Research Of Knowledge-grounded Dialogue Response Generation

As the key to human-computer intelligent interaction,the dialogue system has received extensive attention from academia and industry due to its important practical value.Natural Language Generation(NLG),as a key technology in the dialogue system,determines the quality of the dialogue system's output responses.Therefore,the performance of the NLG model largely affects the degree of intelligence of the dialogue system.However,due to the lack of knowledge guidance,traditional NLG models usually tend to generate general safe responses that lack information,which seriously affect the dialogue experience.Knowledge-grounded dialogue response generation technology introduces knowledge information into the context of the dialogue,and guides the model to generate responses containing knowledge information,which has become a popular direction in the reseach of NLG technology in dialogue system.This article mainly focuses on knowledge-grounded response generation technology in dialogue system,conducts research and experiments on it.Firstly,this article introduces the related technologies of NLG,including word embedding,Transformer and pre-training model.Then,this article introduces the multi-domain multi-turn dialogue dataset Kd Conv that integrates knowledge information.At the same time,this article introduces the commonly used automatic evaluation metrics of NLG methods,such as Hits@K,BLEU,Distinct,etc.,as well as the human evaluation metrics – fluency and correlation.This article uses traditional retrieval and generative models to conduct experiments on the Kd Conv dataset,and uses the performance of traditional models on various metrics as the baseline.Then,this article proposes two novel models,one is called KT-GPT that based on knowledge-tree and GPT,another is KC-GPT based on knowledge-copy and GPT.KT-GPT introduces knowledge triples to input sequence through knowledge-tree structure and encodes the knowledge to representation vector implicitly,to avoid the problem of low information.In order to improve the interpretability of the model,KC-GPT adjusts the encoding of the input sequence of KT-GPT and uses the tokens of knowledge triples explicitly in the generated responses,to solve the problem of OOV words in traditional generated responses.Finally,this article uses the Kd Conv dataset to conduct experiments and analyses on KT-GPT and KC-GPT.Compared with traditional NLG models in the dataset including three different fields,KT-GPT imporves at the automated evaluation metrics Hits@K,BLEU and Distinct,and the average improvement rate has reached 56.82%,65.96% and 36.64%,respectively.As the same time,the average improvement rate of the manual evaluation metrics – fluency and correlation has reached 6.52% and 34.49%,respectively.Compared with the KT-GPT,KC-GPT has a further improvement.The average improvement rate of BLEU and Distinct are 3.19% and 2.83% respectively,and the average improvement rate of fluency and correlation are 3.24% and 2.26%.Through the analysis of actual examples,it is found that the responses generated by the KC-GPT model in a specific context are closer to the human level.