Coreference Resolution Based On Anaphricity Identification And Global Optimization

This paper mainly focuses on the research of using anaphoricity identification and global optimization to improve the coreference resolution. By using coreference determination result feedback and tunable parameter, the coreference resolution can combine with different anaphoricity identification.We treat the anaphoricity identification as a classification problem, and apply maximum entropy model and 70 features to build an anaphoric classifier. In order to tune the precision and recall of anaphoricity in a certain range, a corpus ration control method is proposed and a probability threshold method is introduced in this paper. Use anaphoricity classification before coreference resolution may remove much noise.The baseline system is a coreference resolution system that constructed by maximum entropy model and 65 linguistics features. Two kinds of methods is employed in generating training samples. As the reason that anaphoricity classifier is a kind of filter and baseline system combine with the different anaphoric classifier will generate different results, we can use the corpus ration parameter and probability threshold parameter to tune the anaphoricity identification in a global way, and optimize the coreference resolution.We also tried another global optimization method in this paper: apply integer linear programming (ILP) optimize the coreference resolution. We treat the coreference resolution as a optimization problem and introduce ILP to realize the global optimization with the output of maximum entropy model. We also proposed loose transitive constraints to build the feasible region.In the experiment and analyze part we take a comparison in baseline system, the anaphoric-coreference system, the ILP optimization system and other two coreference resolution system. This paper also take a comparison between local optimization method and global optimization. The experiment results shows that anaphoric-coreference get the best result in average evaluation and upgrade the reuslt of baseline from 50.57% to 53.35%.