The Anaphora Resolution Research Based On Frame Semantic Annotation

The anaphora resolution is important and hard natural languageprocessing.In the text summarizationg、machine translation、multilingualinformation processing and information extraction and many otherapplications,anaphora resolution is applied.In this paper,I studies thetheory of the frame semantic annotation,such as:FrameNet theory、CFNtheory and other theories.In this paper,a large number of literature andscheme are studied at home and abroad,these literatures are on anaphoraresolution of frame semantic annotation.According to the requirement ofthe latest study case and provincial Natural Science Fund,the paperproposes the anaphora resolution research based on frame semanticannotation.According to the shortcomings that the study find,the paperproposes their own research content.This paper analyzes a large numberof studies,finally the following three aspects will be the depth study.Firstly,a framework corpus is to be builded.In anaphora resolutionstudy,corpuses are data that will be processed.During the study,theexperiment need to collect corpus.In this paper,the corpus is collectedfrom two aspects,namely teaching textbooks and network,a collection of121corpus.In this paper,the corpus is preprocessed by LTP preprocessingtools.121corpuses are stored in XML format.Finally,the experimentneeds to give corpus.Secondly,the anaphora resolution based on rules and maximumentropy is studied.The algorithm adopts five kinds of rules,they aresingular and plural consistent,syntax with a consistent,genderconsistency,gender agreement and semantic information consistent.TheMaximum Entropy algorithm adopts13kinds of features.Finally,the experimental results based on the rules and maximum entropy algorithmcompared with the results of maximum entropy algorithm.Bothalgorithms are implemented anaphora resolution.The experimental resultsof rules and maximum entropy is higher than results of the maximumentropy.Lastly,the anaphora resolution based on rules and tree kernelfunction is studied.The algorithm adopts five kinds of simple rules.Thealgorithm extracts the five kinds of structured information tree,they areMCT treeg、CT tree、SPT tree、MT trees and RMLSPT trees.Thealgorithm adopts26features.This paper also studies anaphora resolutionbased on tree kernel function.The best result of based on tree kernelfunction is RMLSPT tree.The result of rule and tree kernel function ishigher than the result of tree kernel function.