Narrative Theories as Contextual Constraints for Agent Interaction

This paper presents the view that narrative may be viewed as a framework for situating interaction (or the distribution of knowledge) in a particular context. The process of constructing this context explicates both interaction and knowledge exchange. Narrative techniques can therefore be placed within a framework for providing contextualization in agents to enhance user interactions, and this, in turn, can be used to inform the design of agent interaction and interactive narratives. Context and Contextualization The paradigm-shift from computer-as-tool to computer-ascommunicator in human-computer interaction emphasises the importance of context in constraining interaction, as context may be thought of as the environment that allows an intended meaning to be correctly ascribed by the communicators. Cooper (1991) emphasises the ability of people to actively construct such an underlying reality in order to resolve contradiction and inconsistency. In communicating, the dialogue partners actively construct context (i.e. establish shared plans, shared knowledge, etc.) there is no guarantee how much sharing will take place; context is as much a dynamic contest as a static goal. This process can therefore be viewed as a process of contextualization (Edmondson and Meech, 1994). Mittal and Paris (1993) identify the following as components of context: 1. The problem solving situation (the Tasks), 2. The participants involved (expertise, beliefs, goals, etc.) 3. The mode of interaction in which communication is occurring, 4. The discourse taking place, 5. The external world. Using these factors, context has been used in the design of adaptive interfaces with the motivation of improving the task-to-tool mapping and subsequently improve interaction (Croft, 1984). Here it is important to recognise that adaptation refers to the ability of the system to act appropriately in a given context. However, this conception of context is static and scope is already constrained by an implicit context, usually represented by a task (Tyler and Treu, 1989). For example, in Croft’s work the ability of the interface to adapt successfully is because the behaviour of the system is already anticipated by the user. (e.g. as “edit”, “form filling”, “email” or “calendaring”). This is because the user has recognised the next (higher) level of context is “office tasks”, and they know what it is possible to do in this context. Agent-based user interfaces essentially duplicate the functionality of intelligent and adaptive user interfaces, the change in terminology reflecting a change in paradigm rather than a change in goals. This paradigm emphasises the move towards active, autonomous support architectures, and brings with it a series of concerns relating to the characteristics which these agent interfaces appear to the user. These agents can take on many roles, adopting a range of human social conventions for the tasks they perform and the way they relate to the user. These roles can be summarised as follows (See also Mase, 1997; Wilson, 1997; Bickmore et al., 1998): Gurus The agent performs some reasoning (possibly prompted by the user) and supplies the results. Colleagues The user and the agent negotiate about the issue under consideration in a way similar to consulting a colleague (co-worker) for an opinion. Assistants Secretary Agents or Guides; These agents allow delegation of tasks from the user, necessarily including elements of Guru and Colleague. Autonomous Personal Representatives Agents that “stand in” for the user. They may express a user’s point of view, give guided tours, presentations, or represent the user’s opinions (including elements of personality). Avatars (virtual physical representations) are included in this category. Companions Agents that exist as believable social entities (including virtual pets). Entertainers Agents which entertain (e.g. Actors in Virtual Theatre). The communicating agent model implies that the system behaves intelligently, and consequently the responsibility From: AAAI Technical Report FS-99-01. Compilation copyright © 1999, AAAI (www.aaai.org). All rights reserved. for recognising what the system can or cannot do, when and how to act, etc. is shared between the agent and the user. The important question now concerns the scope of context and contextualization that an agent should take into account when communicating with a user. This classification of agent types indicates that the criteria include contextual elements (such as social attributes, personality, degree of autonomy in relation to the user e.g. degree of delegation, etc.). Many of there elements are outside the scope that is normally recognised as providing context. Brown and Duguid (1994) term these border resources. These socially shared, dynamic resources become established genres over time. “These genres are socially constructed interpretative conventions that bridge the two sides of a communication. ... on one side producers – architects, authors, designers, speakers and the like – try to invoke a particular genre to establish the conventions that they are putting into play.” p10 For example, calling a movie a “thriller” places it into an established genre and consequently establishes a recognisable context. The previous taxonomy of agent types encompasses a wide range of social and communicative abilities. To provide the abilities to enable an agent to contextualize in the widest sense, the agent will need resources based on the following factors: 1. The external world (the situated context) 2. The set of goals and tasks (the focal context) 3. A model of the user (preferences, etc. – the user context) 4. The form of dialogue (elements of agent role and social context) 5. The representation of the dialogue (physical representations of context). A model of the external world enables border resources to be used when available, and enables other non-task related factors to be used (e.g. social cues). The tasks, goals and user model enable shared context to be established as contextual process through dialogue. The dialogue itself also has elements of concerning social factors (e.g. the degree of initiative the agent is expected to take), coupled with how the dialogue is rendered (e.g. as speech, text, graphics, etc.). Notice that all of these categories have elements connected with social characteristics of the agent. e.g. the representation of the dialogue, even as speech, can be given intonation and gender characteristics which may influence the way the communication is perceived by the recipient. Many of these social factors relate to how human-like the agent is viewed as being by the user (how “believable” the agent seems). Narrative and Believable Agents One approach to the design of agents in order to generate believable behaviour is from the perspective of character. Perlin and Goldbers’ (1998) rationale for this believableness stems from the desire to convey mood and personality in order to express emotional messages or to portray specific characters rather than