In the last decade, the challenges in spoken language systems have been tackled with variable success. Significant progress has been made on sound recognition and semantic interpretation techniques. Other issues - segmentation of discourse units and their interrelations in the overall discourse structure, still remain challenges to researchers. The impediments to processing these are due to speech inherent phenomena: dysfluencies (interjections, false starts, pause fillers); attention shift, and the grounding between the participants (Glass, 1999). We believe that a uniform Discourse Particle (DP) taxonomy supports resolving these and even is, in some instances, a key solution to them.

In this presentation we describe a DP taxonomy based on previous work on dysfluent speech (Meteer, 1995) and DPs (Fisher, 1997, 1998). Basing the taxonomy on Reichman’s dialogue grammar and Marcu’s (1997) discourse relations we demonstrate the effectiveness of DPs in the context of discourse analyses. The presented results deal with the application of the DP taxonomy to 10 of the TRAINS project (Allen et al., 1993) dialogues performed by five annotators who work outside the current project.

The DP taxonomy is based on three main features: their structure, their distribution, and their function. We consider DPs to be all sounds, syllables, words and phrases devoid of semantic meaning. They function in spoken dialogues as syntactical and/or pragmatic elements that support discourse structure. Their structure follows the main language level distinction. Sound DPs are interjections (hm, ah, oh). Word DPs are lexemes of different type: junction words (so, but, then) discourse words (okay, anyway, well) and repeated words. The third type - syntactic DPs are fixed phrases of the type (let me, you know, I mean). Most DPs are not constrained in their distribution: they appear at the beginning, the body or the end of a unit. Their function is to support units’ segmentation (mark out the beginning/end of a turn/dialogue, hold the turn, and so on) or/and to indicate the intentions and understanding of the current speaker. In that sense, their function is twofold. At the first level, immediately related to their neighbouring utterance, they show the boundaries of these neighbouring utterances and the relation between them as a pair. They prepare the listener for the utterance to follow. For example, if the speaker is presented with a plan whether she understands/agrees with its details would be explicit from the interjections at the beginning of the listener’s turn. The higher level function is to place the utterance in the context of a loose dialogue structure. On a higher level, DPs show the relation between the utterance to come with the previous ones, whether the relation is that of subordination or coordination.

The taxonomy is applied to the TRIANS 91 dialogue corpus (Gross et al. 1993), our own dialogues, and Thompson’s (1995) discourse transcript. . It comprises about 120 minutes (1500 turns, 6000 utterances) of spontaneous speech. Further expansion of the analysed data is in progress. The transcribed speech differs in genre- one corpus is scenario based; the other is topic constrained, and the third one unrestricted. All the dialogues are segmented using DAMSL annotation scheme (Core & Allen, 1997).

In this talk we present comparative results of the analysis made by the annotators. At this stage of the project we are focusing on analysing varieties of corpora and investigating the ways in which DPs reflect discourse structure and unit segmentation. The next stage of the project is aimed at augmenting the taxonomy and applying it as a DP expert module in a speech system.

Allen, J., Schubert L., Ferguson G., Heeman P., Hwang C., Kato T., Light M., Martin N., Miller B., Poesio M., and Traum D. (1995). The TRAINS Project: A case study in building a conversational planning agent. Journal of Experimental and Theoretical AI, 7, 7--48

Core, M. & Allen, J. (1997). Coding Dialogues with the DAMSL Annotation Scheme. Workshop Notes of AAAI Fall Symposium on Communicative Action in Humans and Machines.

Glass,V. (1999). Challenges for Spoken Dialogue Systems. Proceedings IEEE ASRU Workshop, Keystone,

Gross D., Allen J. and Traum D. (1993). The Trains 91 Dialogues. Technical Note 92-1, Computer Science Department, University of. Rochester.

Fischer, K. & Brandt, P. (1998): Automatically Disambiguating Discourse Particles. Proceedings of Coling/ACL '98 Workshop on Discourse Relations and Discourse Markers, Montreal, Canada, pp. 107-113.

Fischer, K. & Wrede, B. (1997): Discourse Particles in Female and Male Human-Computer-Interaction. Proceedings of WiC 1997, Intellect Press, Exeter.

Meteer, M. (1995) Dysfluency Annotation Stylebook for the Switchboard Corpus. unpublished.

Marcu,D., (1997) The Rhetorical Parsing, Summarization, and Generation of Natural Language Text. PhD theses. Department of Computer Science University of Toronto, Canada

Thompson, S., (1995) Appease the Monster. Tanscript. Conversation and Grammar. Ling.490.359/554.360. Linguistic Summer Program. University of New Mexico.