Two-level classification for dialogue act recognition in task-oriented dialogues

Proceedings of the 28th International Conference on Computational Linguistics, pages 4915–4925 Barcelona, Spain (Online), December 8-13, 2020 4915 Two-level classification for dialogue act recognition in task-oriented dialogues Philippe Blache1, Massina Abderrahmane2, St´ephane Rauzy1, Magalie Ochs3, and Houda Oufaida2, 1 LPL-CNRS-AMU, Aix-en-Provence, France (firstname.last@univ-amu.fr) 2 ESI, Algiers, Algeria (h oufaida@esi.dz) 3 LIS-CNRS-AMU, Marseille, France (firstname.last@univ-amu.fr) Abstract Dialogue Act classification becomes a complex task when dealing with fine-grain labels. Many appli- cations require such level of labelling, typically automatic dialogue systems. We present in this paper a 2-level classification technique, distinguishing between generic and specific dialogue acts (DA). This approach makes it possible to benefit from the very good accuracy of generic DA classification at the first level and proposes an efficient approach for specific DA, based on high-level linguistic features. Our re- sults show the interest of involving such features into the classifiers, outperforming all other feature sets, in particular those classically used in DA classification. 1 Introduction Task-oriented dialogue systems are specific in many respects. First, and this is the most important characteristic, they focus on a limited semantic domain, rendering the comprehension operation closer to slot filling than deep semantic understanding. Moreover, these applications correspond to very specific interactions, usually with one of the speakers (the human or the machine) leading the dialogue. In such situations, recognizing dialogue acts (DA) becomes extremely useful as a preliminary step of the understanding process: associating speaker’s utterances to DAs makes it possible to identify very efficiently the type of information they bear and the knowledge to be conveyed. Many works on dialogue act classification have been done for a long time (Stolcke et al., 2000). The first question consists in defining the set of relevant dialogue acts for the system. Different generic DA annotation schemes have been proposed, including an ISO standard (Bunt et al., 2012; Bunt et al., 2017). In addition, several dialogue corpora have been annotated and made available, among which two are particularly used: Switchboard Dialog Act Corpus, (SwDA) (Jurafsky et al., 1997) and Meeting Recorder Dialog Act (MRDA) (Shriberg et al., 2004). These datasets have allowed to train different classifiers. In these works, two preliminary questions arise: the identification of the dialogue acts to be classified and the features on which the classifier must be based. In most cases, the classification task targets a limited set of classes (corresponding to very general DAs such as statement, question etc.) and features remain at a low level (n-grams of words, characters, word embeddings, etc.). The performance of these classifiers is generally very good. However, the proposed task (i.e. the targeted tagset) remains often too general for an optimal use in a dialog system. Moreover, and this is a recurring problem in this type of approach, it is sometimes difficult to interpret the results and understand precisely the relative impact of the different features on the model. In this paper, we address the question of DA classification for helping comprehension and dialogue supervision in the context of task-oriented dialogue systems. Several …

Content truncated to 500 words