Research On Knowledge-driven Dialogue Generation Technology

With the rapid development of artificial intelligence,human-machine dialogue systems are expected to replace the current mainstream input devices and become one of the most common used human-machine interaction methods.Dialogue generation is one of the most important links in a dialogue system.It means that the computer automatically generates a dialogue response based on certain dialogue information.Since the generated results will be directly presented to the user,the quality of the task will greatly affect the user experience.In recent years,the rapid development of deep learning technology has broughtnew opportunities for dialogue generation.The models based on deep learning can learn the characteristics from massive dialogue data and automatically generate corresponding responses.However,in practical applications,people find that the dialogue systems trained by the dialogue data cannot meet the needs of users well.These dialogue systems tend to generate less informative responses,such as "I don't know",this type of safe response often makes people lose the desire to chat.The expression of language and knowledge are inseparable.With the support ofknowledge,language becomes rich and colorful.Knowledge can provide additional useful information to the dialogue system,and introducing knowledge into the dialogue system can alleviate its problem of easily generating safe responses.This paper will study the task of knowledge-driven dialogue generation,that is,in the process of dialogue generation,use knowledge reasonably and incorporate it into deep learning models to make the responses more diverse and informative.The main research work of this paper is as follows:This paper proposes a knowledge dialogue generation model based on theattention mechanism(KDG-Att).The knowledge information required in different dialogue contexts is different,so relevant knowledge needs to be selected.The KDG-Att model uses the attention mechanism to assign different weights to each knowledge,so as to filter out the most relevant knowledge information related to the user messages to assist the generation of response.In order for the response to contain more valid information about the knowledge,in addition to the error between the generated response and the real response,the model's objective function also measures the correlation between the generated response and the knowledge to guide the learning of the model parameters during the training process.As a result,the trained model is more inclined to generate knowledge-related responses.This paper designed a series of comparative experiments to verify the effectiveness of the KDG-Att model.The experimental results show that the model can generate higher quality and more informative responses.This paper proposes a dynamic knowledge dialogue generation model(DKDG-MD).During the dialogue generation process,as response is gradually generated,the words that have been generated will change the state of the dialogue,and the knowledge required will change accordingly.The DKDG-MD model can automatically select the knowledge most relevant to the current state of dialogue,and continuously update the required knowledge as the state of dialogue changes.This paper compares the performance of multiple models and the DKDG-MD model on Chinese and English datasets,and the experimental results show that the DKDG-MD model performs better and the responses generated by the model contain richer information.