Sentence Similarity Calculation Based On Semantic Role Labeling

Sentence similarity,as one of the basic problems in natural language understanding,is the measurement of semantic equivalence of sentences.The existing methods of sentence similarity calculation usually analyse from the surface of the sentence,and it is one of the difficulties in the study of the sentence semantic.This paper analyzed the sentences from the semantic level and put forward a method of sentence similarity calculation based on semantic role labeling.The effect of semantic role labeling has a direct impact on the results of sentence similarity calculation.However,the existing semantic role labeling methods have a problem in analying complex sentences,so it is necessary to improve the effect of semantic role labeling.In the aspect of semantic role labeling,the sentences were processed by pruning and the clause extraction combined with the phrase structure syntax tree,then semantic role labeling.Finally,semantic role boundary after restoration was revised combined with phrase tree.The F value of the WSJ dataset in CoNLL2005 is 88.25%.The experimental results show that the introduction of phrase structure syntax can effectively improve the labeling of semantic roles.In the aspect of sentence similarity calculation,this paper adopted two kinds of semantic expressions,Deep Structured Semantic Model and the semantic role labeling,to calculate the similarity of sentences.Based on the Deep Structured Semantic Model expression,the sentences are vectorized and the sentence similarity is computed using vectors.Based on semantic role labeling expression,the matching of predicate pairs and computation of similarity between semantic roles were performed based on automatic labeling of semantic roles.The results of the two semantic expressions are linearly combined as the whole similarity of sentences.In this paper,an experiment was conducted on the SemEval2017 evaluation corpus,and pearson correlation coefficient reached 85.746%.The experimental results show that the method proposed in this paper can effectively improve the calculation of sentence similarity.In the aspect of sentence topic analysis,this paper presents a method of topic classification based on support vector machine.The sentences in the corpus with the scores 1to 5 are classified as the same topic,and 0 to 1 the different topics.Taking the sentence similarity calculated by DSSM,CDSSM,Skipthoughts and so on as input features,the similarity of sentences is classified into two categories by using SVM,and then the similarityis modified on the basis of this.In the same SemEval2017 evaluation corpus,the pearson correlation coefficient increased from 85.746% to 85.921%.