to solve a particular problem. They note that in certain applications the way in which the agent provides information can be as important as the information itself. In terms of a contextual model, the insight here is that human social cues apply to these agent-based systems, and they provide a great deal of contextual information to the user. The findings of Reeves and Naas (1996) concerning the inability of people to distinguish between media and real-life can also be understood from the perspective of media providing social context, which affects how the media are perceived. Such social conventions have a major effect on human-human interaction and consequently on human-computer interaction, especially when the computer exhibits some human-like qualities. Reflexively, the goals of affective computing (Picard, 1997) can therefore be viewed as enabling agents to contextualize these social cues from users. Sengers (1998) examines the transitional behaviour of existing social agents and observes that behavioural changes (between observable states of behaviour) exhibit patterns that are similar to that of schizophrenics. The change between observable actions lacks coherency, and this manifests itself in a parallel fashion to a human mental disorder. Sengers poses the question “how can an agent [behaviour] appear coherent to the user” and answers “by supporting the user in constructing coherent interpretations of the agent”. This is the same as enabling the user to contextualize the behaviour of the agent. Because the user uses the agents’ actions and behaviour to interpret the agents’ activity, the cues that the agent provides form the context in which the user places the whole interaction. By allowing the agent to provide appropriate behavioural cues for the user, the agent is contextualizing for the user, and this in turn facilitates the user’s contextualization. Contextualization and Narrative The other issues that concern the designers of believable agents also benefit from a re-analysis from the perspective of context and contextualization. The improvisational qualities sought by researchers such as Hayes-Roth and van Gent (1997) become the ability to make use of context in a way that situates agent behaviour within a context. This is the central tenant of improvisational theatre, the difficulty being the dynamics of improvisation and the ability to contextualize as the context changes. Other forms of storytelling use the term narrative to describe structure. The arguments about the use of narrative by agents (and in interfaces generally) pivot about leveraging the user’s inherent knowledge about how these structures provide context and understanding. A narrative provides a framework for an audience to construct a context that constrains their expectations of how a medium will behave. In this way, narrative can be viewed as a conceptual framework for providing an interactor with contextual constraints. Narrative (in a structuralist sense) is seen as the communication of a linked set of events (Rimmon-Kenan, 1983). This linking allows both structural and temporal links to be made within a story, and this in turn allows contextualization to take place in the audience (e.g. Mateas, 1997). The basis of Narrative Psychology (Bruner, 1986; 1990) is that intentional understanding is obtained by structuring events into stories (the alternative view being that inanimate understanding the computer as tool paradigm – is understood by cause and effect and logical reasoning). The rationale for understanding intentional behaviour in this way means that systems should incorporate narrative to provide an active, socially constructable context (contextualization). Narrative as Contextual Constraint Narrative and contextualization share many of the same attributes. They are both active processes, and they are may be viewed as being composed of several different elements. Narrative can be decomposed into the structural elements shown in figure 1. (Chatman, 1978 cited in Galyean, 1995). Figure 1. Elements of Narrative. Narrative is seen as being composed of a representation (Story) and the presentation of the story (Discourse). The discourse essentially becomes the rendering of the story onto some form of media. The Story, in turn, is divided into Events and Existents. Each of these elements can then be examined in terms of the contextualization it can provide. For example, character is viewed as an important element of storytelling, and the relationship with believable agents is obvious, as is the context than can be provided using characters as the embodiment of social cues. In a similar way, “setting the scene” is synonymous with providing context. Events may be compared with the concept of tasks, the sequencing, structure and composition of which provide vital contextual information. Mateas (1997) links character with story, and defines drama as consisting of characters, story and presentation. A story (narrative) is an experience with a temporal structure and defined presentation. Interaction, as the ability to influence temporal structure and presentation, raises the question of what is meant by the term “interactive story”? By allowing a communicative dialogue through interaction, the agent must be intelligent enough to guide the story in a way that provides the required narrative as interaction takes place. Oz takes an alternative approach of confining the interaction to certain points in the narrative, and allowing branching to take place. This is an approach taken by many computer games and (semi-) interactive fiction, and constrains the number of paths that the agent must be capable of navigating through. A dynamic approach to interactive narrative essentially provides an intelligent interface that manages the presentation of the story according to some set of criteria. For example in “Dogmatic”, Galyean (1995) presents a virtual environment that changes the events and appearance of the world to prompt a participant to take action that is meaningful to the narrative. Using these narrative devices, a context is provided as an overall story “shape”, but events still change within this framework (and contextualization takes place within these constraints). It is also worth noting that narratives also belong to recognisable genres, and these provide contextual constraints before a narrative begins. It may be that the design of compelling interactive narratives will only be possible when the need to maintain contextualization is appreciated. This will necessitate a different approach to authoring interactive narrative in which contextualization in each possible story-thread is explicitly considered. Narrative in Context Theories of Context and Contextualization can therefore be viewed as encompassing narrative theories from a communicative agent viewpoint. Comparing the elements of context and contextualization with the elements of narrative, an interactive narrative framework may be viewed as implementing many of the required elements for providing the means for an agent to contextualize. Contextualization Narrative