Section: New Results

Anaphoricity detection and coreference resolution

Participant : Emmanuel Lassalle.

Resolving coreference in a text, that is, partitioning mentions (noun phrases, verbs, etc) into referential entities, is a challenging task in NLP leading to many different approaches. Anaphoricity detection, on the other hand, consists in deciding whether a mention is anaphoric (aka discourse-old) or non-anaphoric (discourse-new). This task is strongly related to coreference resolution and has been mainly addressed as a preliminary task to solve, leading to pipeline architectures.

A first line of work compares several methods for learning latent structures encoding coreference clusters that optionally take into account very accurate constraints on mention pairs. We study the relationship between standard decoding strategies used with pairwise models and those used with structured learning of latent structures, providing both topological and empirical comparisons. We also show that further gains can be obtained by the addition of pairwise constraints. Our experiments on the CoNLL-2012 dataset show that our best system obtains state-of-the-art results, and significant gains compared to standard locally-trained models.

Our second line of work introduces a new structured model for learning anaphoricity detection and coreference resolution in a jointly. Specifically, we use a latent tree to represent the full coreference and anaphoric structure of a document at a global level, and we jointly learn the parameters of the two models using a version of the structured perceptron algorithm. This model is refined by the use of pairwise constraints, and our experiments on the CoNLL-2012 English datasets show large improvements in both coreference resolution and anaphoricity detection, compared to various competing architectures. Our best coreference system obtains a CoNLL score of 81.97 on gold mentions, which is to date the best score reported on this setting.

This work has been achieved in collaboration with Pascal Denis, a former Alpage member, now at Inria Lille-Nord-Europe (EPI Magnet).