Research On The Model Of Answer Selection Based On Deep Neural Network

Answer selection is a very important task in natural language processing,and it is also an indispensable component in the question answering system.The basic definition of answer selection is to give the question and the answer candidate pool for this question,the purpose is to find the best answer to the question accurately from the answer candidate pool.The most difficult challenge for this task is that the answers may not directly share the vocabulary units with the questions.The questions and answers may only be semantically related,and the information of the candidate answers is sometimes noisy,and may contain a lot of irrelevant information or even negatively related information.In the early stage,the answer selection task is implemented by traditional machine learning algorithms,but this method is very time-consuming,and the trained model is prone to over-fitting problems.Deep learning has flourished in the past decade,and has achieved good results in answer selection tasks,among which are convolutional neural networks(CNN),recurrent neural networks(RNN),and bidirectional long-term short-term memory neural network(BiLSTM).These neural networks have their own advantages and disadvantages.For example,CNN can obtain the local semantics of the sentence,but it is difficult to obtain long-distance information.RNN can obtain the sequence information of the context,but it is difficult to obtain the local semantics of a sentence.In response to the problems mentioned above,the main work of this research includes the following points:1.Research multiple answer selection models based on deep neural networks.To solve the problem that CNN cannot obtain long-distance information and RNN cannot obtain sentence local semantics,the existing answer selection model is improved by combining multiple neural networks.And the feasibility of multiple combined models was verified through experiments.2.Research the space vector representation model and semantic feature alignment model of the answer selection task,the attention mechanism is introduced under various neural neural network models to establish a variety of combined models,and the effect of the improved model(BiLSTM-ATT)is verified on the data set.The effect of this model is better than most existing models.Then,the model was used as the infrastructure model for further improvement.3.Aiming at the semantic gap between the question and the answer,the external knowledge of the knowledge-enhancing memory network is introduced to fill the semantic gap between the question and the answer.After applying the method to the infrastructure model,MAP increased by 2.55% and MRR increased by 2.85 %.4.Aiming at the problem that the traditional attention mechanism raises unrelated or even negatively related semantic information,design and implement a dynamic threshold mechanism to solve the problem of attention mechanism redundancy encountered in the answer selection task.After applying this method to the infrastructure model,MAP increased by 1.74% and MRR increased by 2.15%.5.The knowledge-enhanced memory network and the dynamic threshold of attention are simultaneously cited in the infrastructure model for experimentation,and compared the model with existing models.The final improved model(KQA-BiLSTM-D-ATT)compared with the infrastructure model,its MAP increased by 4.20% and MRR increased by 4.04%.The model compared with the optimal model,its MAP increased by 2.23% and MRR increased by 2.26%.