Research On Answer Validation Technology Based On Textual Entailment Recognition

In recent years,with the fieryness of artificial intelligence,automatic question-answering techniques have received extensive attention.The fundamental difference between a question-answering system and a traditional search engine is that the question-answering system can return accurate answers to the user rather than a list of candidates.To ensure the accuracy of the answer,the question-answering system needs to verify the returned results before returning the answers to the user.How to apply textual entailment recognition method to verify the answer is the main research content of this paper.Because of the lack of large-scale Chinese entailment corpus,in order to apply the textual entailment recognition method based on deep learning,first we use Baidu Translate translates English SNLI corpus into Chinese to construct Chinese entailment corpus.Then we train the textual entailment recognition model based on the translated Chinese entailment corpus.This paper separately trains three textual entailment recognition models based on matching encoding,which are applied to the research of answer validation technology.The results of the question-answering system are mainly divided into lexical-level and sentence-level answers.This paper examines the lexical-level and sentence-level respectively.For the lexical-level answer,the answer may come from the knowledge base,Baidu Zhidao,we use knowledge base triples to assemble,know the question and answer to rewrite the declarative sentences as premise,the questions and answers rewritten as declarative sentences as hypothesis implies validation of answers.For the sentence-level answer,the answer may come from Baidu Zhidao and Baidu Encyclopedia,using the answer as an premise to determine the entailment,from the answer to the question and then determine the entailment relationship with question to verify the answer.In addition,some question answering systems return answers that are neither lexicals nor sentences,but rather longer paragraphs,so and at the same time research verification of paragraph-level answers.Because paragraphs are generally long and difficult to model,an alignment-based textual entailment recognition method is used for paragraph-level answer validation.In terms of evaluation,it is necessary to construct lexical and sentence-level answer test sets,which are based on an question-answering system of the QA group of the Harbin Institute of Technology's Social Computing and Information Retrieval Research Center.At the same time,an online question-answering system of the QA group of Social Computing and Information Retrieval Research Center of Harbin Institute of Technology was used to construct a test set for evaluating paragraph-level answers.According to the test set,evaluate textual entailment recognition effect applied to the validation of the answer and the imp rovement effect of the question-answering system.