Learning Interpretable Latent Dialogue Actions With Less Supervision

Proceedings of the 2nd Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics and the 12th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 297–308 November 20–23, 2022. ©2022 Association for Computational Linguistics 297 Learning Interpretable Latent Dialogue Actions With Less Supervision Vojtˇech Hudeˇcek and Ondˇrej Dušek hudecek@ufal.mff.cuni.cz, odusek@ufal.mff.cuni.cz Charles University, Faculty of Mathematics and Physics Malostranské námˇestí 25, 118 00 Prague, Czechia Abstract We present a novel architecture for explain- able modeling of task-oriented dialogues with discrete latent variables to represent dialogue actions. Our model is based on variational re- current neural networks (VRNN) and requires no explicit annotation of semantic information. Unlike previous works, our approach models the system and user turns separately and per- forms database query modeling, which makes the model applicable to task-oriented dialogues while producing easily interpretable action la- tent variables. We show that our model outper- forms previous approaches with less supervi- sion in terms of perplexity and BLEU on three datasets, and we propose a way to measure dialogue success without the need for expert annotation. Finally, we propose a novel way to explain semantics of the latent variables with respect to system actions. 1 Introduction While supervised neural dialogue modeling is a very active research topic (Wen et al., 2017b; Lei et al., 2018; Peng et al., 2021), it requires a signif- icant amount of work to obtain turn-level labels, usually with dialogue state annotation. We argue that in many real-world cases, it is very expensive to obtain the necessary labels or even to design an appropriate annotation schema. Consider a call center with various dialogues that has a lot of tran- scripts available, including the corresponding API queries, but has no capacity to label them. This mo- tivates our research of approaches that minimize the need for expert annotation. While most recent research focuses on pretrained language models (PLMs) and reaches state-of-the- art performance in standard supervised (Peng et al., 2021; Zhang et al., 2020b) or even few-shot (Peng et al., 2020; Wu et al., 2020) settings, these mod- els still require full supervision. Furthermore, they lack the potential to interpret the model de- cisions. Some recent works try to address PLM interpretability with some success (Lin et al., 2019; Stevens and Su, 2021), but still face considerable difficulties due to PLMs’ huge number of param- eters and their structure. On the other hand, dia- logue models using latent variables are able to infer interpretable attributes from unlabeled data (Wen et al., 2017a; Zhao et al., 2019). These models are mostly trained using variational autoencoders (VAE; Kingma and Welling, 2014; Serban et al., 2017). Improvements with discrete variables (Zhao et al., 2018; Shi et al., 2019) offer better inter- pretability, but the approaches are not directly ap- plicable to task-oriented response generation as no distinction between the system and user roles is made, and database access or goal fulfillment are not considered; most research on unlabeled data only applies to a chit-chat setting. Since …

Content truncated to 500 words