Research On Intent Recognition Based On Task-oriented Multi-round Dialogue System

Intent detection is a sub-task of the language understanding,which is the part of the human-machine dialogue system.To understand what users want to express,we need to primarily identify the behavioral intention expressed by them firstly.The most effective method for intent detection is using classification.The aim of intent classification is to classify language queries into a predefined intent category according to the domain of user's natural language or the intent of the expressions.The accuracy of the intent detection is of great importance to dialogue system quality.In the early days,template methods and traditional statistical machine learning techniques have been applied to the study of intent detection tasks.In recent years,deep neural network models have performed well on intent detection tasks.However,differently from the general classification task,in multi-round dialogue,the historical background and contextual information should be taken into account to understand what users want to express.There are still many problems need further research.In this article,the following work is carried out in view of the existed problems in historical background and contextual information in multi-round dialogue.Firstly,aiming at the lack of high-quality annotated corpus for the intent-detection task of the multi-round dialogue dataset,we selected the task-oriented dialogue dataset about restaurant reservation at Cambridge University for labeling.676 dialogue segments are included in this dataset,and a total of 10 tags are labeled.High-quality annotated datasets will be conductive to futher study of intent detection in multi-round dialogue systems.Secondly,existing intent detection methods have only considered simple contextual information,and the memory module can just solve a part of the problem.In this article,we propsed the classification model based on memory network and attention mechanism,which includes the gate unit.Attention mechanism can make better use of the historical information.The ground unit can be used to select whether historical information is needed,which further enhanced the classification effect.The experimental analysis based on the classification model of memory network and attention mechanism proved its validity.Thirdly,the memory module considers the similarity between information supplement and intention between the contexts,and does not consider the dependency between the intentions.In this article,we designed an enhancement module to learn the transfer between intentions by training the Q-learning network.Experiments show that the Q-learning combination neural network model is effective in the intent detection.This have provided a new idea for intent detection.