ISO-Standard Domain-Independent Dialogue Act Tagging for Conversational Agents

Proceedings of the 27th International Conference on Computational Linguistics, pages 3539–3551 Santa Fe, New Mexico, USA, August 20-26, 2018. 3539 ISO-Standard Domain-Independent Dialogue Act Tagging for Conversational Agents Stefano Mezza, Alessandra Cervone, Giuliano Tortoreto, Evgeny A. Stepanov, Giuseppe Riccardi Signals and Interactive Systems Lab, University of Trento, Italy name.surname@unitn.it Abstract Dialogue Act (DA) tagging is crucial for spoken language understanding systems, as it provides a general representation of speakers’ intents, not bound to a particular dialogue system. Unfor- tunately, publicly available data sets with DA annotation are all based on different annotation schemes and thus incompatible with each other. Moreover, their schemes often do not cover all aspects necessary for open-domain human-machine interaction. In this paper, we propose a methodology to map several publicly available corpora to a subset of the ISO standard, in order to create a large task-independent training corpus for DA classification. We show the feasibility of using this corpus to train a domain-independent DA tagger testing it on out-of-domain con- versational data, and argue the importance of training on multiple corpora to achieve robustness across different DA categories. 1 Introduction The correct interpretation of the intents behind a speaker’s utterances plays an essential role in determin- ing the success of a dialogue. Hence, the module responsible for intents classification lies at the very core of many dialogue systems, both in research and industry (e.g. Alexa, Siri). Moreover, although the task of intent recognition is traditionally linked to task-based systems, recently it has also proved crucial for non task-based conversational systems. According to the results of the Amazon Alexa Prize challenge (Ram et al., 2017), the most successful conversational systems in the competition relied on a strong spoken language understanding module, while more than 60% of the approaches explicitly used intents. Nevertheless, automatic intent recognition is hard, since participants’ intents in a dialogue are im- plicit. Intent classification has therefore been mostly modeled as a supervised machine learning problem (Gupta et al., 2006; Xu and Sarikaya, 2013; Yang et al., 2017), with the consequent definition of intents taxonomies. Over time this led to the creation of expensive annotated resources (Price, 1990; Henderson et al., 2014) with the related time-consuming design of multiple intent schemes. In most cases, however, intents taxonomies are defined specifically for a given application or a dataset and are not generalizable to other systems or tasks, making these resources difficult to reuse (e.g. the popular Air Travel Information Services (ATIS) corpus include heavily domain-dependent intents such as Airfare or Ground Service). Dialogue Acts (DA), also known as speech or communicative acts, represent an attempt to create a formalized and generalized version of intents. As such, DAs have been investigated by the research community for many years (Stolcke et al., 2000) and have been applied successfully to many tasks. In particular, their aspiration to generality makes them an appealing option for non task-based application (e.g. more than 20% of the teams in Amazon Alexa Prize Challenge explicitly used DAs (Cervone et al., 2017; Bowden et …

Content truncated to 500 words