The ADELE Corpus of Dyadic Social Text Conversations:Dialog Act Annotation with ISO 24617-2

The ADELE Corpus of Dyadic Social Text Conversations: Dialog Act Annotation with ISO 24617-2 Emer Gilmartin1, Brendan Spillane1, Maria O’Reilly2, Christian Saam1, Ketong Su1, Leigh Clark3, Benjamin R. Cowan3, Killian Levacher4, Arturo Calvo5, Loredana Cerrato6, Nick Campbell1, Vincent Wade1 ADAPT Centre Trinity College Dublin1, Phonetics Laboratory, Trinity College Dublin2, University College Dublin3, IBM Research Ireland4, Accenture Norge5, EIT Digital Sweden6 gilmare@tcd.ie Abstract Social or interactional dialog is less well described than task-based or instrumental dialog, although there is increasing interest in the genre, particularly in light of new spoken and text dialog applications which aim to relate to the user as well as perform tasks. Dialog act annotation aids understanding of interaction structure; essential to the design of sucessful artificial dialog. Much social text interaction may be closer to social talk than to traditional written language. In this paper we briefly describe social or casual talk, and review how current dialog annotation schemes and particularly the ISO standard 24617-2 (Semantic annotation framework, Part 2: Dialogue Acts) treat non-task elements of dialog. To aid in training a casual talk system, we collected a corpus of 193 dyadic text dialogs, based on a novel ‘getting to know you’ social dialog elicitation paradigm. We describe the annotation of the dialogs, and propose additional acts to better cover greeting and leavetaking. We report on preliminary analyses of the corpus, and provide an overview of the corpus content and its relationship to spoken language. The corpus, coding manual, and annotations are being packaged and will be made available to interested researchers. Keywords: Casual Conversation, Dialog Corpus, Dialog Acts 1. Introduction Many dialogs and indeed parts of dialogs can be described as task-based or instrumental, with clear goals, as in the case of a service encounter or business meeting. Others, such as friendly chats or longer casual conversations, are more interactional in nature. Acknowledging that mod- elling casual conversation would prove difficult, early di- alog technology research focused largely on practical goal- oriented dialog (Allen et al., 2001). Recently, more atten- tion is being paid to social aspects of spoken and text in- teraction, with the desire for systems which can engage, entertain, and provide the illusion of companionship to users. Successful design of such systems requires relevant data. Dialog act annotation aids understanding of interac- tion structure; such understanding is essential for designing artificial spoken or text dialog. Below, we briefly review accounts of casual social talk and discuss how current dia- log annotation schemes and particularly the ‘ISO standard 24617-2 Semantic annotation framework, Part 2: Dialogue acts’ (ISO, 2012) (henceforth ISO standard) treat non-task elements of dialog. We then describe the collection and an- notation, using the ISO standard, of 193 text dialogs elicited using a novel ‘getting to know you’ paradigm. We describe the distribution of dialog acts in the corpus as a whole, re- port on annotation of the dialog acts used in greeting and leave-taking sequences, and describe proposed new social dialog acts for these sequences. We also outline our ongo- …

Content truncated to 500 words