Research On Answer Selection Ranking Based On Attention Mechanism

The use of a search engine facilitates people's lives,but its query results still need to be manually filtered.Given a question,how to get the correct answer accurately is crucial.Answer selection is a starting point for solving this problem.The key is to calculate the semantic similarity between the question and its candidate answers.Under certain conditions for the calculation method of semantic similarity,the semantic similarity between question and its candidate answers mainly depends on the grammatical semantic information inside the question and its candidate answers and the semantic information between question and its candidate answers.This paper builds a new answer selection ranking model based on the concept of semantic relevance distance,and improves the model performance from the above two aspects.The main contributions of this research are as follows:1.This paper puts forward the concept hypothesis of the semantic relevance distance between question and its candidate answers,explaining in detail how to distinguish the correct answer from the wrong answer in the candidate answer set,so as to select the correct answer.Then,a new answer selection ranking model is constructed based on the theory.2.For the answer selection ranking task,this paper use a combination of statistics and deep learning to extract the part-of-speech features of a sentence.First the deep learning model is used to capture context semantic of the sentence,and the statistical model is used for part-of-speech tagging and solve the tagging problem of unregistered words.Then,efficient part-of-speech feature is extracted by integrating the advantages of the two models,and verificated the tagging performance.Finally,combining partof-speech features with the ranking model,the internal grammatical and semantic information of the question and its candidate answers is enhanced.3.Introduce the attention mechanism through the fused semantics of part-ofspeech and original corpus.Different from the existing attention usage methods,such as construct attention matrix by the original corpus or a feature of question and its candidate answers,this paper first fuses the part of speech with the original corpus,then uses the fused vector to construct the attention matrix,and then fuses Into the ranking model,enhance the semantic information between question and its candidate answers.Based on the above research,the ranking model in this paper is experimented on the NLPCC 2017 DBQA dataset.The results show that after integrating part-of-speech and attention in the baseline ranking model,its three performance indicators of MAP,MRR,and ACC@1 reach 79.34%,79.42%,and 70.13%,respectively,which are significantly higher than the performance of other ranking models.Thereby,the correctness of the theoretical assumptions and the rationality,effectiveness and advancedness of the model were verified.