Research On Anaphoricity Detectionation In Coreference Resolution

Coreference resolution has been drawing more and more attention in recent years due to its importance in NLP applications, such as information extraction, text summarization. As an essential part of coreference resolution, anaphoricity determination restricts the performance of coreference resolution.This paper extensively explores anaphoricity determination of English by various methods, including rule-based, feature-based and kernel-based methods. In particular, this paper proposes anaphoricity determination by exploring dependency theory. It also presents a composite kernel to combine both flat features and structural information.In rule-based method, heuristic rules getting from grammar and syntactic perspective are proposed. In feature-based method, we focus on exploring various kinds of flat features which contain characteristic and contextual information of anphor. Experiments on the ACE2003 show that our feature-based anaphoricity determination filter helps coreference resolution with an improvement of 0.8-2.2 in F1 measure.In kernel-based method, we explore a new approach to dynamically determine the tree span by proposing three pruning strategies. Experiments show that dynamical extension strategy achieves the best performance. It also shows that our kernel-based anaphoricity determination filter enhances coreference resolution 0.6-1.7 improvement in F1 measure.Given the convolution tree kernel, the key issue is how to extract a pare tree structure from the parse tree. Compared with other researchers'results, this paper proposes a new approach to dynamically determine the tree span for anaphoricity determination by exploiting constituent dependencies to remove the noisy information, as well as keep the necessary information in the parse tree. Our approach achieves best performance with an improvement of 2.4-3.6 in F1 measure for coreference resolution. Finally, this paper furthers the research on kernel-based method by focusing on employing anaphoricity determination result as a feature for coreference resolution, exploring dependency tree pruning strategy and using composite kernel to combine structural information with flat feature information.