Multi-features Based Textual Entailment Recognition In Chinese

Textual entailment recognition aims to build a common framework for textual inference andsolve the problem of variability of semantic expressions in natural language. If we changeinference problem into the problem that recognize the textual entailment relationship betweentwo text fragments, any kind of natural language processing applications can realize textualinference with the method of textual entailment recognition. As the foundation of many appliedresearches in the field of natural language processing, the textual entailment recognition has beenwidespread concerned.Textual entailment recognition is essentially a classification problem, so this paper usessupport vector machine to design textual entailment classifier, and the output of classifier is2-way label or5-way label. The2-way classifier is used to identify the entailment andnon-entailment relationship, and the5-way classifier is used to identify forward entailment,reverse entailment, bidirection entailment, contradiction and independece relationship. In orderto improve the accuracy of classification method, this paper uses a variety of features, includingsurface statistical feature, lexical semantic feature, syntactic feature and event semantic feature.The event semantic feature based textual entailment recognition approach firstly analysizesthe event labeled text pair and the event graph generated. And then with the event graphs, theentailment recognition between text pairs can be changed to entailment recognition betweenevent graphs, and the event semantic feature can be computed based on max-common subgraph.At last, the method of support vector machine based textual entailment recognition using eventsemantic feature, combined with the surface statistical feature, lexical semantic feature andsyntactic feature, can get the preliminary experimental result, and the correction module basedon event semantic rules handles preliminary experimental result to obtain the final experimentalresult.In the evaluation of the experimental results, we use three common standards, Precision,Recall and F-measure. The experiment results show that multi-features based textual entailmentrecognition approach can achieve good recognition performance, and surface statistical feature,lexical semantic feature, syntactic feature and event semantic feature are suit for the textualentailment recognition in Chinese.