Are Word Embedding and Dialogue Act Class-based Features Useful for Coreference Resolution in Dialogue?

Samarth Agrawal, Aditya Joshi, Joe Cheri Ross, Pushpak Bhattacharyya, Harshawardhan M. Wabgaonkar

Abstract: Due to the rise in the popularity of chatbots, there is a need to revisit the past work in coreference resolution in dialogue. Dialogues pose unique challenges to coreference resolution. This paper introduces a novel set of features based on word embeddings and dialogue act classes for the task. We show that our system with these novel features gives an improvement 24.8% in F-score over previous work on the same dataset. Additionally, we also evaluate our system using the CoNLL metrics and report the best CoNLL score of 75.93. This paper establishes the importance of these features for coreference resolution in dialogues and points to further work that can be done for the task.