Research On Context-based Dialogue Sentence Understanding And State Tracking

The task-based multi-turn dialogue system is designed to help users accomplish specific goals through human-computer interaction.The machine in the system determines the user’s specific intentions and mission points through multiple rounds of dialogue and interaction with the user,and at the same time gives feedback,and finally completes the purpose(such as booking tickets,querying the weather,etc.).The subtasks can be divided into dialogue sentence understanding tasks(Spoken Language Understanding,SLU),dialogue state tracking(DST)and natural language generation(NLG).Dialogue Sentence Understanding(SLU)is the core technology in the task-based dialogue field.Its main task is to understand the semantics of dialogue sentences,and perform intent recognition and slot filling.It is a prerequisite for the task-based multi-round dialogue to start.Dialogue state tracking(DST)is an important module in the task-based dialogue system.It inherits the dialogue state understood by the dialogue sentence understanding(SLU)and passes it to the subsequent natural language generation(NLG).It is the strategy management part of task-based multi-round dialogue.The task-based multi-round dialogue system can be widely used in smarter human-computer interaction fields such as smart robots and smart hardware,so it has strong research value and practical value.In recent years,with the advancement of pre-training models for natural language understanding and more high-quality multi-round dialogue data sets have been proposed,task-based multi-round dialogue systems have made considerable progress.Despite such remarkable achievements,there are still some challenges: 1)The current model does not distinguish between effective information and ordinary information in the contextual information in multiple rounds of dialogue.Insufficient consideration of timing information 3)The current model cannot effectively use the position of the key information in the dialogue sentence when tracking the dialogue state.4)The current model fixes the importance ratio of the dialogue sentence and the system response in the expression of the dialogue state.The expression of the location.Based on the above problems,this article focuses on dialogue language understanding and dialogue state tracking in task-based multi-round dialogue,and conducts a number of key method research and experimental analysis from dialogue state transfer,continuous sample training,and key information location prediction tasks.The main contributions of this paper are as follows:First,the paper proposes a network based on dialogue hidden state transfer,which can effectively improve the accuracy of slot filling and intent detection in multi-turn dialogues.Multi-turn dialogue is challenging because semantic information is not only contained in the current sentence,but also in the context of the dialogue.In fact,understanding multiple rounds of dialogue is a dynamic process.As the number of conversations increases,the user’s understanding is also changing.In response to the above problems,this paper proposes a network based on dialogue hidden state transmission(USET).The model uses the extracted understanding information as prior knowledge to analyze the semantics in the dialogue context.The USET model first extracts the hidden state of the previous sentence as the previous understanding information.Then pass the understanding information to the next round to assist in understanding the next round of dialogue.At the same time,the model proposes a continuous sample training method in which all sentences in multiple rounds of dialogue are trained in batches according to time sequence,and the multiple rounds of dialogue are understood as a whole.The experimental results on the Stanford University multi-domain data set show that the method has a greater performance improvement compared with the previous model,which verifies the effectiveness of the USET model.Second,the paper proposes a joint learning model for dialogue state tracking based on position importance prediction,which can improve the accuracy of dialogue state value.The existing dialogue state tracking(DST)equally treats the three types of information,namely,user utterance,system response,and ontology.This approach limits the expression of key information in user sentences and system responses.Traditional methods treat user sentences as a whole to filter all possible dialogue states in the ontology,but seldom pay attention to the position of key information(such as slots)in the sentence.In fact,due to human grammatical habits,when expressing the same intention,the position of key information in the sentence structure is relatively fixed.In response to the above problems,this article proposes a dialogue state tracking method based on location importance prediction(TMask).The additional masking information prediction task set by the TMask model and the original DST model are jointly trained to learn where the key information exists,and further use The learned location information influences the semantic expression of user sentences,and dynamically adjusts the expression of user sentences and system responses.On the Wo Z2.0 data set,the performance of TMask surpasses most models.At the same time,on the DSTC2 data set,the TMask method also has a greater performance improvement than the previous model.