Study On The Answer Ranking Based On Deep Learning Methods

Limited by the complexity of natural language,the traditional search engine which is based on keyword search method can not meet the users' need for exact problem.And the question-answering system based on common question sets can not solve the problem of open domain.Therefore,intelligent search is popular among the researchers.In this paper,we study the solution of answer ranking task based on the open-domain questionanswering system.After the effective retrieval process,we need to rank the rich candidate answers,which is the last step of the question-answering system.Answer ranking is also one of the most important parts of a question-answering system,since its returned results directly decide if the question-answering system is good or not.This paper focuses on the task of ranking answers by their semantic relevance to the questions text.In order to solve this problem,traditional machine learning methods rely on external resources,which is time-consuming and ineffective to extract features from the training data set.This paper studies the deep learning method to accomplish this task.Recent years,the deep learning method has made breakthroughs in many tasks of natural language processing,while we lack practical experience in the application of question answering system.Therefore,this paper study neural networks to get the intermediate representations of question-answer pairs,and further improvement to the deep learning model has made to complete the task of answer ranking.The main contributions of this paper are as follows.A bidirectional long-short-term memory(BiLSTM)based deep learning model was designed and implemented to rank answers.The bidirectional long-short-term memory neural network used in this model can increase the interaction among the text sequence,and the convolution neural network layer can convolve the local information.The input question-answer pairs are processed independently by the neural network model,and we get the intermediate vector representations of the pairs.Then we use similarity matrices to correlate the intermediate vectors of the input pairs and calculate their distances.Experimental results on a public benchmark dataset from TREC demonstrate that our system outperforms previous work which requires syntactic features and some deep learning models.What's more,the attention mechanism applied to the deep learning model has been studied in this paper,and we designed and implemented an improved deep learning model with attention mechanism for answer ranking.The long-short-term memory neural network(LSTM)layer was rewrote to combine attention mechanism.The question representation is the input parameter,acting as the weight to influence the answer representation so as to keep more information related to the question.In this paper,the attention mechanism is added to two deep learning models which contain BiLSTM network,and we conduct experiments to compare the two models when the attention mechanism is added or not.In summary,this paper studies and implements an improved deep learning model for answer ranking.Experiment results on the public data set prove its validity.Some solutions are also proposed to solve the problems happened in the practical answer ranking submodule of the question-answering system.