Shadows and Headless Shadows: an Autobiographical Approach to Narrative Reasoning

Ladislau BölöniarXiv 2012 · OpenReview

TL;DR

This paper introduces Xapagy, a cognitive architecture for narrative reasoning that exclusively uses autobiographical memory. It proposes a novel "shadowing" mechanism where current events are matched against past experiences, leading to "headless shadows" (HLSs) that enable surprise detection, inference of missing information, and story summarization. Xapagy's approach offers a unique perspective on how agents can understand, recall, and even confabulate stories.

What Problem Does This Paper Address?

The paper tackles the challenge of building artificial agents that can understand and reason about stories in a human-like way. Traditional approaches often rely on structured knowledge bases or complex reasoning engines. Xapagy, however, aims to mimic human narrative abilities by grounding all reasoning in a raw, autobiographical memory – essentially, a personal history of events. This approach seeks to address how an agent can infer meaning, predict continuations, and even generate new narratives solely from its past experiences, without pre-defined story schemas or explicit world models.

Methods

Xapagy's core innovation lies in its shadowing mechanism. When a new event (a "Verb Instance" or VI) occurs, Xapagy searches its autobiographical memory for similar past events. These matched past events form "shadows" of the current event. These shadows are then extrapolated into Headless Shadows (HLSs), which are collections of related past VIs that are not directly paired with a current event.

HLSs are crucial for several functions:

  • Tracking Surprise: By comparing current events to predicted continuations from HLSs, Xapagy can gauge the level of surprise.
  • Inferring Missing Information: HLSs can suggest hidden actions or relationships that might have occurred but weren't explicitly observed.
  • Summarization: HLSs can be used to condense sequences of events into more abstract representations.
  • Recall and Confabulation: In a "recall mood," HLSs can be used to reconstruct past stories, ranging from exact recall to free-form confabulation (generating new, potentially fictional, narratives).

The system uses a custom "pidgin" language called Xapi for representing stories and events. Xapi sentences are translated into Verb Instances (VIs), which are the fundamental units of Xapagy's internal representation.

Datasets

Not discussed in the paper. The paper uses the "Little Red Riding Hood" story as a running example to illustrate its concepts, but does not mention formal datasets for training or evaluation.

Key Insights and Ideas

  • Autobiographical Memory as Sole Knowledge Source: The radical idea that an agent can perform sophisticated narrative reasoning without explicit knowledge bases or ontologies, relying solely on a raw collection of past events.
  • Shadowing as a Core Reasoning Mechanism: The concept of matching current events to past experiences to create "shadows" provides a powerful way to leverage historical data for understanding and prediction.
  • Headless Shadows (HLSs) for Predictive and Inferential Power: HLSs are a key abstraction that allows the system to make predictions, infer missing information, and summarize events without direct current event anchors.
  • Mood as a Control Mechanism: The agent's "mood" (e.g., story-following vs. recall) significantly influences how HLSs are used, allowing for flexible narrative behavior.
  • Fragmentation of Self: The architecture's inability to reload past experiences and its reliance on recreating representations of the "self" leads to a unique concept of a fragmented self, where each recall can be a new story.

Key Results

The paper primarily focuses on describing the architecture and its mechanisms rather than presenting quantitative results. The key "results" are qualitative demonstrations of how Xapagy can:

  • Track the level of surprise in a narrative.
  • Infer hidden actions or relationships (e.g., inferring that LRRH might pick up a basket).
  • Summarize events.
  • Generate new stories through recall and confabulation.

The paper includes figures illustrating the scene and identity structure of the "Little Red Riding Hood" story within Xapagy and screenshots of the shadowing and HLS panels, visually demonstrating the system's internal state.

Limitations and Future Work

  • Lack of Explicit Learning: Xapagy's autobiographical memory is a recording mechanism; it doesn't explicitly "learn" in the traditional sense of extracting general rules or updating its core knowledge structures. The paper notes that learning is "only a recording of the experiences."
  • Fragmentation of Self: While presented as a unique implication, the fragmentation of the self could be seen as a limitation in achieving a stable, consistent agent identity over time.
  • Computational Complexity: The paper doesn't deeply explore the computational challenges of managing and searching a vast autobiographical memory, especially for complex narratives.
  • Future Work: The authors suggest exploring how the agent's "mood" can be dynamically set and how HLSs can be used for more complex reasoning tasks. They also highlight the potential for Xapagy to serve as a foundation for more advanced cognitive architectures.

Summary updated

Abstract

The Xapagy architecture is a story-oriented cognitive system which relies exclusively on the autobiographical memory implemented as a raw collection of events. Reasoning is performed by shadowing current events with events from the autobiography. The shadows are then extrapolated into headless shadows (HLSs). In a story following mood, HLSs can be used to track the level of surprise of the agent, to infer hidden actions or relations between the participants, and to summarize ongoing events. In recall mood, the HLSs can be used to create new stories ranging from exact recall to free-form confabulation.
BibTeX
@misc{boloni2012shadows,
  author = {Ladislau Bölöni},
  title = {{Shadows and Headless Shadows: an Autobiographical Approach to Narrative Reasoning}},
  year = {2012},
  eprint = {1211.6097},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1211.6097},
}
Figure