Annotation and Classification of French Feedback Communicative Functions

Annotation and Classification of French Feedback Communicative Functions Laurent Pr´evot Aix-Marseille Universit´e Laboratoire Parole et Langage Aix-en-Provence, France laurent.prevot@lpl-aix.fr Jan Gorisch Institut f¨ur Deutsche Sprache Mannheim, Germany gorisch@ids-mannheim.de Sankar Mukherjee Istituto Italiano di Tecnologia Genova, Italy sankar1535@gmail.com Abstract Feedback utterances are among the most fre- quent in dialogue. Feedback is also a crucial aspect of all linguistic theories that take social interaction involving language into account. However, determining communicative func- tions is a notoriously difficult task both for human interpreters and systems. It involves an interpretative process that integrates vari- ous sources of information. Existing work on communicative function classification comes from either dialogue act tagging where it is generally coarse grained concerning the feed- back phenomena or it is token-based and does not address the variety of forms that feed- back utterances can take. This paper intro- duces an annotation framework, the dataset and the related annotation campaign (involv- ing 7 raters to annotate nearly 6000 utter- ances). We present its evaluation not merely in terms of inter-rater agreement but also in terms of usability of the resulting reference dataset both from a linguistic research per- spective and from a more applicative view- point. 1 Introduction Positive feedback tokens (yeah, yes, mhm ...) are the most frequent tokens in spontaneous speech. They play a crucial role in managing the common ground of a conversation. Several studies have at- tempted to provide a detailed quantitative analy- sis of these tokens in particular by looking at the form-function relationship (Allwood et al., 2007; Petukhova and Bunt, 2009; Gravano et al., 2012; Neiberg et al., 2013). About form, they looked at lexical choice, phonology and prosody. About com- municative function, they considered in particular grounding, attitudes, turn-taking and dialogue struc- ture management. Despite the previous attempts to quantify that form-function relationship of feedback, we think that more work needs to be done on the conversa- tional part of it. For example, Gravano et al. (2012) used automatic classification of positive cue words, however the underlying corpus consists of games, that are far off being “conversational” and there- fore do not permit to draw any conclusions on how feedback is performed in conversational talk or talk- in-interaction. What concerns the selection of the feedback units, i.e. utterances, more work that clar- ifies what consists of feedback is also needed, as an approach that purely extracts specific lexical forms (“okay”, “yeah”, etc.) is not sufficient in order to account for feedback in general. Also, the question of what features to extract (acoustic, prosodic, con- textual, etc.) is far from being answered. The aim of this paper is to shed some more light on these issues by taking data from real conversations, anno- tating communicative functions, extracting various features and using them in experiments to classify the communicative functions. The study reported in this paper takes place in a project (Pr´evot and Bertrand, 2012) that aims to use, among other methodologies, quantitative clues to decipher the form-function relationship within feedback utterances. More precisely, we are …

Content truncated to 500 words