Research On Automatic Question Answering Technology Based On Attention Mechanism

This is an era of vigorous development of information technology.People can instantly browse the information they care about through the Internet.With the rapid development of technology and the prosperity of the Internet,the amount of data generated on the Internet today is far beyond the traditional text era.Although rich information resources can satisfy people's thirst for knowledge,how to quickly screen for valuable information has also become a difficult problem.The automatic question answering system can mine the connection between the latent semantic information in the sentence and the candidate answer,and find the best answer quickly and accurately,which can meet the needs of users who want to accurately locate the answer.We study the answer selection technology in the automatic question answering system.The specific process of this task is as follows: Given a question and multiple candidate answers,the goal is to use deep neural network learning to find the most relevant answer from the candidate answer set.The key technique of this task is to calculate the degree of similarity between the question and the candidate answer.Deep neural networks do not need to manually extract auxiliary features such as features,language tools or external knowledge,and use the network itself to extract semantic relationships.This paper uses deep neural networks to study answer selection techniques.The basic model used in this paper is: a single-layer BiLSTM network model based on the attention mechanism.Based on this model,this paper has improved from two perspectives: attention mechanism and feature extraction using multi-layer networks.First,for the attention mechanism,this paper proposes a two-way attention mechanism that co-occurs between the question and the answer step by step to highlight the important part of the question and answer.Secondly,for the single-layer network structure of the basic model,this paper has actively explored the use of multi-layer networks.In addition,this paper also explored the use of BERT pre-trained models.We used BERT pre-trained models for feature extraction to generate word vectors rich in contextual information instead of word vectors generated using Word2 Vec.Comparing the experimental results of the improved model and the basic model on the 2016 NLPCC QA test set and the Wiki QA data set,it is found that the improved answer selection model is superior to the basic model.Among them,the optimal model combining multiple improved technologies reached 83.73%,83.67% and 75.10% on the 2016 NLPCC QA test set evaluation indicators MRR,MAP and ACC.Compared with the basic model,our model has improved by 5.20%,5.28% and 6.30% respectively.