A Joint BERT-based Approach with Explicitly Incorporated Slot into Intent
Xincheng Tang, Yiyang Chen · 2023
Intent detection and slot filling are two prominent tasks to capture the conversational utterance. Extant studies have shown that joint learning models can promote the two tasks mutually. However, the previous joint models mainly think over the guideline of intent detection for slot filling and show little improvement about the opposite way. This study fully considers the guidance of slot filling applying on intent detection and proposes a joint approach on the basis of BERT (Bidirectional Encoder Representation from Transformers) with max pooling layer and CRF (Conditional Random Field) for intent detection and slot filling. The joint approach fuses the max-pooled slot embedding representation with the beginning of the sentence token vector for intent detection and replaces the softmax layer with CRF for slot filling. The comparative experiments by using two common public datasets demonstrate that the proposed joint model achieve competitive performance for both tasks.