Research On Mongolian Automatic Question And Answer Combining Retrieval And Generation

Question and answer system(Q?A)is a highly regarded research direction in the field of artificial intelligence and an important branch of natural language processing.With the upsurge of Internet data,improved performance of hardware devices and the maturity of deep learning technology,more and more intelligent products have been integrated into people's lives.At present,the mainstream automatic question and answer systems are mainly in Chinese,English and other languages.Due to the complexity of Mongolian script itself,the small number of researchers and the lack of publicly available question and answer corpus,the research on Mongolian question and answer systems is still in the initial stage.In this paper,the following studies on Mongolian automatic Q?A are conducted.1.Building a Mongolian Q?A corpusBy collecting,filtering,translating and correcting the existing Chinese Q? A corpus,a Mongolian daily Q ? A corpus containing 100,000 Q ? A pairs is constructed,which provides data support for the subsequent research.2.A retrieval-based Mongolian Q?A model based on machine learning was studiedThe retrieval-based Q?A model uses the TF-IDF method to achieve text vectorization,and calculates the similarity between user input questions and questions in the corpus,and then matches the most appropriate answers back to the user.The experiments show that the model has high response accuracy and simple implementation features.3.A deep learning-based generative Mongolian question and answer model is investigatedThe retrieval-based question-and-answer model cannot provide valuable responses to questions beyond the scope of the corpus.In this paper,a deep learning-based generative Mongolian Q?A model is investigated,and a recurrent neural network-based model and a Transformer-based model are implemented respectively.The recurrent neural network-based model uses LSTM,Bi LSTM,and Bi GRU as network units,and also introduces pre-trained Word2 vec word vectors and Attention mechanism.It was found that the model based on the recurrent neural network model with Bi GRU as the network unit and the Attention mechanism was able to understand the user input better,and the replies were more fluent and had the best results.The Transformer model was improved in all aspects compared to the recurrent neural network model,and the quality of the generated Mongolian text was better,with an improvement of 38.51 in the confusion index.4.An automatic Mongolian question and answer model incorporating Mongolian word slicing is investigatedIn view of the problem of OOV in Mongolian question answering due to the scarcity of corpus,.In this paper,the Mongolian Q?A corpus is experimented with Mongolian partial cut,BPE subword cut and Unigram subword cut based on Transformer Mongolian Q ? A model.The experimental results show that the Mongolian word slice can solve the problem of unregistered words and improve the performance of the Q ? A system;comparing the three slice methods,the Mongolian Q ? A model based on Unigram subword slice has the best effect,and the confusion indexes are improved by 0.8 and 0.52 compared with partial slice and BPE subword slice,respectively.5.The study implements a Mongolian automatic question and answer model that integrates retrieval and generationIn order to make full use of the advantages of the two models and improve the performance of the Mongolian Q?A model,a fused Mongolian Q?A model is studied by fusing the retrieval Mongolian Q?A model with the generative Mongolian Q?A model.The experimental results show that the fusion model has improved the performance compared with the other two models.Finally,in order to facilitate user interaction,the trained model is deployed to a server and an automatic Mongolian question and answer platform is developed.