Research on the Joint Learning Method of Intent Detection and Slot Filling by Fusing Tag Semantic Information

Intention recognition and slot filling are the main functions of natural language understanding. In view of the insufficient consideration of the semantic information of intention and slot label in most current models, this paper proposes a method to integrate the semantic information of intention and slot label into the joint model. First, in order to make full use of the semantic information of intention and slot, we transform intention, slot label and label "B, I, O" into natural language; Secondly, the pre-training model BERT is used to encode the natural language and interact with the coded representation of the sentence, so that the sentence fully contains the intention and semantic slot information, and the intention and semantic slot that best match the sentence and word are output. Finally, we conducted experiments on three data sets, ATIS, SNIPS and Facebook. The experimental results show that the method proposed in this paper can effectively use the semantic information of tags, and the accuracy of intention recognition and F1 value of slot filling are improved, which verifies the effectiveness of the method in this paper.