Towards a Model for Discourse Marker Annotation in spoken French: From potential to feature-based discourse markers

Starting from the common observation that there is no recognized closed class of discourse markers (DMs) and that a number of linguistic markers may or may not count as DMs according to the definitions at stake (Schourup 1999: 228), we aim to present an empirical method for the identification and annotation of DMs in spontaneous spoken French (MDMA project). Central to our proposal is that DMs may be described as clusters of features that, in specific patterns of combination, allow distinguishing DM use from other uses. We proceeded in three steps: (i) using a very broad definition of DMs – i.e. items that “provide instructions to the hearer on how to integrate their host utterance into a developing mental model of the discourse in such a way as to make that utterance appear optimally coherent” (Hansen 2006: 25) – three analysts identified all potential DMs in an 800 words transcript; (ii) all types found were then extracted from a balanced 10,000 words corpus; and (iii) analyzed according to more than 10 features (including syntactic, semantic, collocational, and prosodic features). The hypothesis underlying our annotation experiment is that the analysis of the distributional constraints imposed on specific markers should uncover reliable features for the identification and categorization of DMs. The potential DMs extracted at the first step of analysis refer to those linguistic expressions that can, in one context or another, fulfill a DM function, i.e. be used at either of the following “levels” or “domains”: “the sequential structure of the dialogue, the turn-taking system, speech management, interpersonal management, the topic structure, and participation frameworks” (Fischer 2006: 9). For example, tu vois ‘you see’ is defined as a potential DM because it can occur in contexts where it serves to manage the relationship between speaker and hearer (1), although in other contexts it does not (2). In line with the objectives of the workshop, our endeavor seeks to establish (more) reliable criteria for the categorization as DMs, and how to distinguish them from other linguistic items fulfilling a non-propositional function, such as modal particles (Degand et al. 2013) or pragmatic markers (Brinton 1996). Disagreement between annotators also reveals borderline cases where, although some discursive, pragmatic, indexical function is commonly detected, there seems to be some hesitation as to what category these items belong to (such as c'est ça ‘that’s it’, quand même ‘still’ in examples 3-4). In this presentation, we first briefly go over our methodological choices and issues, and then uncover problematic examples of inter-coder disagreement, as well as the first results of the statistical analysis of clusters of features. The latter suggests that there is a certain hierarchy between the different features under scrutiny, regarding their relevance, reliability, or usefulness in the process of identifying DMs in context