Research On Intent Classification In Dialogue Systems

With the development of mobile internet and artificial intelligence(AI),people more desire to interact with computers by natural language.The human-computer dialogue systems will greatly facilitate people's life and work,and have attracted increasing attention.Building dialogue systems usually requires great effort,because natural language understanding(NLU)and generation modules are all essential to conduct conversations.NLU typically involves identifying speaker's intent,one task that are often referred as intent classification,which plays an important role for dialogue systems in order to understand the user's utterance.In recent studies,deep learning is extensively used in Natural Language Processing tasks,which also can be applied to intent classification.This paper studies and analyzes the basic principles of traditional machine learning classification method,deep learning model and character-level embedding.Using deep learning to solve the problem of intention classification has been deeply studied.The main research work is as follows:Firstly,for the lack of high-quality Chinese datasets in the intent classification task,a Chinese dataset of intent classification is built in this paper based on real-world conversational texts which collected by HUAWEI NOAH'S ARK LAB and uploaded to the GitHub community for reference by researchers.Secondly,to represent dialogue texts,in this paper,pre-trained Chinese character-level embedding has been trained by word2vec tools on corpus derived from Wikipedia articles.Character-level embedding can effectively avoid problems caused by out-of-vocabulary words and word segmentation errors.Our research and experiment notices that character-level neural embedding is a good alternative for deep neural network intent classification model on Chinese datasets.Lastly,this paper presents a model to classify the intent of a dialogue utterance.This model is named as Character-CNN-BGRU which combines hybrid of Convolutional Neural Network and bidirectional gated recurrent unit neural network architecture based on the character-level neural embedding.The model is able to recognize and classify user's dialogue intent in an efficient way.First,convolution is used to extract local features from each utterance.Second,using a max-pooling layer to select the most important latent semantic factors,in the meanwhile,bidirectional gated recurrent unit(BGRU)layer architecture can capture the contextual semantic information.So that this model can utilize more higher-level and abstract features to perform classification.We evaluate the proposed models on two datasets.The experimental results show that our proposed model outperforms the baselines.