Enhancing Joint Multiple Intent Detection and Slot Filling with Global Intent-Slot Co-occurrence

Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 7967 - 7977 December 7-11, 2022 ©2022 Association for Computational Linguistics Enhancing Joint Multiple Intent Detection and Slot Filling with Global Intent-Slot Co-occurrence Mengxiao Song, Bowen Yu, Quangang Li∗, Yubin Wang, Tingwen Liu, Hongbo Xu Institute of Information Engineering, Chinese Academy of Sciences. Beijing, China School of Cyber Security, University of Chinese Academy of Sciences. Beijing, China {songmengxiao,yubowen,liquangang}@iie.ac.cn {wangyubin,liutingwen,hbxu}@iie.ac.cn Abstract Multi-intent detection and slot filling joint model attracts more and more attention since it can handle multi-intent utterances, which is closer to complex real-world scenarios. Most existing joint models rely entirely on the train- ing procedure to obtain the implicit correlation between intents and slots. However, they ig- nore the fact that leveraging the rich global knowledge in the corpus can determine the in- tuitive and explicit correlation between intents and slots. In this paper, we aim to make full use of the statistical co-occurrence frequency between intents and slots as prior knowledge to enhance joint multiple intent detection and slot filling. To be specific, an intent-slot co- occurrence graph is constructed based on the entire training corpus to globally discover cor- relation between intents and slots. Based on the global intent-slot co-occurrence, we propose a novel graph neural network to model the inter- action between the two subtasks. Experimen- tal results on two public multi-intent datasets demonstrate that our approach outperforms the state-of-the-art models. 1 Introduction Spoken language understanding (SLU) aims to en- able the machine to understand user’s utterance, which contains two typical subtasks: intent de- tection and slot filling (Tur and De Mori, 2011). Intent detection obtains the user’s intent from the input utterance and slot filling recognizes entities carrying detailed information of the intent. Gan- gadharaiah and Narayanaswamy (2019) found that complex scenarios of real-world may often include multiple intents in an utterance. Take an example in Figure 1, the utterance indicates two intents: "GetWeather" and "BookRestaurant". Jeong and Lee (2008) suggested that solving the two subtasks jointly makes better perfor- mance because of the correlation between in- ∗Corresponding author. Utterance Tell me the weather in timbo and book a bakery ↓ ↓ ↓ ↓ ↓ ↓ ↓ ↓ ↓ ↓ Slot �� �� �� �� �� ������� �� �� �� ������������������ Intent ��������������������������� Figure 1: An example of joint multiple intent detection and slot filling for an utterance. tents and slots (e.g. "BookRestaurant" and B-restaurant_type). Therefore, re- searchers in recent years (Gangadharaiah and Narayanaswamy, 2019; Qin et al., 2020, 2021b) pay more attention to joint multiple intent detec- tion and slot filling. Gangadharaiah and Narayanaswamy (2019) first tried to explore the joint model for multiple intent detection and slot filling based on attention with a slot-gated mechanism to model dependencies be- tween intents and slots. Qin et al. (2020) proposed an adaptive graph-interactive framework (AGIF) where the core is an adaptive intent-slot interaction multilayer graph attention network. It makes it pos- sible for each token to capture different relevant intent information, thus …

Content truncated to 500 words