Humanoid Robot Dialogue And Sentiment Analysis Model Based On Deep Neural Network

The traditional humanoid robot dialogue system is generally based on ‘pattern matching',which can make reasonable answers in the designed dialogue topics,but can't generate good responses to the content outside the domain.The system's dialog rules rely on manual design and can't detect the emotional state of an interactive object.Aiming at the shortcomings of traditional methods,a sentiment analysis model based on deep neural network is designed to detect the emotions of interactive objects,and an open domain dialogue system of humanoid robots is constructed.The main research contents and innovations of this paper are as follows:(1)Text-based sentiment analysis model.The “Sentiment Classification with Deep Learning Data Set” published by the International Conference on Natural Language Processing and Chinese Computing(NPLCC)in 2013 and 2014 is used as the sentiment analysis model dataset.This paper evaluates the classification of recurrent neural networks,convolutional neural networks,support vector machines,naive Bayes and other algorithms on the dataset.The advantages and disadvantages of the algorithms and the correlation of the classification results are compared.In order to integrate the advantages of each algorithm,the soft voting strategy and Stacking strategy are compared in this paper.Finally,the sentiment analysis model is built through Stacking strategy with an F1 value of 0.804,which can detect the emotional state precisely.(2)Multi-indicator dialogue cleaning model based on attention mechanism.Aiming at the defects of low quality and high noise of the dialogue database on the network,this paper proposes a multi-indicator dialogue cleaning model based on attention mechanism for unsupervised cleaning of dialogue data.The innovation of the model is integrating the two layers attention mechanism and using multiple indicators to judge whether the question and answer pairs have semantic relevance.The experimental results show that the multi-indicator model has higher accuracy in the current data set than the traditional model,which can reduce the time cost of data cleaning.The heat map is used to visualize the weight of the attention mechanism,which verifies the good interpretability of the model.(3)Open domain dialogue system.A dialogue model based on sequence to sequence(Seq2seq)for humanoid robot is constructed in this paper.In order to solve the problem that the model will generate meaningless answer,the maximum mutual information is used as the objective function for the dialogue model in this paper.And the dialogue system is constructed by integrating dialogue model,sentiment analysis model,speech recognition module and speech synthesis module.Experiments show that the dialogue system can generate human-like answers and conduct emotional monitoring of interactive objects during the conversation.