Discovering Dialogue Slots with Weak Supervision

Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 2430–2442 August 1–6, 2021. ©2021 Association for Computational Linguistics 2430 Discovering Dialogue Slots with Weak Supervision Vojtˇech Hudeˇcek,1 Ondˇrej Dušek1 and Zhou Yu2 1Charles University, Faculty of Mathematics and Physics, Institute of Formal and Applied Linguistics 2 Columbia University, Department of Computer Science {hudecek,odusek}@ufal.mff.cuni.cz, zy2461@columbia.edu Abstract Task-oriented dialogue systems typically re- quire manual annotation of dialogue slots in training data, which is costly to obtain. We propose a method that eliminates this require- ment: We use weak supervision from existing linguistic annotation models to identify poten- tial slot candidates, then automatically identify domain-relevant slots by using clustering algo- rithms. Furthermore, we use the resulting slot annotation to train a neural-network-based tag- ger that is able to perform slot tagging with no human intervention. This tagger is trained solely on the outputs of our method and thus does not rely on any labeled data. Our model demonstrates state-of-the-art per- formance in slot tagging without labeled train- ing data on four different dialogue domains. Moreover, we find that slot annotations dis- covered by our model significantly improve the performance of an end-to-end dialogue re- sponse generation model, compared to using no slot annotation at all. 1 Introduction Task-oriented dialogue systems typically use anno- tation based on slots to represent the meaning of user utterances (Young et al., 2013). Slots are at- tributes relevant to completing the task (e.g., price, food type, area). The sets of slots and their val- ues typically need to be designed in advance by domain experts. Slots and their values are tracked over the course of the dialogue, forming dialogue state, which allows a dialogue system to plan the next actions effectively (Williams et al., 2013). Getting raw data for dialogue system training is not difficult, especially if we restrict the tar- get domain. A requirement for dialogue state la- bels makes this process much more costly. How- ever, both traditional pipeline systems (Young et al., 2013) and end-to-end task-oriented architectures (Wen et al., 2017) typically require such annotation. While some systems use implicit, latent state rep- resentation and do not require annotation (Serban et al., 2016), the behavior of such systems is hard to interpret or control. There are several works aiming at keeping interpretability and reducing the anno- tation needs by automating it (Chen et al., 2014, 2015) or transferring annotation across domains (Zhao and Eskenazi, 2018; Coope et al., 2020), but they still require significant manual effort. In this paper, we present a novel approach to discovering a set of domain-relevant dialogue slots and their values given a set of dialogues in the target domain (such as transcripts from a call cen- ter). Our approach requires no manual annotation at all in order to tag slots in dialogue data. This substantially simplifies dialogue system design and training process, as the developer no longer needs to design a set of slots and annotate …

Content truncated to 500 words