Research On Multi-Round Dialogue System Based On Deep Reinforcement Learning

In recent years,man-machine dialogues have been emerging in various fields such as mobile terminals,smart homes,smart medical care,and smart networked cars,etc.Agents that understand human language and can communicate with human beings have become the needs of most people.The task-driven multi-round dialogue is one of the main forms of manmachine dialogue and also one of the difficulties.This paper focuses on the construction of task-based dialog system in the field of life,and focuses on the two sub-tasks of dialog state tracking and strategy management.Based on the advanced achievements of deep learning and reinforcement learning in natural language processing,a robust,high-quality and accurate intelligent multi-round dialog system is realized.The specific work is as follows:1.For the problem of data preprocessing,word vector and character vector are used to represent the text.Word segmentation is based on directed acyclic graph and hidden markov model.Word vector is obtained by Word2 vec model and mixed word vector is obtained by concatenating part of speech and word order.At the same time,we use pre-training BERT model and fine tuning this model to obtain the character vector,and finally the text information will be represented by word vector and character vector.2.To solve the problem of state tracking,the text state is extracted through the Convolutional Neural Network and the switched Long Short-Term Memory.We build and train a CNN model to realize the intention recognition of this dialogue.We build two bidirectional LSTM and input word vector and character vector respectively.Train the shallow neural network to switch the output of the two models,and finally the dialog state tracker with slot filling function and intention recognition function is obtained.The experimental results show that the slot filling model with switch is the best,and the F1 value reaches 97.65%.3.To solve the problem of strategy management,we constructe a deep reinforcement learning model,and the action of dialogue is select through the intermediate state after feature extraction.By using the approximate action value function of a CNN model,two parallel Q networks were constructed and the objective parameters of bellman were solved by adjusting the Q network.Finally,a multi-round dialog model that can select dialog actions based on historical information is obtained.