Reference production in human-computer interaction: Issues for Corpus-based Referring Expression Generation

Reference production in human-computer interaction: Issues for Corpus-based Referring Expression Generation Danillo da Silva Rocha, Ivandr´e Paraboni University of S˜ao Paulo, School of Arts, Sciences and Humanities S˜ao Paulo, Brazil rochadan60@gmail.com,ivandre@usp.br Abstract In the Natural Language Generation field, Referring Expression Generation (REG) studies often make use of experiments involving human subjects for the collection of corpora of definite descriptions. Experiments of this kind usually make use of web-based settings in which a single subject acts as a speaker with no particular addressee in mind (as a kind of monologue situation), or in which participant pairs are engaged in an actual dialogue. Both so-called monologue and dialogue settings are of course instances of real language use, but it is not entirely clear whether these situations are truly comparable or, to be more precise, whether REG studies may draw conclusions regarding attribute selection, referential overspecification and others regardless of the mode of communication. To shed light on this issue, in this work we developed a parallel, semantically annotated corpus of monologue and dialogue referring expressions, and carried out an experiment to compare instances produced in both modes of communication. Preliminary results suggest that human reference production may be indeed affected by the presence of a second (specific) human partic- ipant as the receiver of the communication in a number of ways, an observation that may be relevant for studies in REG and related fields. Keywords: Natural Language Generation, Referring Expression Generation, Corpus 1. Introduction In Natural Language Generation (NLG) studies, the collec- tion of referring expressions - usually in the form of definite descriptions - produced by human subjects is a common task in Referring Expression generation (REG) and related fields (Krahmer and van Deemter, 2012). Descriptions of this kind are usually elicited from visual stimuli represent- ing a context in which there is one particular target and ad- ditional distractor objects. Figure 1 illustrates a stimulus image from the Stars2 corpus (Paraboni et al., 2017a). Figure 1: A stimulus image from the Stars2 corpus. Given a context of this kind, the task of the human subject - who acts as a speaker or writer - is to produce a uniquely identifying description of the intended target. This could be accomplished, for instance, by producing a definite de- scription as in ‘The cone next to a grey box’. Experiments involving human subjects for the collection of referring expression corpora are often implemented as a web-based data collection task, that is, without a particu- lar addressee in mind. When there is no risk of confusion, we will hereby call these monologue situations1. So-called monologue situations are the method of choice for collecting data in TUNA (Gatt et al., 2007), GRE3D3 (Dale and Viethen, 2009), GRE3D7 (Viethen and Dale, 2011), Wally (Clarke et al., 2013) and other similar re- sources. By contrast, a number of data collection tasks make use of participant pairs in some form of dialogue. These include GIVE-2 (Gargett et al., 2010), ReferIt (Kazemzadeh et al., 2014), Stars2 (Paraboni et al., …

Content truncated to 500 words