Summarizing Narratives

Wendy G. Lehnert, John B. Black, Brian J. ReiserIJCAI 1981

Abstract

Traditional models of text summarization often rely on "story grammars"—hierarchical tree structures that prioritize information based on its structural level. This research challenges that paradigm, proposing a "plot unit" model based on heterarchical graphs of emotional states and events. Our findings suggest that the connectivity of an element within a story graph, rather than its position in a hierarchy, more accurately predicts its inclusion in human-generated summaries.

Introduction: The Limitations of Story Grammars

In the early 1980s, experimental psychology was dominated by the "story grammar" approach. This model posited that stories possess a rigid, hierarchical syntax similar to sentences. Under this theory, nodes higher in the tree are deemed more "important" and thus more likely to be remembered or summarized. However, AI researchers observed that minor structural tweaks could significantly alter human summaries in ways these hierarchical models failed to explain. The core issue was a lack of representation for the complex interconnectedness of character goals, affect states, and events.

The Plot Unit Model: Moving from Trees to Graphs

To address these shortcomings, the authors introduced the Plot Unit Model. Unlike the rigid hierarchy of story grammars, this model utilizes a heterarchical graph where nodes represent "plot units"—complex structures built from three primitive affect states: mental states (intentions), positive events, and negative events.

Definitions of key plot units used in the heterarchical graph model, including 'denied requests' (DR) and 'request honored with condition' (RHC).

A defining feature of plot units is their ability to overlap. For instance, a single event might simultaneously represent a "success" for one character and an "honored request" for another. This overlapping creates a dense web of connections that better reflects the narrative's emotional and causal logic.

Experimental Methodology

The researchers tested their model using a series of controlled narrative variations:

  • The "Mike & Paul" Baseline: A story involving Mike’s attempt to surprise his friend Paul with a patio, requiring a complex series of subgoals, including bribing an accountant with marijuana.
  • Structural Variations (S1, S2, S3): Three versions of the story with subtle modifications designed to test specific predictions about connectivity and importance.
  • The "Farmer/Donkey" Control: A structurally identical story with different content (a farmer wanting to attend a dance) used to isolate the effects of "self-containment" on summarization.

Data was collected from 72 human subjects, who provided concise one- or two-sentence summaries for each narrative.

Results: Connectivity as a Predictor of Importance

The plot unit model significantly outperformed story grammars in predicting which details would appear in summaries. The primary insight was that highly connected nodes—regardless of their "level" in a tree—act as "summary magnets."

A key test case involved the "bribe" (categorized as a 'request honored with condition', or RHC). While story grammar predicted the bribe would be mentioned with equal frequency across all versions, the plot unit model predicted that its inclusion would depend on its connectivity strength within the graph.

Connectivity strength of the 'bribe' plot unit (RHC1) across three story variations, showing a direct correlation with summary inclusion rates.

As shown in the data, the bribe's connectivity (and the resulting 'generational frame' strength) increased from S1 to S3, and human summaries followed this trend precisely.

Content-Driven Factors: Self-Containment and Subsumption

The study also identified two critical factors that override pure structure:

  1. Generational Subsumption: This occurs when one plot unit is naturally "swallowed" by another in natural language. For example, stating that a character "arranged a deal" might implicitly summarize several underlying subgoals, effectively removing them from the explicit summary.
  2. Self-Containment: Humans are driven to explain "mysterious" or non-self-contained goals. In the "Mike & Paul" story, the patio goal was always included because it explained Mike's otherwise bizarre behavior. When the goal was changed to something self-contained (like a farmer going to a dance), inclusion rates plummeted, proving that content-driven psychological needs play a major role in summarization.

Conclusion

While story grammars provide a useful framework for top-level goals, the plot unit model offers a far more robust explanation for the nuances of human summarization. By focusing on connectivity and the psychological drive for explanation, researchers can develop more sophisticated algorithms that capture the true essence of a narrative.

Summary updated

BibTeX
@inproceedings{lehnert1981summarizing,
  author = {Wendy G. Lehnert and John B. Black and Brian J. Reiser},
  title = {{Summarizing Narratives}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {1981},
}
Figure