Building an RST-based multi-level dialogue context and structure

1 downloads 0 Views 207KB Size Report
We present a new initial approach to modeling dialogue which, based on two levels of discourse planning, builds structure and context incrementally in multiple ...
Building an RST-based Multi-Level Dialogue Context and Structure T. Daradoumis Departament de Llenguatges i Sistemes Inforniltics Universitat Politknica de Catalunya, 08028 Barcelona, Spain Abstract

discourse proceeds, using a set of (non-RST) context forming methods(CFM). In addition, the scope of context is further extended to incorporate pragmatic and stylistic goals(aspects)of the conversation (in the style of [Z]). The use of two levels of planning is similar to the use of contenr and dialogue plans in generating explanatory dialogues [ 13 and to several tutoring planning systems. However, unlike those systems, our context forming planning level focuses in building and organizing a multi-level differentiated context of the whole dialogue. This is an important issue (as argued by researchers in discourse understanding), failed to be treated adequately by most generation systems. On the other hand, our dialogue planning level emphasizes the need to take a more global view of the dialogue and account for its overall coherence. Then, building an RST-based hierarchical dialogue model, we manage (unlike the model of [ 13) to recognize and make explicit the rhetorical relations(means) that hold between dialogue actions on each conversational level, thus being able to explain the role each action plays with respect to another of the same level and to the higher-level goal both actions contribute to. The overall dialogue can be represented by a tree structure. In our model, in addition to the principal context (intentional, attentional and belief knowledge) built by the CFMs, DPMs also contribute to context forming on each conversational level, by filling in context with further complementary information, such as rhetorical knowledge, pragmatic/stylistic information and a record of the current dialogue state. First of all, in order to account for the overall coherent organization of the dialogue, rhetorical relations are used as DPMs to structure a conversation, dialogues and subdialogues on a topic, exchanges with the user, moves within an exchange and speech acts within a move; that is, we distinguish six levels of description for a conversation. These methods actually function as metaplans and define the coherent combinations of lower-level conversational actions allowable for an

We present a new initial approach to modeling dialogue which, based on two levels of discourse planning, builds structure and context incrementally in multiple levels as the dialogue progresses. This fact allows speaker’s intentions and beliefs as well as attentional, rhetorical and pragmatic knowledge to be represented at each level in a more specific manner than previous models.

1. Introduction Important issues in generating interactive explanations are how to structure the resulting dialogue and how the changing context (intentions, beliefs, attentional and pragmatic information) is to be monitored (maintained and updated) with an explanation. Our dialogue model offers four major advantages over previous approaches [l] [3] [5]: (1) it provides a more sophisticated account of dialogue context and structure built incrementally on multiple interconnected levels: (2) it uses rhetorical relations (based on Rhetorical Structure Theory (RST) [4]) that ensure overall coherence and explicitly identify the role of an action in dialogue; (3) it provides a computational approach that uses two interleaving levels of discourse planning,; and (4) it makes explicit the general-purpose explanation strategies used and distinguishes them from the RST relations implicit in the explanation.

2 . A two-level dialogue

plan-based model of

The basic planning approach includes two levels of planning: One level is concerned with organizing the structure of the dialogue, using a set of dialogue planning methods(DPM) which are rhetorical relations. The other level is concerned with building and maintaining the context of the dialogue(exp1anation) incrementally as a

465

1043-0989/94 $03.00 0 1994 IEEE

[ l]), global coherence of the explanatory dialogue is explicitly indicated and maintained through the rhetorical relations which give the directives of how the outlined future plan could be realized to blend in a more natural way with the previous discourse. Finally, each DPM contains what we call the rhetorical structure of the conversation level it models. This structure is useful since it not only houses all the principal discourse processes that influence the current conversational level but also makes their effects consistent, thus ensuring dialogue overall coherence; moreover, it explicitly indicates how two dialogue subactions are related and which purposes each one serves in achieving a higher-level action.

immediate higher-level action. In other words, a DPM is used to pursue a conversational goal by decomposing it into subgoals that are to be achieved at the immediately lower conversational level. Before proceeding to achieve the fist posted subgoal, control switches to a CFM which is engaged to set the context underlying that subgoal. The type of context that is adequate at that point of discourse depends on criteria such as, the conversational level, the particular dialogue structure that results from applying the current RST relation (i.e., the current and future goals to be achieved), as well as the semantics of this relation and the principles that motivated its choice. In general, context should express knowledge, such as speaker’s beliefs (needs and preferences) about the current activity as well as speaker’s beliefs about hearer’s mental states (e.g., knowledge, attitudes, etc.). Dialogue planning and context forming methods are used alternatively and jointly (by letting one influence the other as the discourse proceeds) down to the speech act level to plan a discourse (i.e., to achieve a top-level conversational goal). The model is currently used to generate cooperative and negotiative tutorial dialogues. We distinguish two types of tasks that correspond to the two levels of planning: dialogue tasks, posted by a DPM, that represent the conversational actions(leve1s) a tutorial discourse is decomposedand intentional tasks, posted by a CFM, that represent the pedagogical, teaching and communicative (exchange, move and act) goals of the participants. Based on the function of and alternation between DPMs and CFMs, a dialogue task is always decomposed into an intentional task and it is during this process when the system sets the intentional task’s surrounding context; in contrast, an intentional task is decomposed into dialogue tasks, which serves to indicate the necessary conversation actions needed to bring the intentional task into completion. If the intentional task is eventually achieved, then, all the intentions and beliefs conveyed by the task’s context become true; discourse context is updated on each particular level of conversation. Context evolves from general enough to more specific as we come down the dialogue hierarchy: the principles guiding the inclusion of intention, belief, attentional and pragmatic information evolves from general pedagogic down to more specific instructional and explanatory, involving linguistic and cognitive ones at the lower levels. The system does not plan too much ahead. Future explanation is always outlined by being simply represented by a number of high-level goals still to be satisfied and the way it is realized will depend on the updated context (as in [ 11). However (unlike the system of

3. Conclusions We have presented an initial RST-based planning approach to building a multi-level context and structure of tutorial explanatory dialogues. The system is able to capture the kinds of interactions and interrelations among levels of context, among context types within each level, as well as, between discourse context and structure as a whole. This results in building a richer and more fine-grained representation of the dialogue structure and context than previous explanation generation systems. Acknowledgements This work is being supported by a European Economic Community fellowship under contract no. ERBCHBICT920213, within the framework of the Human Capital and Mobility program.

References [ 11 Cawsey, A. Generating communicative discourse. In: Current Research in Natural Language Generation. Dale, R., Mellish, C. and Zock. M. (Ed.). Academic Press. Boston. pp. 75-102, 1990. [2] Hovy, E. Pragmatics and Natural Language Generation. Artificial Intelligence Journal. Vol. 43, No. 2, 153-198, 1990. [3] Lambert, L. & Carbeny, S . A tripartite plan-based model of dialogue. In: Proc. of the 29th Annual Meeting of the Association for Computational Linguistics. Berkeley. California. pp. 47-54, 1991. [4]Mann, W.& Thompson, S . Rhetorical Structure Theory: A theory of text organization. 1n:The Structure of Discourse. Livia Polanyi. (Ed.). Ablex Publishing Corporation. Norwood. NJ. 1987. [ 5 ] Moore, I. & Paris, C. Planning text for advisory dialogues. In Roc. of the 27th Annual Meeting of the ACL. pp. 203-211, 1989.

466

Suggest Documents