Research On Feature-based Coreference Resolution

Coreference resolution plays an important role in natural language processing. It involves the resolution of various kinds of noun phrases, such as named entities, nominal phrases and pronouns. This paper recasts coreference resolution as a classification problem, and focuses on feature-based machine learning approaches on both English and Chinese languages.Our feature-based coreference resolution system is first built via a pipeline of NLP techniques, including POS tagging, named entity recognition and noun phrase chunking. Then, a number of effective features and their combinations are explored using the maximum entropy model and the SVM model. Evaluation on the English MUC-6 corpus shows that our system achieves the F1-measures of 68.0 and 68.1 using maximum entropy and SVM respectively. For evaluation on the English ACE2003 corpus, the SVM model achieves the F-measure of 53.1, 58.4 and 54.2 on the BNEWS, NPAPER and NWIRE portions respectively.We also adapt the above system on the Chinese language. Evaluation on the ACE 2005 Chinese corpus (with annotated named entities and 200/98 articles as the training/test data) shows that our system achieves the F1-measure of 70.83.In particular, this paper discusses the effectiveness of various features and machine learning models on both Chinese and English coreference resolution with detailed experiments and error analysis.Evaluation shows that our system provides a good platform for both English and Chinese coreference resolution.