Design And Implementation Of Mongolian Speech Interaction System

Voice interaction systems for major languages such as Chinese and English have been applied to all aspects of technology and life,greatly improving the efficiency of information acquisition.The Mongolian people also have a wide range of needs for voice interaction systems,hoping to use Mongolian to interact with smart devices.Therefore,this thesis designs and implements a Mongolian-oriented voice interaction system.This system is composed of Mongolian voice recognition,question and answer system,and speech synthesis.The question answering system is the core part of the voice interactive system,and the performance of the question answering model directly affects the language quality of the answers generated by the voice interactive system.There is no publicly available high-quality Mongolian question answering corpus,and related research on the combination of Mongolian question answering model and deep learning is in its infancy.In response to the above problems,this article has conducted the following research:(1)In response to the lack of Mongolian question and answer corpus,a Mongolian question and answer corpus(Mon QA Corpus)was constructed.This thesis collects a large number of open-field single-round Chinese question and answer corpora,performs a series of processing such as Chinese-Mongolian translation,automatic screening,and manual correction,and builds a high-quality Mongolian question and answer corpus containing about 15,000 pairs of daily question and answer dialogues for follow-up Modeling research of generative question answering model.(2)In order to generate more diverse answers under limited corpus,this thesis uses a sequence-to-sequence-based generative Mongolian question answering model.According to the characteristics of long and short-term memory network(LSTM),a Mongolian question answering baseline model based on LSTM-LSTM is constructed.In order to improve the language quality of the model's generated answers,this thesis introduces a gated recurrent unit(GRU)network structure in the encoder to build a GRU-LSTM question and answer model.It is found through experiments that this model has less than the baseline model in improving the quality of the generated answers.Not obvious.In order to make the answers generated by the question answering model more reasonable and contextual,this thesis is based on four different encoder structures of LSTM,GRU,BiLSTM,and Bi GRU,and adds an attention mechanism to the decoding process of the decoder to construct LSTM-AM-LSTM,GRU-AM-LSTM,BiLSTM-AM-LSTM and Bi GRU-AM+LSTM four question answering models.Through experiments,it is found that the Mongolian question answering model with the GRU-AM-LSTM network structure has the best effect in generating answers.(3)In order to make the question answering model obtain richer semantic expression,this thesis will add the sequence-to-sequence question answering model of the attention mechanism to the final layer of word embedding output of the pre-trained Mongolian BERT model to construct the Bmo GRU-AM-LSTM question answering model.Experiments show that this model can effectively improve the contextual semantic relevance and richness of answers.(4)In this thesis,the BmoGRU-AM-LSTM Mongolian question answering model is applied to the construction of the Mongolian question answering system,combined with Mongolian speech recognition and speech synthesis services,to design and implement an Android-based Mongolian speech interactive system.The test results show that the Mongolian voice interaction system developed in this thesis can perform a single round of Mongolian voice human-computer interaction,and the generated answers have high semantic relevance to the questions.