Intent Detection and Slot Filling from Dependency Parsing Perspective: A Case Study in Vietnamese

Intent Detection and Slot Filling from Dependency Parsing Perspective: A Case Study in Vietnamese Phu-Thinh Pham, Duy Vu-Tran, Duc Do, An-Vinh Luong, Dien Dinh University of Science, Ho Chi Minh city, Vietnam Vietnam National University, Ho Chi Minh city, Vietnam {phpthinh18, vtduy18}@apcs.fitus.edu.vn {dotrananhduc, anvinhluong}@gmail.com ddien@fit.hcmus.edu.vn Abstract Spoken language understanding (SLU) sys- tems using deep learning techniques require ef- fective intent detection and slot filling models. Previous studies in this field taking advantage of sequence to sequence models have achieved good results. However, they focus on the local- ity of words and thus are sensitive to surround- ing terms. In this paper, we introduce a new ap- proach for this problem inspired by the depen- dency parsing techniques via a biaffine model to give the system a global view of the input. The experiments on PhoATIS dataset for Viet- namese have shown that our joint model for in- tent detection and slot filling obtains potential results. 1 Introduction Spoken language understanding (SLU) has been ap- plied to many chatbot applications in recent years. Intent detection and slot filling are two main tasks in this field for building task-oriented dialog sys- tems. The purpose of intent detection task is to classify users’ intent and that of slot filling task is to extract semantic constituents from the natural language utterances (Tur and De Mori, 2011). The most common approach for intent detection task is using a classifier based on [CLS] context repre- sentation. In parallel, the slot filling task is usu- ally considered as a sequence to sequence problem, with the help of conditional random fields (CRFs) and recurrent neural network (RNN). Normally, these two tasks are considered as two distinct tasks, thus implemented separately, although the slots in- tuitively depend on the intent (Goo et al., 2018). Hence, some studies have proposed joint models based on the correlation between two tasks, enhanc- ing the performance of each other (Goo et al., 2018; Dao et al., 2021; Wu et al., 2020; Wang et al., 2018; Chen et al., 2019). Briefly summarized, most of the previous studies take advantage of autoregressive model or sequence to sequence architecture to solve the problems (Wu et al., 2020). For example, conditional random field (CRF) is a common approach for slot filling tasks since it considers the correlations between tags. However, we argue that this approach heavily relies on the locality of words and we need to provide the model with a global view of the input. In this study, we propose a joint model for intent detection and slot filling tasks inspired by the de- pendency parsing technique. For slot filling task, we reformulate it as the task of identifying the span of a slot and assigning its category, following the study of Yu et al. (2020). In parallel, we consider the intent detection task as the task of classifying the intent labels of the span from the beginning to the end of an utterance. Furthermore, we in- corporate the intent context information with an intent-slot attention …

Content truncated to 500 words