| With the rise of artificial intelligence in the past two years,voice intelligent chat robots have become sought after objects of all walks of life,Among them,a task-based dialogue system with a strong domain is widely used.Including the bank’s intelligent assistants,hospital guidance robots,mobile communication business handling robots,etc.,all reflect the increasing demand for task-oriented human-computer dialogue systems.The natural language understanding module of the human-computer dialogue system is an important work in the human-machine dialogue system.This module includes intent recognition and semantic slot filling.The accuracy of the intent recognition directly affects the user’s experience of the system,that is,whether the user can get reasonable information from system feedback.Based on deep learning technology,this paper does the following work on the problem of intent recognition in task-based human-computer dialogue system:First of all,for the general deep learning method,when performing intent recognition,it is mainly based on the text feature representation of a single word vector,but cannot fully express the information contained in the text.A model MCCL(Mulity Channels CNNLSTM)is proposed.This method is based on deep learning model Convolutional Neural Networks(Convolutional Neural Networks,CNN),combined with the LSTM(Long-Short Term Memory)algorithm idea,uses a variety of features to represent the text to reduce the impact of word segmentation errors and inaccurate semantic understanding.The MCCL model can effectively overcome the single problem of general feature representation,thereby improving the accuracy of intent recognition.Secondly,aiming at the problem that the original word context semantics are lost when the text is classified with the label information,and the serious problem of information loss in the CNN max pooling layer,an intent recognition model LICL(with Lable Information Cnn-lstm)based on the improved CNN pooling layer is proposed.In this model,words and labels are embedded in the same space,so that text and labels share the representation of words.In order to solve the serious problem of data loss of the max pooling function in CNN,a method of combining max pooling and average pooling is proposed,which provides a new solution for intent recognition task.Finally,the effectiveness of the proposed method is verified through experiments. |
