Research On Answer Selection Algorithm Based On Deep Learning

The Question-Answering is a hot research field in Natural Language Processing and answer selection is an important part of the Question-Answering.There are two key questions to be solved in the Question-Answering system: one is how to make semantic representations about questions and answers,and the other is how to achieve semantic matching between questions and answers.In recent years,deep learning has achieved good results in the field of natural language processing.Using deep learning technology can learn the rich semantic information of sentences and show great advantages in dealing with answer selection tasks.This thesis focuses on deep learning based answer selection tasks,and proceeds from two main perspectives: semantic representation and semantic matching.Finally,based on the research results of this subject,the visualization tool of the model is realized.The main work of this thesis includes the following aspects:1.Deep learning techniques are used to study the answer selection model based on sentence-to-semantic representation.The thesis implements the representation model of sentence pair based on CNN and BiLSTM respectively.The two models are compared by experiments,the optimal network structure of the CNN model and the BiLSTM model in the answer selection task is obtained.At the same time,the experimental results show that the performance of sentence-to-semantic representation model based on CNN is better than that of the semantic representation of sentence pair based on BiLSTM.2.The attention-based model is used to study the answer selection model of semantic representation based on sentences.This thesis mainly focuses on the learning of alignment features,and proposes to construct alignment feature matrix by means of residual connection.It is proved by experiments that the proposed model shows good results in the answer selection task.3.The answer selection model based on sentence matching is studied,and the two aspects of the construction of the matching matrix and the feature extraction of the matching matrix are explored.In this thesis,a deep matching model is proposed,and the semantic matching matrix is constructed at different semantic levels.The DenseNet network is used to extract the matching matrix.The validity of the model in the answer selection task is proved by experiments.4.A client tool is implemented which can visualize the training,validation,and testing of the model.