Research On Opinion Target Extraction Based On Syntactic Structures And Dependency Relationships

Syntactic structure and dependencies play an important role in guiding the capture of correspondence between opinion targets and emotional words.However,it is rarely used in current machine learning-based opinion target extraction.Aiming at this problem,this paper defines new features by comprehensively analyzing the syntactic structure and dependencies between words,and combines existing features such as words,part of speech and emotional words to extract the opinion targets.Firstly,different components of opinion targets and emotional words are analyzed in the syntax,besides the possible combinations,to summarize the syntactic structures between opinion targets and emotional words,opinion targets and opinion targets,and emotional words and emotional words.Secondly,according to different syntactic structures,the dependency relationships between the emotional words and the opinion targets,such as the subject relationship and the attribute relationship,are analyzed.Then,emotional words are marked by part of speech and dependencies,and the dependency relationships are defined as the features.Finally,the new features and existing features are applied to the conditional random field model for opinion target extraction.The comparative experiments have conducted on the three data sets of COAE2011 task3.Firstly,the influences of different templates on the extraction are analyzed under the same characteristics.Secondly,the performance of different feature combinations in the opinion target extraction in the case of the same template.Thirdly,under the same data set,the results of two machine learning method are compared.The experimental results show that the optimal extraction results in this paper have different degrees of improvement in evaluation indicators,which proves that the features based on syntactic structures and dependency relationships can be fully captured between the opinion targets and the emotional words.The opinion targets extraction effect can be improved.