Transition-Based Mention Representation for Neural Coreference Resolution

Coreference resolution plays an important role in text understanding. Recently, various neural approaches have been proposed and achieved success. Although most researches agree that mention extraction and representation much impact the performance of coreference resolution, existing neural architectures consider all possible spans up to a maximum length and only employ a simple ffat word-level span representation. In this way, those information which has been proved to be effective in previous non-neural coreference resolution, such as structural information, has been largely ignored. In this paper, for coreference resolution, we propose a uniffed transition-based approach to extract and represent mentions simultaneously. In particular, we propose a Simplified Constituent Parse Tree (SCPT) scheme for each sentence by only keeping the local detail inside the mentions and the coarse frame structure outside the mentions. That is each mention corresponds to a constituent of the obtained SCPT. Then we employ a transition-based strategy to construct the SCPT structure in a bottom-up manner. In this way, various potential mentions (i.e., constituents) can be obtained and the corresponding transition action sequences embedded with internal structural information can be viewed as their proper representations. After that, we employ such achieved potential mentions and their transition-based representations for neural coreference resolution.