoun pair, then weights the retrieved triplets so that only contextually useful facts contribute to the coreference score.

The model was tested on CoNLL-2012 (news) and i2b2 (medical records) and beat strong baselines, including End2end, by a large margin. The gap was largest in cross-domain transfer. This suggests that learning to use external knowledge, not just fitting labelled data, makes coreference systems more robust and portable.

Key points

• Datasets: CoNLL-2012 shared task data drawn from OntoNotes 5.0, and the i2b2 shared task corpus of de-identified medical records from Partners HealthCare and Beth Israel Deaconess.
• Pronouns and candidates: Three pronoun types are covered: third-person personal, possessive, and demonstrative. Candidate antecedents come from the current sentence and the two preceding ones, giving on average 1.3 (CoNLL) and 1.4 (i2b2) correct references per pronoun.
• Knowledge sources: All resources are merged into one triplet set. Examples include OMCS, syntactic co-occurrence triplets such as ('magazine', dobj, 'drop'), and type facts such as erythema IsA disease.
• Cross-domain results: Trained on CoNLL and tested on i2b2, the model reaches 80.9 F1 versus 75.2 for End2end. In the reverse direction (i2b2 to CoNLL) it leads End2end by 22.7 F1, although absolute scores remain low.
• Availability: Code and data are released at github.com/HKUST-KnowComp/Pronoun-Coref-KG.