Joint Training Model of Intent Detection and Slot Filling for Multi Granularity Implicit Guidance

Abstract

Spoken Language Understanding (SLU), which includes intent detection and slot filling, is an important module in task-based dialogue system. Most traditional models only take single intent and guide the slot filling explicitly. However, in a real conversation scenario, users may have more than one intentions in a sentence. In this paper, we propose an implicit guidance framework for joint intent detection and slot filling. We use token-level and sentence-level intent information to guide slot filling. The token level realizes fine-grained intent guidance, and the sentence level realizes coarse-grained intent guidance. The experimental results show that compared with the baseline model, our model improves the accuracy of sentence-level semantic framework on both multiple intents dataset MixSNIPS and single intent dataset SNIPS by 1.2% and 0.7%, respectively.