Research And Implementation Of Generative Adversarial Intelligent Conversation Robot Based On Attention Mechanism

With the rapid development of hardware and software technologies such as artificial intelligence and intelligent equipment,more and more intelligent interactive robots are applied in industry,medical treatment,education and other fields.The representative intelligent chatbots can freely interact with users through various means such as interface,voice and gesture.Man-machine dialogue is an important branch of artificial intelligence technology.In early dialog systems,machine conversations were set by databases and templates,and application scenarios were limited.With the rapid development of deep learning,people want to be able to communicate with machines naturally.In recent years,researchers have used Transformer model to implement an open domain dialogue system and achieved certain research results.However,due to the complexity of natural language and the different requirements of application scenarios and professional fields,the existing methods still have shortcomings or defects,such as the lack of coherence of the replies generated by the model,which is difficult for users to understand.It is easy to generate security reply,but the semantic information of generating reply is small.Although the large general model has made a certain breakthrough,it is difficult to continuously expand the model due to the large reference number and training data set,and it is not universal in the application of small and medium-sized robot systems.To solve the above problems,this paper systematically analyzes the methods of traditional dialogue generation model.In order to improve the generation and reply quality of the dialogue system,the generated dialogue is deeply studied and the following main work is done:(1)Aiming at the large gap in similarity between generated dialogue and real dialogue,a dialogue model based on generated adversarial network Seq2 Seq is proposed.Firstly,LSTM and GRU networks are used to construct generators,which effectively retain semantic information.The encoder compresses the long distance sequences into fixed intermediate vectors,and introduces attention mechanism in the decoder part of the model to enhance the semantic expression ability.In order to better automatically discriminate the similarity between generated dialogue and real dialogue of the model,the discriminator using RNN as classification function is improved.The ability of the decision finder to recognize true and false dialogues is constantly improved through training,and the partial sequence generated for the whole reward and decision is generated by using Monte Carlo search and strategy gradient algorithm.Improve the similarity between generated statements and real statements to avoid safe answers.The experimental comparison between Little Yellow Chicken corpus and some special service terms shows that the introduction of generative adversarial network makes the dialogue more close to human dialogue.(2)Aiming at the problem that traditional Seq2 Seq is insufficient in extracting semantically relevant features from multi-round and long sentence generation dialogues and poor in response expression,a generative antagonistic TGAN model based on attention mechanism is proposed.Firstly,multi-layer Decoder structure of Transformer model is adopted to construct generator part.In the coding stage,position coding and word vector fusion are used to make the model can be trained in parallel.Using multi-head attention mechanism and self-attention mechanism in Decoder structure,more semantic features can be learned in different high-dimensional Spaces.The mask mechanism covers up the part that does not appear in the sentence to improve the task of dialog generation.In addition to the basic discriminator,artificial reinforcement learning evaluation mode is added to the generative adversarial network of the improved model.By adding the output positive and negative feedback mechanism to the generated dialogue,the discriminator is improved to conform to the human evaluation,and the generator parameters are adjusted by rewards.After several rounds of interaction between the model and the user,the dialogue effect can be generated in accordance with human approval.In the experiment,comparison between automatic evaluation and manual evaluation with Transformer model proves the effectiveness and feasibility of joining the generation countermeasure network,and verifies that the generation and recovery quality of the model proposed in this paper is higher.(3)According to the improved dialogue model,the human-computer interaction system based on robot hardware and software is realized.Based on the dialogue system of TGAN model,combined with various technologies such as speech recognition,speech synthesis and computer network,real-time human-computer dialogue is completed,and various questions are answered through the user’s instructions.In the process of robot interaction,the data collection function of man-machine dialogue is realized,which provides strong support for the subsequent man-machine dialogue learning.