📄 Using Type Information to Improve Entity Coreference Resolution

Summary

Neural coreference resolution (CR) systems built on contextual encoders such as BERT usually ignore the entity-type labels that many CR corpora already contain. Khosla and Rose test whether making these types explicit helps. They start from the mention-ranking model of Bamman et al. (2020), which derives from Lee et al. (2017), and add type knowledge in two places. First, each mention's span vector is extended with its type (+ET-self). Second, the pairwise scorer receives a binary flag marking whether two candidate mentions differ in type (+ET-cross). This flag acts as a soft filter rather than a hard one, so bridging-anaphora cases are not ruled out. Experiments use gold mentions on LitBank, EmailCoref, WikiCoref and OntoNotes. Each corpus is tested with its own tag set and with a shared coarse set. Both gold and automatically predicted types give modest but consistent gains over the baseline, and predicted types still beat it significantly. The study shows that explicit semantic types remain useful even for BERT-based CR.

Key points

• The baseline is the BERT-based model of Bamman et al. (2020), which handles only mention linking over gold mentions. The combined system, +ET, merges +ET-self and +ET-cross.
• Four corpora are used:
  • LitBank: 100 literary texts, with 83.1% of mentions referring to persons.
  • EmailCoref: 46 threads containing 245 messages.
  • WikiCoref: 30 Wikipedia documents, split 24/3/3, with types taken from Freebase.
  • OntoNotes: a much larger type inventory.
• Tag granularity varies widely across corpora, from 4 types in EmailCoref to 18 in OntoNotes. The authors therefore also map all corpora to a common set of PER, ORG, LOC, FAC and OTHER.
• A separate type predictor uses the mention span and its sentence context. It struggles with demonstrative pronouns (this, that, it), because these carry little type signal.
• An EmailCoref case study shows type-aware models fixing errors that the baseline makes. Predicted types still fail when sentence-level context is insufficient.