A Task-oriented Multi-round Dialogue Semantic Understanding Model

Abstract:
Task-oriented multi-round dialogue is one of the hot topics in the field of artificial intelligence in recent years. Due to the time sequence relationship of the dialogue context, the possible change of user intentions and insufficient data, the existing models have the disadvantages of poor reliability and scalability. In this paper, an end-to-end task-oriented multi-round dialogue semantic understanding model PIENet is designed based on the encoder-decoder modeling idea, in which the encoder uses a single-layer bi-directional GRU and the decoder uses three standard GRUs to resolve the confidence domain, behavior domain and system response respectively. In the data preprocessing stage, rich state-behavior pairs are enhanced to improve the diversity of system response. The semantic information is enhanced by embedding the relative position vector between tokens into the word vector combined with the attention mechanism. Experimental results based on MultiWOZ2.3 dataset show that the model achieves high conversational fluency and rationality while ensuring certain accuracy of domain, intention and slot identification, and has a good performance improvement.