Research And Application Of Answer Selection Algorithm Based On Deep Learning

At present,answer selection is one of the research hotspots in the field of natural language processing,and it plays a very important role in automatic question answering and search applica-tions.The existing answer-selection methods focus more on the matching between short texts,lack of research on long-text application scenarios,and it is difficult to solve the problems of ”semantic migration” and ”semantic gap” in the field of long-text applications.Transfer learning and graph neural networks have been widely applied and studied in recent years.Transfer learning has the characteristics of incorporating external knowledge.Graph neural networks are suitable for long-text modeling.Therefore,this article aims at the long-text matching problem,and tries to combine transfer learning and graph neural network to effectively solve the long-text answer-selection prob-lem.The specific contributions of this article are as follows:1)The purpose of the answer-selection method is to correctly match the question with the target answer.The dataset is mainly a set of questions and candidate answers.The existing answer-selection dataset in the question answer matching task lacks the feature of ”short question and long answer”,and the application scenarios of long answer matching widely appear in the field of automatic question answering.Therefore,in order to study the answer-selection method with the long-text feature,we constructs a dataset called CMASD in medical answer-selection field,which consists of 5 million short-question-long-answer data.2)Study transfer-learning methods to enrich long-text semantic representations.We proposes an answer-selection method called BertAttTL based on BERT-sentence vectors,and introduces a transfer-learning method to obtain sentence-level semantic vectors with closer semantics through the siamese network structure and attention mechanism.Through the analysis of the results of sev-eral comparative experiments and ablation experiments,its performance exceeds the existing answer-selection methods.Under a large number of answer search scenarios,it directly obtains all sentence vectors,which avoids the time-consuming calculation between multiple text pairs,and the calcula-tion efficiency is higher.3)Study graph neural network to model long-text answers,we proposes an answer-selection method called BertAttTL-GCN.By constructing a fine-grained data network graph,GCN was used to effectively obtain fine-grained semantic information for long-text answers.Combined with the BERT-sentence vector,BertAttTL-GCN's MRR value on the CMASD dataset reaches 77.26 %.Based on the sentence-level semantic information of BERT-sentence vectors,we add more fine-grained word-level semantic information,and finally achieves higher matching performance through the semantic interaction of questions and answers.4)Based on the answer-selection method proposed in this article and the medical dataset called CMASD,an answer-search system in the medical field is designed and implemented,which gets higher search performance of long-text answers.