Mateas and Sengers: Narrative Intelligence

[Readings] (02.28.09, 5:53 pm)

Narrative Intelligence is a bridge between narrative and artificial intelligence. The book is a compilation of papers that touch on the theory of computation and narrative in various ways. The approach to AI is heavily influenced by Schank, particularly in the sense that to develop a computational model of something is to develop a theory of it. Working with narrative crosses many disciplinary boundaries, and borrows from several conventions of computer science. Additionally, the papers in the book borrow from many other disciplines, such as art, psychology, cultural studies, drama, and “humanistic AI”. The approaches and goals may be wildly different. Ultimately NI is many and not one, and the richness and diversity of narrative intelligence is consistent with the richness and diversity of narrative.

Jerome Bruner: The narrative construction of reality

The subject of this article is how reality is constructed. This originates as a philosophical question, namely: how does one achieve knowledge of the world? Dominan approaches to this are empirical and rationalist. Both of these assume an immutable world ready to be observed. This ties again to psychology and cognitive science. Bruner’s goals are to look at the cultural and personal dimension to constructing reality, which is very different. Bruner uses narrative as a framework for the construction of reality. To do so, he outlines ten features of narratives and outlines how they relate to this central theme:

  1. Diachronicity: Multiplicity of dimensions of time.
  2. Particularity: Narrative is about the specific and not the general. Lessons may be gained from it, but not general schemas or rules for knowledge generalization.
  3. Intention: The reader of a narrative is interested in the intentions and motivations behind the characters: reasons rather than causes.
  4. Hermeneutic composability: There is an interdependence between the whole of the story and its parts. The structure is not organized in a clear context-free tree structure.
  5. Canonicality and breach: Narratives tend to focus on breaches of canonical behavior and circumstances.
  6. Referentiality: Narrative creates an independent and internal world, where truth is defined by verisimilitude, rather than verifiability. Logic is about internal consistency, not factuality.
  7. Genericness: Narratives frequently fall into genres, which structure human plights and circumstances, and present a formula for interpreting the meanings of events and circumstances, as though they were within the genre’s model.
  8. Normativeness: Narrative is concerned with conflicts and breaches, but ultimately these result in a return to a normative state. Narrative shapes cultural legitimacy. While it may not resolve real world problems, it may articulate plights (Kermode’s “consolation of narrative”), which does not provide comfort of happy ending, it may make a plight bearable by being made understandable.
  9. Context sensitivity and negotiability: Narrative is dependent on context and its interpretation is negotiable, not absolute.
  10. Narrative accrual: Stories are put together to shape whole worlds.

Brenda Laurel: Vital narratives

Laurel’s paper is on the types of narratives found in culture and how they affect individuals’ understanding of the world. These are understood with respect to three axes: Personal relevance, strategies and outcomes, and epistemologically. Laurel reviews several kinds of narratives that are culturally present and shape the listener’s relationship to the world outside the story. There are a variety of these: religious, folk and spiritual, scientific, historical, journalistic, political and geopolitical. She challenges the tendency of many of these to generalize and provide a narrowing of ways of looking at the world, which are made more prevalent by the pervasiveness of these kinds of stories.

Steffi Domike, Michael Mateas, and Paul Vanouse: The recombinant history apparatus presents Terminal Time

This gives a general description of both the architecture of Terminal Time and its motivations. The project adopts and parodies the “cookie-cutter documentary” format, which uses presentation of historical narratives to convey presentational assertions as facts. These documentaries make use of filmic conventions of juxtaposition, the “Kuleshov effect”, to convey biases with authority. This is emphasized with an example of two documentaries, “The River Ran Red”, and “The Richest Man in the World”, which presented very different perspectives of Andrew Carnegie’s influence on the steel industry in Pittsburgh, though they used much of the same footage.

Because Terminal Time makes use of the generation of these narrative segments, it arguably lacks a preexisting bias in terms of its narrative construction. Additionally, the means for engaging with it is participatory and forum-like format, inviting interaction and then discussion. This is very different from the generally passive format of documentaries, where the viewers are not invited to be critical of the conclusions being drawn by the documentary itself. While Terminal Time does not take inherent positions on the material out of which it composes documentaries, it can be said to have a meta-intention as a critique and deconstruction of the documentary format.

The architecture of Terminal Time is quintessentially recombinant, it is not at all a simulation of a world or story, but a reordering and a presentation of one. It could perhaps be argued that it simulates a documentary alone. It uses rules of telling, but not rules of the ideological content. Instead, history is presented as content, as evidence of an ideology. History is presented as though it were the result of a systemic simulation, where the visible material was the result of a whole self-contained simulation. This format is different from simulation where, having adopted an ideology, a new history would need to be written that is the product of that model.

