The Research Of Multi-Turn Dialogue System Based On Scene Context-Awareness

Dialogue system,also known as chatbot,virtual assistant,which provides services for users in the form of dialogue,such as search,ordering,shopping and so on.In recent years,with the development of deep learning,recurrent neural network-based dialogue models have gradually become the mainstream trend.The end-to-end Seq2 Seq model automatically encodes the input as a contextual representation and then decodes it into its corresponding output,yielding significant results in dialog systems.However,it can only capture the contextual information of the current input,while the context information of multiple conversations cannot be obtained,therefore it cannot handle the multi-turn dialogues.In multi-turn dialogue systems,extracting the global contextual information of dialogues in real time and giving a reasonable response plays an important role in providing long-term and friendly human-computer interaction.During long-term conversations,it usually involves the switching of multiple dialogue scenes.Inconsistencies in the context of different scenes make that not all contextual information is conducive to the generation of current responses.Therefore,apperceiving the change in dialogue scene in real-time,and making the correct response to the current scene,is a serious problem that needs to be considered in multi-turn dialogue systems.The purpose of this paper is to implement a non-goal-driven multi-turn dialogue system that can fully understand the context of the conversation based on the end-toend deep neural network technology.The research work is divided into the following aspects:Firstly,a fixed-scene multi-turn dialogue model based on hierarchical encoding and neural attention mechanism is constructed on the basis of multi-turn dialogues in a fixed scene,to learn the sentence-level and scene-level context vector representations of the dialogue.In a fixed scene,the context of the scene changes as the conversation progresses.Scene-level context vector vectorizes the global context of a conversation,which is a key element of dialogue scene context awareness.Secondly,a scene-aware multi-turn dialogue model based on scene context is constructed,for multi-turn dialogues of long-term and multi-scenes.The scene context vectors are recorded periodically in a long turn of dialogues.Different scenes are assigned with different weights according to the scene context similarity,and the context information of multiple scenes is combined to generate the response which is close to the scene context.Thirdly,a multi-turn dialogue corpus with scene segmentation is constructed based on the movie and TV scripts,to solve the problem of the lack of open corpus.Based on the crawler technology and text preprocessing method,the script is converted into structured multi-turn dialogue data.Through comparative experiments,the model proposed in this paper has achieved good results in a number of evaluation indicators.