Research And Application Of Slot Filling In Task-oriented Human-machine Dialogue System

In the dialogue understanding of human-machine dialogue system, the slot filling is an important task. The slot filling extracts information related to task from users' utterances. The performance of slot filling has an important influence on the quality of the whole dialogue system.Slot filling has received extensive attention and research, from the early adoption of template methods to the current use of statistical sequence labeling model. In recent years, deep neural network models are also used to solve the slot filling task. However, there are still many problems need to be further studied. Aim at two problems, this paper carried out the following work:First, the performance of sequence labeling models such as conditional random field and recurrent neural network model in the slot filling task remain controversial, this paper systematically compares and analyzes the performance of several sequence labeling models in slot filling tasks, and get some more rich and clear conclusions in the experiments of three dialogue corpus.Second, the features used in the present slot filling methods are often sentence internal features such as word, part of speech and NER features.While other sentence external features are rarely involved. Actually, some sentence external information such as dialogue history information has important value to the slot filling task, sentences from last turn always have supplementary information of this turn. This paper puts forward to use several sentence internal and history features in the slot filling task,experiments show that these features have a significant performance improvements.According to the experimental results, we train a new slot filling model and apply it to the meeting room reservation dialogue system.