Algorithm Research For Dialogue System Based On Deep Learning

The dialogue system is a kind of computer intelligent agent,which can carry on the fluent conversation with human beings,and plays an important role in the development of machine intelligence in the future.There are many tasks in the field of dialogue system research.The thesis mainly studies system how to automatically respond to the user's speech in a chitchat.The model of dialogue system response algorithm mainly includes retrieval model and generation model.The thesis proposes an improved algorithm from these two perspectives.1.Retrieval model.The retrieval model mainly encodes the sentences,and then selects the answers with higher scores in the answer library as the output through matching.Therefore,the encoding and matching of the sentences are particularly important.In the thesis,a BERT based multi-round dialogue retrieval model is proposed,which encodes the context statement and the response to be matched in the presentation layer by combining the word embedding and position embedding of the statement.In the selection layer,the key information of context is selected.In the thesis,the first state output by BERT module is used as the filter and the other states are used as the encoding to realize the selection function and encoding function.Based on the idea of interaction,the matching layer matches the encoding of the response to be matched with the encoding of the statement in the context to form a matching matrix,and splices multiple matching matrices into a 3D matrix.The aggregation layer uses convolution pooling to compress the information of the 3D matrix and calculate the matching score.The experimental analyzes and compares the performance of the existing excellent baseline model with the model in the thesis,and finds that the model in the thesis is superior to some of the existing models.BERT module is used in the retrieval model to extract important features.Compared with the existing excellent models,the structure is simpler and clearer.The selection layer can ignore irrelevant information and noise in the conversation history,making the model more efficient.2.Generative model.The generative model is mainly based on the encoder-decoder framework to generate appropriate sentences.The thesis proposes a multi-round dialogue answer generative model combined with retrieval model,which not only overcomes the problem of the retrieval model's answer deviation,but also improves the "security" problem of the general generation model's answer.The model retrieves the response R' and the corresponding context C' at the retrieval layer for a given context C using the aforementioned retrieval model.The edit vector layer computes the delete and insert vectors based on the context C and context C'.These vectors respectively delete the useless information in the response R' in the rewrite layer and extract the key information in the context C,and then input it into the decoding layer to calculate the response R.The experiment in the thesis analyzes and compares the performance of the existing excellent baseline model and the model in the thesis based on the two evaluation criteria of information and diversity,and finds that the model in this paper is superior to most of the existing models.To further explore the sources of model performance improvements,the relationship between the quality of response R and contextual similarity,the case of the test model,and the extent to which the model replicates retrieved responses R' were investigated.Finally,it is concluded that the deletion vector of the model in the thesis can reduce unnecessary duplication in the response,and the model can extract useful information from the retrieved response and generate more interesting response,and it is informative.The model even learned about the relationship between contextual similarity and the distance between two responses.