Tree Kernel-based Pronoun Coreference Resolution

As a critical and hot research topic in NLP,coreference resolution plays an important role in many NLP applications,such as text summarization,machine translation and multi-language information processing.Having drawn great attention in the machine learning community,tree kernel methods are expected to have broad application prospects, and have been successfully applied in NLP to a certain extent.While previous research focuses on application of lexical and syntactic information in coreference resolution,this paper systematically explores various kinds of semantic information which automatically acquired from convolution kernel.With syntax tree considered as a feature,a number of effective features are explored to resolve pronominal resolution using the SVM model.In the research of tree kernel based English pronoun resolution,this paper first explores the sub tree through a suitable expanding and pruning strategy of syntactic parsing tree.Pruning strategy can be divided into static and dynamic parse tree.Evaluation on the ACE 2004 NWIRE corpus shows that dynamic parse tree outperforms than static parse tree with its F-measure 79.3%.Due to the lack of distinction,this paper considers semantic expansion of parse tree by adding some semantic nodes to the parse tree.The experimental results show that the semantic expansion of parse tree can significantly improve the performance of the system.Furthermore,the paper takes into account the filter of the training examples as well as the redundancy of 'it',to optimize the performance of the classifier.The filter to training examples refers to a number of examples obviously coreferential,which will reduce the noise of the classifier significantly;while the filter of 'it' refers to those that do not need to resolution.The dependence of convolution tree-kernel on the classifier and cross-sentence anaphora are also analyzed.Moreover,a machine learning based Chinese pronoun resolution system is also developed.Because it's an alpha system,so we only consider the pruning in the Chinese pronominal resolution.Experimental results both in English and Chinese ACE2004 NWIRE corpus show that the convolution kernel can effectively improve the resolution of pronouns,with F-measure 82.1%and 50.3%respectively.Evaluation shows that our system provides a good platform for pronominal resolution with high compatibility of algorithm.