Research And Implementation Of Modular Dialogue Generation Model Based On Hybrid Neural Network

Dialogue generation has always been a popular project in machine learning and natural language processing.It is one of the core technologies to realize machine intelligence.The dialogue model based on rule setting,information retrieval and data statistics tends to mature and encounters many technical bottlenecks,such as the difficulty of extracting knowledge in specific areas,completing rules and subject to specific fields of expertise.At present,there are already work to combine these models with the deep learning algorithm[9][11]However,neural network model for chatting based on the generation is still less.Compared with the traditional dialogue generation model,model based on the neural network can save a lot of manual participation.However,there are a lot of deficiencies and defects in the existing dialog generation model.In terms of performance,the model uses the vocabulary as a category,and the computational complexity of the model increases greatly as the vocabulary becomes larger.The use of random sampling to reduce the vocabulary sum in training,lead to increase the number of training steps.At present,the neural network models for dialogue generation is a whole.They have complex structure and numerous parameters and train slowly.These factors lead to the huge cost of changing these models.In effect,the existing dialogue generation models tend to output common words and the resulting statement always lack diversity.In addition,the generated statements are often not fluent or semantics unknown.Moreover,the statements often have obvious lexical syntax errors,far from the traditional models in effects.Based on the idea of modularization of conversation generation tasks and neural network technology,in order to solve the performance problem in the dialog generation model and to improve the effect of the generated statements,we studied a data preprocessing process for variable sharing,a vocabulary recommendation method for conversation generation and a comprehensive vocabulary sequence generation method based on recommended sampling in this paper.On this basis,we designed and implemented a modular dialogue generation model based on hybrid neural network.The main contributions of this paper include:(1)In order to solve the problem of sharing variables in neural network,we studied and implemented a data preprocessing process for variable sharing.The study realized the sharing of partial variables of neural network models.The repeated training will be avoided and the amount of mediate data will be decreased.These speed up the training process.To some extent,the reuse of data reduces the waste of resources.(2)In order to solve the performance problem caused by excessive vocabulary in the training process,we studied and implemented a vocabulary recommendation method for dialogue generation.This method effectively defines the vocabulary scope,improves the accuracy of the final generated statement,and provides a more efficient vocabulary-sampling basis.(3)On the basis of the vocabulary recommendation method,we studied and realized a comprehensive vocabulary-sequence generation method based on vocabulary recommendation sampling.Using the vocabulary recommendation as sample method,instead of the traditional random sampling,not only the complexity of the training generated sequence module is reduced from the length associated with the vocabulary to a constant,but also making the model training time-consuming shorter for the same effect.(4)By combining the above main technical methods,we design and implement a modular dialogue generation model based on hybrid neural network.Experiments show that,this modular dialogue generation model avoids using the same statements or commonly statements in result,increases the diversity of statements,reduces the unreasonable statements,and improves the quality of the generated conversations.In addition,modular implementation reduces the overall re-training of the model and the cost of modifying parameters,adjusting structures,and replacing datasets.