Research On Dialogue Emotion Generation And Emotion Recognition Based On Deep Neural Network

Building a dialogue system capable of communicating with humans has long been the ultimate goal of researchers.Both the task of emotion dialogue generation and the task of emotion recognition in dialogue are important research directions to achieve this goal.At present,the emotional dialogue generation model has been able to generate sentences containing emotions,but the emotional expression intensity of the sentences is not enough.In this context of specifying emotions,the emotional accuracy of the generated sentences still cannot meet the actual needs.At present,the task of emotion recognition in dialogue is in its infancy,and it has only begun to receive attention in recent years.The accuracy of emotion recognition is also at a low level.Accurately identifying emotional changes in dialogue is also of great significance for the task of emotional dialogue generation.For these two research tasks,this thesis will achieve the following objectives:1.Propose an emotional dialogue generation model based on emotion embedding and emotion reinforcement mechanism,using the Seq2 Seq model as the basic dialogue generation model,in the encoding process,encode the specified emotion category,and then fuse it into the LSTM neural network of the encoder,make the output hidden state and unit state have emotional factors;then in the decoding process,use the same method to fuse the emotional factors into the LSTM neural network of the decoder.In addition to this,an emotion reinforcement mechanism is proposed.The sentiment is reinforced in the dimension of the hidden state of the LSTM neural network.Experiments are carried out on NLPCC2017,XHJ and Opensubtitles datasets,and the results show that the proposed model has significantly improved the emotional accuracy compared with other models,and with greatly improved sentiment accuracy,there are also better results in terms of the quality of the generated utterances.2.A dialogue sentiment recognition model based on commonsense knowledge and a gating mechanism is proposed.The commonsense knowledge in the dialogue is incorporated into the sentiment recognition of the dialogue.The information of the commonsense knowledge is modelled using BIGRU neural network.To further remove unimportant information from the commonsense knowledge,the noise is then filtered using a CNN gating mechanism.The filtered information,dialogue utterances and sentiment information are modelled by a separate BIGRU neural network,and finally softmax is used to complete the task of classifying sentiment in dialogue using this information.Experiments are performed on IEMOCAP,Daily Dialog,MELD and EmoryNLP datasets,and the results show that the proposed model has the best performance compared to other models.3.Design and implementation of a prototype system for emotional dialogue.The system is modelled using Pytorch.Using VUE and JAVA to build the system page and backend respectively.