A Cognitive Model of Coherence-Driven Story Comprehension

Elliot SmithACL 1998

TL;DR

This paper proposes a new cognitive model for story comprehension where the primary driver is the reader's need to achieve a "coherence threshold." Instead of relying solely on prediction or fixed inference rules, this model treats inferences as "coherence generators," with their application determined by their contribution to the overall coherence of the story representation. This approach offers a more flexible and experimentally grounded explanation for how humans understand narratives.

What Problem Does This Paper Address?

Current models of story comprehension often struggle to explain how humans actually process narratives. Three key deficiencies are highlighted:

  1. Lack of Experimental Support for Inference Processes: Many models rely heavily on predictive inferences (guessing what comes next), but experimental evidence suggests readers are more cautious, only making predictions when strongly supported by the text.
  2. Indifference to Kinds of Coherence: Models often don't distinguish between "local coherence" (connections between adjacent sentences) and "global coherence" (overarching themes). Some argue only local coherence is consistently established, while others suggest readers aim for both.
  3. Inability to Handle Variable Depths of Comprehension: Existing models tend to assume comprehension is an "all-or-nothing" process, either fully instantiated or not. They lack flexibility to represent understanding at different levels of detail or to stop processing when a satisfactory level of coherence is reached.

This paper aims to bridge these gaps by proposing a model where the need for coherence drives comprehension, allowing for more nuanced and flexible interpretation.

Methods

The paper introduces a novel metric for measuring the "coherence contribution" of a schema (a mental representation or knowledge structure) when applied to a set of observations (textual input). This metric is defined as:

C = (V x U) - (P x S)

Where:

  • C is the coherence contribution.
  • V is Coverage: How many elements of the input are matched by the schema.
  • U is Utility: The inherent usefulness or "size" of the schema (larger schemas are considered more useful, reflecting chunking principles).
  • P is Completion: The percentage of the schema's conditions that are matched by the input.
  • S is Skepticism: A parameter representing the comprehender's unwillingness to "jump to conclusions."

This metric is applied within a framework where comprehension is driven by the need to reach a "coherence threshold." Inferences are seen as "coherence generators," and their application depends on their calculated coherence contribution. This allows for "satisficing" – achieving a representation that meets a minimal coherence constraint, rather than necessarily an optimal one.

Datasets

Not discussed in the paper. The paper focuses on theoretical modeling and provides a hypothetical example rather than empirical evaluation on specific datasets.

Key Insights and Ideas

  • Coherence as the Driving Force: The core idea is that the reader's goal isn't just to predict or understand every detail, but to build a representation that feels coherent. This "coherence threshold" is the primary motivator.
  • Inferences as Coherence Generators: Instead of being passive steps, inferences are active tools that create coherence. Their utility is judged by how much they contribute to this goal.
  • The Role of Skepticism: The "Skepticism" parameter (S) is crucial. High skepticism means an inference must be strongly supported to contribute positively to coherence, preventing hasty or unsupported interpretations. This aligns with experimental findings about cautious inference-making.
  • Satisficing vs. Optimizing: The model embraces "satisficing" – reaching an acceptable level of coherence – rather than striving for a perfect, complete representation. This acknowledges that real-world comprehension is often efficient and pragmatic.
  • Variable Precision Comprehension: By adjusting the coherence threshold and skepticism, the model can account for how readers might understand a story at different levels of detail, depending on their goals and the text's demands.

Key Results

The paper does not present quantitative results from empirical studies. Instead, it offers a conceptual framework and a hypothetical example to illustrate its principles:

  • Example of Skepticism's Impact: A hypothetical scenario demonstrates how increasing the "Skepticism" parameter can significantly alter the coherence contribution of a schema, making it less likely to accept an assumption (like "leave(john, restaurant)") without sufficient evidence. This shows how the model can modulate inference acceptance.
  • Theoretical Framework: The primary "result" is the proposed model itself, which offers a more nuanced explanation for phenomena like differential coherence establishment (local vs. global) and the cautious nature of predictive inferences.

Limitations and Future Work

  • Experimental Validation: The authors acknowledge that the proposed metric and model require experimental validation to demonstrate their empirical adequacy.
  • Incremental Derivation of Coherence: The current metric derives coherence from a static set of observations. Future work needs to address how coherence is incrementally derived during the comprehension process.
  • Cost of Assuming Conditions: The paper mentions that the cost of assuming various conditions is not fully addressed. Future work will involve developing methods for deriving weights from condition distributions over the knowledge base to address this.
  • Global Skepticism Parameter: Skepticism is currently a global parameter. The authors suggest exploring local weights attached to individual conditions as a future direction.

Summary updated

BibTeX
@inproceedings{smith1998cognitive,
  author = {Elliot Smith},
  title = {{A Cognitive Model of Coherence-Driven Story Comprehension}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {1998},
}

Loading related content...

Figure