Bi-directional Joint Model for Intent Detection and Slot Filling
Bin Luo, Baiming Feng · 2024
Natural language understanding plays a very important role in textual information processing systems, and is a necessary module for systems such as recommendation, question and answer, and search. Intent detection and slot filling are two crucial tasks for natural language understanding. Traditionally the two tasks were proceeded independently. Recently some studies have shown that the two tasks are correlative strongly, and some joint models have achieved better performance. However, they still have difficulty in capturing the mutual information between the two tasks adequately. To address the underutilization of mutual information in the above two tasks, this papers proposes a new bi-directional joint model for intent detection and slot filling, which contains an input embedding layer, a slot enhancement layer and an intent enhancement layer, where the input embedding layer uses BERT as the underlying encoder, the slot enhancement layer is to add the intent information of the whole sentence as a guideline to establish a deeper connection between the slot and the intent in the process of slot filling, and the intent enhancement layer is similar. Experimental results on Snips dataset demonstrate that our model achieves state-of-the-art results in slot F1(97.1%) and sentence-level semantic frame accuracy(93.1%), indicating that our model is able to capture the mutual information between the semantic components in the NLU tasks more adequately.