Natural Language Understanding Model Based On Deep Learning

Due to the progress of technology and the increase of labor cost,the demand for high-performance intelligent dialogue system in daily life is increasing.As an important part of intelligent dialogue system,natural language understanding model has become a research hotspot in scientific research and industrial application.Natural language understanding is mainly divided into two tasks intention recognition and semantic slot filling.Previous work has proved that joint processing of the two tasks will achieve better results.In this paper,we propose a model of joint processing of two tasks,and explore the methods to improve the performance of the model.In this paper,through the comparison with the traditional structure such as recurrent neural network,the special model structure of S-LSTM can extract the word hidden state and global hidden state at the same time,strengthen information exchange between overall and local information,and meet the needs of joint processing intention recognition and semantic slot filling task.Therefore,S-LSTM model is innovatively introduced as the basic model of this paper to deal with the Natural language understanding task.On this basis,this paper finds that the origin S-LSTM model's data preprocessing is relatively rough,and it can not effectively distinguish the influence degree of different words when refining the global state.So we introduce refined data preprocessing and attention mechanism to solve these problems.These measures make the model performance further improved.In the above model,many methods are used to improve the accuracy of slot filling task.For the intention recognition task which is often more important in practical use,the final classification result is obtained by passing the global hidden state through softmax function.This kind of processing method is relatively rough.Therefore,this paper continues to explore the methods to improve the accuracy of intention recognition.Based on the mature feature extraction network,this paper attempts to explore methods to improve the accuracy of intention recognition.First of all,this paper creatively proposes the idea of using the semantic information contained in the tag.In order to preliminarily verify this idea,this paper obtains the tag embedding based on the simple splicing of the word embedding vector.Although the experimental results have declined,it still has practical significance,which opens up ideas for the follow-up research.Then,in order to enhance the ability of network to deal with extraterritorial intention,this paper introduces outlier detection LOF algorithm to detect possible outlier statements in the test phase,so as to enhance the practical value of the model.The model proposed in this paper not only has good natural language understanding task performance,but also has certain practical application ability.At the same time,it also provides ideas for the follow-up research in the process of research.