Chris Crawford: Assumptions underlying the Erasmatron storytelling system

An interesting observation in this essay is that the format of choice and interactivity implies a certain kind of agency within a story, and requires an action-oriented story structure. It is possible to compare Crawford’s valuing of player choice (as verbs), with other narrative domains, for instance Pride and Prejudice, where much of what a player might do is dependent on emotional responses. The role of emotional responses does not seem to be as sturdily supported as other means of interactions.

R. Raymond Lang: Story Grammars

This section describes a story generation system via grammars. The format of this consists of a world model and a story grammar. The execution of this model is through “rational intention”, which has a structure of arranging episodes. The system is still problematic because of its uninteresting world and the general vapidness of the stories. This uses emotions, but as tokens. Emotional responses need to be deeper as having an affect on characters, more than being mere justifications for behavior.

The use of story grammars also needs to address the question of human relevance. Why tell stories in the first place? Stories must have some value to somebody. Turner addresses this, but through moralizing, which is fine for what it is. Many story worlds have value and meaning in them already, and may be conveyed ironically or genuinely to convey authorial intent.

Andrew Stern: Virutal Babyz

Stern’s discussion presents Petz and Babyz as narrative systems, but I do not think that is accurate as a general structure (nor should it be). Rather, it is a simulation system, composed of many pieces, which may then emerge into narratives. They contain a degree of flexibility, ambiguity and cartoon referentiality that helps scaffold narrative emergence. Dimensions of emergence can come from long and short term behaviors, but primarily live in the head of the user. This raises the question of how a simulation system needs to be “pre-loaded” with a narrative basis.

Pheobe Sengers: Schizophrenia and narrative in artificial agents

Sengers presents the problems of agents as that they miss a narrative dimension to their interactions. This problem is similar to the handling of schizophrenics by psychiatrists. Schizophrenics suffer from a certain lack of narrative consistency in their behavior, they see themselves as machine-like, composed of many pieces that do not quite work together, often very painfully so. However the psychiatric institution seeks to mechanize the treatment for them, looking at their symptoms individually as separate problems to be diagnosed. The state of the schizophrenic is one of dissociation, mechanization, and a lack of interrelation between behaviors.

Sengers argues that the problems suffered by schizophrenics are eerily similar to the problems experienced by AI controlled agents. They are jumbles of behaviors and systems, that often do not quite come together. An agent’s plans might shift when one goals is dropped in favor of another. To seek a resolution to this problem, she turns to look at the anti-psychiatric movement which suggested a change in treatment for schizophrenics. This movement sought to instead of looking at patients as being composed of symptoms, to look at them as phenomenological wholes, and constructed narratives in the process of working through problems. Sengers uses this as a rationalization to look at agents narratively.

To compose agents narratively, Sengers observes: “if humans understand intentional behavior by organizing it into narrative, then our agents will be more ‘intentionally comprehensible’ if they provide narrative cues.” (p. 266) Sengers identifies three main principles for a narrative agent architecture, which is strongly supportive of the work I have done so far:

  • Context-sensitivity and negotiability: Meaning of events is dependent on context, and is subject to interpretation on the part of the observer.
  • Intentional state entailment: In addition to seeing what is being done, it is important to see why it is being done.
  • Diachronicity: Events and behaviors take place over time, an agent cannot change its behaviors quickly according to what is optimal.

Another interesting observation is that behaviors should be as simple as possible, with minimal cues. The observer should be assumed to do most of the interpretation. The implementation of this system involves a system of signifiers and transitions, rather than plans.

Philip E. Agre: Writing and Representation

Agre is looking to expand and establish the relationship between writing and representation as problematic. A lot of this is oriented toward understanding the contextual dependence of writing, and challenging the idea of world models. He explains that people make symbolic representations as a process of interpretation, and the interpretation of a text is dependent on setting, it is never simply transplanted.

For the most part, the given texts are instructions: recipes, directions, toner loading procedures, and instructions of how to observe a performance. These are not understood properly until given the right contextualization. The problem and content of a text does not lie within the text itself, but rather in identifying the situation, and determining how to connect the situation to the text in the correct way. While Agre is opposed to world models, I have to raise the question of what is the heart of situational understanding? How does one relate a situation to a narrative? I think that model theory is sufficient to handle these circumstances, by abstracting situational cues and elements. The text requires a situational model to be understood.

It is best to see a model as a lens, and not as a miniature.

Reading Info:
Author/EditorMateas, Michael and Sengers, Pheobe
TitleNarrative Intelligence
Tagsspecials, digital media, ai, narrative
LookupGoogle Scholar, Google Books, Amazon

No Comments »

No comments yet.

RSS feed for comments on this post. TrackBack URI

Leave a comment

You must be logged in to post a comment.