Candidate Answer Sentences Selection Based On Deep Learning

The question and answer system usually consists of four parts: the problem analysis module,the retrieval module,the answer extraction and the answer matching.Answer selection is a key component of a typical question-and-answer system.Answer selection questions can be expressed as follows: Given questions and candidate answers,our goal is to find the best candidate answer,among them.The answer sentence is a sequence of tokens of arbitrary length,and the question can correspond to multiple substantial ly correct answer sentences.Traditional methods usually use feature engineering and language tools to study sentence selection tasks,sentence segmentation,sentence part-ofspeech tagging,grammar analysis,syntax analysis,manual extraction of text feat ures,and then calculate the semantic similarity between answers and questions,and based on similarity Sorting selects the answer sentence with the most similar problem as the final answer to the question.Although these methods have achieved certain effe cts,they have these shortcomings:(1)The artificial extraction of text features has human subjectivity and cannot be very comprehensive.Objectively understand all the problems;(2)In order to obtain good text features,you need to manually adjust the feature extraction strategy.(3)The impact of system complexity on the use of feature engineering and language tools.Compared with the use of feature engineering and language tools,deep learning can actively learn the semantic information in the text,and better extract the characteristics of the text.This paper proposes a deep learning framework that does not require any language tools or external resources for the answer selection task in the question and answer system.First,the basic model,long-term and short-term memory network(LSTM)model is built to construct the embedded problem and answer,and based on this,Attention mechanism,which can better understand the feature representation of the problem and answer.In this paper,we call this model Attention-biLSTM,which still has limitations because the model's attention can only focus on the embedded problem at a time.And a candidate answer,can not simultaneously pay attention to all candidate answers and the information of all the words in the candidate answers,propose a complex attention mechanism to enhance this design,the proposed attention mechanism increases the global feature representation of the candidate answer,combined Information about all the words in the answer.Combine the local information of a particular part of the answer with the global representation of the entire problem.Experimental analysis shows that the model in this paper is superior to other models,and the experimental results of the improved Attention-biLSTM model show that it is significantly better than Attention-biLSTM.At maximum pooling,the Attention-biLSTM and the improved Attention-biLSTM increased by 5.6% and 7.8%,respectively,on test set 1 compared to the base model biLSTM,and by 4.2% on test set 2,respectively.7%.