Group is better than individual: Exploiting Label Topologies and Label Relations for Joint Multiple Intent Detection and Slot Filling

Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 3964 - 3975 December 7-11, 2022 ©2022 Association for Computational Linguistics Group is better than individual: Exploiting Label Topologies and Label Relations for Joint Multiple Intent Detection and Slot Filling Bowen Xing1,2 and Ivor W. Tsang2,1 1Australian Artificial Intelligence Institute, University of Technology Sydney, Australia 2Centre for Frontier Artificial Intelligence Research, A*STAR, Singapore bwxing714@gmail.com, ivor.tsang@gmail.com Abstract Recent joint multiple intent detection and slot filling models employ label embeddings to achieve the semantics-label interactions. How- ever, they treat all labels and label embeddings as uncorrelated individuals, ignoring the depen- dencies among them. Besides, they conduct the decoding for the two tasks independently, with- out leveraging the correlations between them. Therefore, in this paper, we first construct a Het- erogeneous Label Graph (HLG) containing two kinds of topologies: (1) statistical dependencies based on labels’ co-occurrence patterns and hi- erarchies in slot labels; (2) rich relations among the label nodes. Then we propose a novel model termed ReLa-Net. It can capture benefi- cial correlations among the labels from HLG. The label correlations are leveraged to enhance semantic-label interactions. Moreover, we also propose the label-aware inter-dependent decod- ing mechanism to further exploit the label corre- lations for decoding. Experiment results show that our ReLa-Net significantly outperforms previous models. Remarkably, ReLa-Net sur- passes the previous best model by over 20% in terms of overall accuracy on MixATIS dataset. 1 Introduction Intent detection and slot filling (Tur and De Mori, 2011) are two fundamental tasks in spoken lan- guage understanding (SLU) (Young et al., 2013) which is a core component in the dialog systems. In recent years, a bunch of joint models (Liu and Lane, 2016; Goo et al., 2018; Li et al., 2018; Liu et al., 2019a; E et al., 2019; Qin et al., 2019; Zhang et al., 2019; Qin et al., 2021b) have been proposed to tackle single intent detection and slot filling at once via modeling their correlations. In real-world scenarios, a single utterance usu- ally expresses multiple intents. To handle this, multi-intent SLU (Kim et al., 2017) is explored and Gangadharaiah and Narayanaswamy (2019) first propose to jointly model the multiple intent detec- tion and slot filling in a multi-task framework. Qin 0.52 0.32 0.2 0.97 0.95 0.95 Label_B P(Label_B|Label_A) atis_flight_time atis_flight atis_airline B-depart_date.day_name B-fromloc.city_name B-toloc.city_name Figure 1: Illustration of top-3 high-probability occur- ring intent (in blue) and slot (in green) labels when B-depart_date.date_relative (Label_A) oc- curs in the training samples of MixATIS dataset. et al. (2020) and Qin et al. (2021c) further utilize graph attention networks (GATs) (Velickovic et al., 2018) which model the interactions between the embeddings of predicted intent labels and the se- mantic hidden states of slot filling task to leverage the dual-task correlations in an explicit way. And significantly improvements have been achieved. However, we argue that previous works are es- sentially limited by ignoring the rich topological structures and relations in the joint label space of the two tasks because they treat the …

Content truncated to 500 words