Research On Answer Generation Based On Sequence-to-sequence Model

Question Answering(QA)attracts much concern and has broad prospects for development in the field of artificial intelligence and natural language processing,which is also one of the most difficult research directions.It involves natural language processing,knowledge reasoning and so on.With the development of deep learning technology,question answering has become the focus of discussion in scientific and technological media and researchers' community,and more and more researchers are keen to apply recurrent neural network(RNN)to question answering to generate answers.The research object of answer generation is text.The length of the text varies,and there is a sequence relationship between the elements in the text,so that the text data belongs to the variable length sequence data,so the answer generation problem can be regarded as a sequence to sequence problem.The characteristic of the sequence-tosequence problem is that the input and output are variable in length,and there is a sequential relationship between the input and output elements.Therefore,in order to solve these problems,this paper chooses the sequence-to-sequence model.The sequence-to-sequence model can obtain output sequences of different lengths for input sequences of different lengths,conforming to the variable length characteristics of the text,and considering the order relationship between text elements.This paper studies the answer generation in Chinese restricted domain.Because the implementation of question answering is divided into retrieval and generation,this paper generates answers based on these two methods.Among them,the retrieval method is based on the fusion of retrieval.Firstly,using the vector space model achieves text matching based on the keywords,and then returning the answer candidate list according to the matching results.Finally,based on the synonym forest,the semantic similarity of the text is realized,and the answer fusion is realized for the candidate list.Finally,the semantic similarity of the text and the answer fusion for the candidate list are realized based on the synonym forest.The generation is based on sequence-tosequence model.Using long short-term memory(LSTM)and combining with attention mechanism to generate answers,and then incorporating sentence embedding module which is based on self-attention and bidirectional long short-term memory.Finally,achieving sequence-to-sequence model answer generates answers based on selfattention.The experimental results show that the system based on sequence-tosequence model and self-attention mechanism is effective.