Research On Intent Detection Of Dialogue System Based On Deep Learning

With the rapid development of the Internet and the popularity of smartphones,the development of the dialogue system has entered an important historical stage.Intent detection is an important task in spoken language understand(SLU)system.This paper further studies and explores from the perspective of multi-turn dialogue document-level intent classification and single-turn dialogue sentence-level intent classification.In recent years,some models started jointly processing intent detection(ID)and slot filling(SF)task.However,most joint models require a lot of manually annotated slot data.Therefore,we propose a joint model for intent classification and named entity recognition.We use the industry-strength natural language processing tool spaCy to generate named entity tags for the benchmark datasets,without need to manually label the named entity recognition tags.Our model can learn more semantic information in named entity tags,which can improve the performance of the model on intent classification tasks.The NER tags are more versatile than slot tags in many different domains and intents.The same NER tag standard can be used in many different domains and intents.On the three benchmark datasets,our joint model of ID and NER achieved better or similar performance than that of ID and SF.In many real-world scenarios,users maybe need to dialogue with the agent multi-turn for the agent to complete one task correctly.That is to say,in multi-turn dialogue,sometimes the user's real intent is only one,and we can also refer to this intent as the original intent.Therefore,we propose a document-level intent detection task in multi-turn dialogue,for detecting the original intent of the user in a multi-turn dialogue.We refer to the labeling system of Harbin Institute of Technology and relabel this dataset with refined classes,enlarging the number of intents from 3 to 21.We explores the different performance between the hierarchical BiLSTM and non-hierarchical BiLSTM model on document-level intent detection task and proposed a hierarchical self-attention model.Experiments have proved that our model achieved competitive performance compared with single-turn single-task model.