Cue Phrase Selection in Instruction Dialogue Using Machine Learning

Cue Phrase Selection in Instruction Dialogue Using Machine Learning Yukiko I. Nakano NTT Information and Communication Systems Laboratories 1-1 Hikari-no-oka, Yokosuka-shi, Kanagawa 239-0847 Japan yukiko~nttnly.isl.ntt.co.jp Tsuneaki Kato NTT Communication Science Laboratories 2-4 Hikaridai, Seika-cho, Soraku-gun, Kyoto 619-0237 Japan katoQcslab.kecl.ntt.co.jp Abstract The purpose of this paper is to identify effective fac- tors for selecting discourse organization cue phrases in instruction dialogue that signal changes in dis- course structure such as topic shifts and attentional state changes. By using a machine learning tech- nique, a variety of features concerning discourse structure, task structure, and dialogue context are examined in terms of their effectiveness and the best set of learning <features is identified. Our result re- veals that, in addition to discourse structure, already identified in previous studies, task structure and di- alogue context play an important role. Moreover, an evaluation using a large dialogue corpus shows the utihty of applying machine learning techniques to cue phrase selection. 1 Introduction Cue phrases are words and phrases, such as "first", "and", "now", that connect discourse spans and add structure to the discourse both in text and dialogue. They signal topic shifts and changes in attentional state (Grosz and Sidner, 1986) as well as expressing the relation between the individual units of discourse (Moore, 1995; RSsner and Stede, 1992). In this study, we focus on the former kind of cue phrases, organization cue phrases that signal the structural organization of discourse. In instruction dialogue, the organization cue phrases play a crucial role in controlling dialogue and making the material easy to understand. More- over, in dialogue systems, the user cannot compre- hend the structural organization of the dialogue un- less the appropriate cue phrases are included in the system's utterances. Therefore, for dialogue gener- ation, we must identify the determining factors of organization cue phrases and select the cue phrases appropriately. In previous studies that have investigated the rela- tionship between cue phrases and the types of struc- tural change (e.g. pop, push), the taxonomies of cue phrases have been presented (Grosz and Sid- ner, 1986; Cohen, 1984; Schiffrin, 1987). These tax- onomies are, however, not sufficient for generation because the correspondence between cue phrase and structural change is many-to-many quite often. For example, "now","and", and "next" are all classified as the category signaling push in attentional state. Therefore, the indication of structural shifts in dis- course is not sufficient to fully constrain cue phrase selection. In this study, we reveal what factors affect or- ganization cue phrase selection, and establish more precise selection rules for generating instruction dia- logues. As factors for cue phrase selection, we exam- ine a variety of features concerning discourse struc- ture, task structure, and dialogue context. The rea- son that we examine these three factors is as fol- lows. First, discourse structure is indispensable for selecting cue phrase as claimed in previous stud- ies (Grosz and Sidner, 1986; Cohen, 1984; Euge- nio et al., 1997). We examine some features con- cerning this factor such as the global structure of …

Content truncated to 500 words