Coreference Resolution as Span Boundary Alignment

Anonymous

Abstract: We propose a new fast and accurate solution for coreference resolution, in which the task is formulated as a span boundary alignment problem. In this solution, a mention is linked to another one via two edges modeling how likely two linked mentions point to the same entity. Specifically, for each mention, its head word (left boundary) needs to be well aligned with the head words of all other mentions that refer to the same entity, so does its tail word (right boundary). Such a ``head-to-head'' and ``tail-to-tail'' alignment strategy greatly reduces the computational complexity of coreference decisions on any pair of mentions, mitigates the error propagation problem caused by mention pruning, and encourages the sharing of features across all mentions that refer to the same entity. Experimental results show that our solution achieves close to state-of-the-art performance on the CoNLL-2012 and GAP benchmarks with much less computational cost.