Research On Open-domain Chatbot Dialogue Generation Algorithm

In recent years,with the rapid development of the Internet and social media,chatbot have been widely used in life.It can not only solve user's communication needs,but also reduce the cost of providing services for users.Therefore,it is of great significance to study the open-domain chatbot dialogue generation algorithm,and it is also one of the hot researches in the present.With the development of deep learning,open-domain chatbot based on Seq2 Seq model have made great progress.However,there are certain gaps in the dialogue generation model.For example,an oversized vocabulary will reduce the dialogue effect,and easily generate non-committal responses and lack diversity,and it is difficult to conduct multi-turn dialogue.In response to the above issues,the main work of this paper is as follows:(1)Under the Seq2 Seq model,use BPE to segment words,which reduce the size of the dialogue vocabulary,and improve dialogue quality.Because the size of the network for the dialogue generation model is limited,and rare words appear in people's daily conversations.if the direct expansion of the vocabulary or use UNK instead of a rare word,it will reduce the dialogue effect.Based on the Seq2 Seq model,this paper uses BPE to segment words and split the words into subword units with a common structure,thereby reducing the vocabulary size.Compared with the unprocessed mode,under the Seq2 Seq model,the dialogue generated after the BPE processing has a higher BLEU scores,and the dialogue generation effect is better.(2)Dialogue generation algorithm based on reinforced adversarial learning.Since conversations consist of discrete words,the resulting Generative Adversarial Nets cannot be directly used for dialogue generation.This paper updates the generator parameters by using the policy gradient so that the Generative Adversarial Nets can be used for dialogue generation.In order to make the generator easier to converge in the training process,this article uses a supervised learning method to assist the generator network update.Compared with the Seq2 Seq model,the Reinforced Adversarial Learning dialogue generation algorithm has better response diversity.(3)Improve the HRED model by using the maximum mutual information as the loss function of the dialogue generation model,and improve the ability of multi-turn of dialogue.Limited to the network structure,the Seq2 Seq model is difficult to conduct multi-turn of dialogue based on contextual history information.At the same time,lack of dialogue diversity in the Seq2 Seq model will also lead to premature termination of the dialogue.This paper improves the HRED model and uses the maximum mutual information as the loss function to reduce the probability of non-committal responses in the dialogue model which increase the diversity of the dialogue,and then increase the number of dialogues turns in the model.Compared to the Seq2 Seq and HRED model,the improved HRED model enables more turn of dialogue.