Mixed-Initiative Retrieval Dialogues Using Abductive Reasoning

Using intelligent dialogue components, we can relate the isolated functions of information retrieval systems to larger tasks and help the user satisfy her information needs. Such dialogue components monitor the interaction between user and system and suggest what the user should do on the basis of her goals and the dialogue history. However, for these components to have the necessary flexibility, they must allow both user and system to take the initiative and control the direction of the dialogue. In our retrieval system, certain dialogue scripts decide who is in charge of the interaction and what are the recommended user acts in a goal-directed dialogue. The user can take the initiative at any point of the dialogue by issuing an act not recommended to her. The system uses abduction to interpret this act in light of the dialogue history and initiates a new dialogue addressing the user’s choice of interpretation. Introduction To offer effective interaction facilities, any advanced system must rely on an--implicit or explicit--method for allocating initiative to the interactants. Modeling mixed initiative (MI) interaction requires that the system apply strategies and mechanisms for deciding when to take/pass initiative depending on the current dialogue situation and how to adapt its behavior accordingly. Hence, the system uses knowledge representation and reasoning techniques to plan its own strategy and co-operate with the user. MI also means that the roles of the agents are not predetermined and that the agents must be enabled to negotiate the control of initiative (Allen, 1994). The agents not only interact solve domain problems, but also collaborate at the level of "interaction management" (Bunt, 1996) and negotiate the problem-solving strategy, e.g., to build up "SharedPlans" (Grosz & Sidner, 1990; Sidner, 1994). * Jon Atle Gulla has been involved in the research reported here during his employment at GMD-IPSI. Issues of initiative and dialogue control are related to some basic issues of intelligent human-computer collaboration (cf. Terveen, 1995). Research relevant to the topic concentrates on AI-oriented approaches to collaborative activity/planning, discourse modeling/planning, and adaptive user modeling. Focusing on the agents’ beliefs, goals, and intentions, most computational models of collaborative discourse have been developed for natural language applications, for example, text planning systems for advisory/explanation dialogues (e.g., Moore & Paris, 1993; Haller, 1996) and other task-oriented systems such as many spoken dialogue systems (see e.g., Dalsgaard et al., 1995 for a collection). In addition to goals and plans, some existing computational dialogue models explicitly take initiative and other social factors such as conversational conventions, expectations and obligations into account (e.g., Traum & Allen, 1994; McRoy & Hirst, 1995; Jokinen, 1996). This paper presents a theoretically motivated approach to the management of mixed-initiative dialogues in intelligent information retrieval systems. Most of the above mentioned approaches and systems presuppose well-defined (planning) tasks, e.g., transportation planning or repair of technical devices. Information retrieval in large databases is a highly interactive--and possibly very complex--task, but the information needs of users of such systems are usually less well defined and users often opportunistically change their goals and strategies during interaction. Hence, system support should be situation and context dependent without relying on a strict plan-goal hierarchy. State-of-the-art retrieval systems provide many usersupport mechanisms, such as query construction aids or access to thesauri, but these are mostly treated as isolated extra-dialogic functions or tools offered to the user. It has been suggested (e.g., Belkin et al., 1995; From: AAAI Technical Report SS-97-04. Compilation copyright © 1997, AAAI (www.aaai.org). All rights reserved.