Research On Intent Detection Method Based On Capsule Network

In recent years,with the continuous development of artificial intelligence technology,man-machine dialogue system has been widely concerned.Spoken language understanding is a vital part of the humanmachine dialogue system,which aims to enable the machine to understand the user?s specific needs and give reasonable answers.In order to clarify the user's expressed intent requirement,intent detection is a key sub-task in spoken language understanding,and the accuracy of intent detection is directly related to the performance of semantic slot filling,and it is helpful for the following research of dialogue system.Considering the frequent interaction between human and machine and the diversity of user's intent expression,sometimes the user?s discourse contains not only one intent,but multiple intents.So the intent detection task can be divided into two tasks: single intent detection and multi-intent detection.The thesis analyzes the research of traditional machine learning methods and deep learning methods on the single-intent detection task and multi-intent detection task,and further considers how to reasonably apply the deep learning model to the intent detection tasks.Capsules in the capsule network contain rich feather information.Dynamic routing can dynamically learn the relationship between neural network layers,it can not only retain the semantic features with low occurrence probability,but also has the ability to fit the features well,which is suitable for small data sets.Therefore,this thesis mainly uses capsule network to study the single-intent detection task and multi-intent detection task respectively.The specific work is as follows:(1)Aiming at the problem that the pooling operation in convolutional neural network cannot fully utilize all the feature information in the sentence.This thesis uses the capsule network to solve this problem and applies it to single intent text.By using dynamic routing algorithm,all feature information in the intent text is reasonably dynamically allocate to the intent capsule,thus completing the single intent detection task,and comparing its performance results with the traditional machine learning method and various deep learning methods on the single intent detection task,further prove the advantages of the capsule network.(2)Considering the scarcity of user's multi-intent expression data,this thesis collects the Chinese and English multi-intent test sets based on the single-intent tags respectively,and constructs the multi-intent classifiers based on the single-intent tags for multi-intent detection.At the same time,the capsule network is used for the multi-intent detection task.On the one hand,in order to ensure feature quality of intent text,this thesis proposes to add convolution capsule layer to capsule network to extract the deep semantic information of the intent text,and use dynamic routing algorithm in capsule networks to dynamically allocated high-level feature capsule to the intent capsule categories.The probability of multiple intents is determined by setting threshold value,thus promote the accuracy of multi-intent detection.On the other hand,on the basis of adding convolution capsule layer,this thesis proposes to use three different convolution kernels to extract semantic information of different phrase collocations in sentences,and explore the influence degree of different convolution kernels on the multi-intent detection performance results.Experiments results show that the better multi-intent classification performance can be obtained when n-gram value is 3.The improvement of the performance of multi-intent detection is helpful to the research of spoken language understanding and the construction of dialogue system.