Font Size: a A A

Research On Adversarial Learning For Dialogue Generation

Posted on:2019-09-03Degree:MasterType:Thesis
Country:ChinaCandidate:H P ZhangFull Text:PDF
GTID:2428330548966894Subject:Computer system architecture
Abstract/Summary:PDF Full Text Request
In recent years,deep learning has made great achievements in the field of natural language processing and image processing,making the intelligent dialogue system no longer satisfied with rule-based or retrieval-based methods.The corpus-based dialogue generation model is more intelligent and more general than the rule-based and retrieval-based dialog system.This thesis focuses on generating a coherent,enjoyable,and engaging conversation that responds to user input.Based on the RNN generative model,we added a discriminator that can discriminate between real data and synthetic data,and used adversarial ideas to make the generators better.In which the generator strives to improve itself to generate data sufficient to confuse the discriminator,the discriminator also constantly improves itself and strives to distinguish the real data from the synthetic data.This adversarial training greatly enhances the performance of the generator.Since the original GAN cannot fine-tune discrete data,we combine reinforcement learning methods to conduct adversarial training on dialogue generation.We treat the generated sequence as the current state,and treat the next token to be generated as an action.The state transition that occurs after an action is considered as a strategy.We also use the Monte Carlo tree search to complement the various possibilities of each action,so that the discriminator can score the complete sequence to generate reward and then pass it back to the generator,which is update the generator by the policy gradient method.Then,Based on the King's Glory data set,this thesis first built a solid knowledge base and a relational library for the game,and collected a wealth of data on the discussion board of the game.An intelligent dialogue system that combines different response strategies is designed and implemented.The system combines a variety of strategies including rule-based,retrieval-based,and adversarial training generation models to form a non-task-oriented Mixture model intelligent dialogue system for specific data sets,which can have a good conversation on almost all topics in the King's glory game with human.
Keywords/Search Tags:Chatbot, GAN, RNN, Reinforcement Learning
PDF Full Text Request
Related items