DARER: Dual-task Temporal Relational Recurrent Reasoning Network for Joint Dialog Sentiment Classification and Act Recognition

Findings of the Association for Computational Linguistics: ACL 2022, pages 3611 - 3621 May 22-27, 2022 c⃝2022 Association for Computational Linguistics DARER: Dual-task Temporal Relational Recurrent Reasoning Network for Joint Dialog Sentiment Classification and Act Recognition Bowen Xing1 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@ihpc.a-star.edu.sg Abstract The task of joint dialog sentiment classifica- tion (DSC) and act recognition (DAR) aims to simultaneously predict the sentiment label and act label for each utterance in a dialog. In this paper, we put forward a new framework which models the explicit dependencies via integrating prediction-level interactions other than semantics-level interactions, more consis- tent with human intuition. Besides, we propose a speaker-aware temporal graph (SATG) and a dual-task relational temporal graph (DRTG) to introduce temporal relations into dialog un- derstanding and dual-task reasoning. To im- plement our framework, we propose a novel model dubbed DARER, which first generates the context-, speaker- and temporal-sensitive utterance representations via modeling SATG, then conducts recurrent dual-task relational rea- soning on DRTG, in which process the es- timated label distributions act as key clues in prediction-level interactions. Experiment results show that DARER outperforms exist- ing models by large margins while requiring much less computation resource and costing less training time. Remarkably, on DSC task in Mastodon, DARER gains a relative improve- ment of about 25% over previous best model in terms of F1, with less than 50% parameters and about only 60% required GPU memory. 1 Introduction Dialog sentiment classification (DSC) and dialog act recognition (DAR) are two challenging tasks in dialog systems (Ghosal et al., 2021). DSC aims to predict the sentiment label of each utterance in a dialog, while DAR aims to predict the act label. Recently, researchers have discovered that these two tasks are correlative and they can assist each other (Cerisara et al., 2018; Kim and Kim, 2018). An example is shown in Table 1. To predict the sentiment of ub, besides its semantics, its Disagree- ment act label and the Positive sentiment label of Utterances Act Sentiment ua: I highly recommend it. Really awe- some progression and added difficulty Statement Positive ub: I never have. Disagreement Negative Table 1: A dialog snippet from the Mastodon dataset. its previous utterance (ua) can provide useful refer- ences, which contribute a lot when humans do this task. This is because the Disagreement act label of ub denotes it has the opposite opinion with ua, and thus ub tends to have a Negative sentiment label, the opposite one with ua (Positive). Similarly, the opposite sentiment labels of ub and ua are helpful to infer the Disagreement act label of ub. In this paper, we term this process as dual-task reasoning, where there are three key factors: 1) the semantics of ua and ub; 2) the temporal relation between ua and ub; 3) ua’s and ub’s labels for another task. In previous works, different models are proposed to model the correlations between the …

Content truncated to 500 words