📄 Adapted End-to-End Coreference Resolution System for Anaphoric Identities in Dialogues

Summary

Most neural coreference systems are developed and evaluated on OntoNotes, which omits singletons and split antecedents and contains little conversational text. As a result, reported scores may overstate how well these systems handle real dialogue. Xu and Choi address the anaphoric identity track of the CRAC 2021 shared task, whose data use the Universal Anaphora (UA) format. They start from the end-to-end mention-ranking model of Lee et al., built on a pretrained Transformer encoder, and make three adaptations:

• Singletons: a jointly trained mention-score objective lets confident candidates that have no antecedent become singleton clusters.
• Speakers and turns: a special speaker token is placed before each utterance, so the encoder can track who speaks in each turn.
• Knowledge transfer: the model is pretrained on OntoNotes (CoNLL format) and then fine-tuned on ARRAU data in the UA format.

These simple changes improve results by up to 27 F1 over the baseline. The final system ranked first on all four test corpora. The study shows that handling singletons and encoding dialogue structure are central to coreference resolution in conversation.

Key points

• Data: UA training data come from ARRAU sub-corpora (TRAINS-93, PEAR, RST, GNOME), with TRAINS-91 used for development. Evaluation covers AMI, LIGHT, Persuasion for Good (PSUA) and Switchboard (SWBD). The additional non-UA resource is OntoNotes.
• Singletons: singletons make up at least 73% of gold clusters in every test set. Recognizing them adds 17–22 average F1 over the plain mention-ranking model and raises mention recall by 26–36%.
• Speaker encoding: speaker-token augmentation adds roughly 2–3 F1 on three of the four datasets.
• Knowledge transfer: pretraining on OntoNotes followed by fine-tuning (SE +P) outperforms simply mixing the corpora. Gains are about 6–8 F1 on LIGHT and PSUA, and smaller on AMI and SWBD (0.5 and 2.6 F1).
• Final results: the submission, which also trained on the dev sets, scored 63.96 F1 on AMI, 80.33 on LIGHT, 78.41 on PSUA and 74.49 on SWBD, taking 1st place in the track. Split antecedents are left for future work.