Narrative Text Summarization

Wendy G. LehnertAAAI 1980

Abstract

This paper proposes a novel approach to narrative summarization by mapping the "emotional trajectory" of characters. By representing plots as graphs of affect units—interlocking mental states and events—the system identifies central plot points, or "pivotal units," to generate concise, high-level summaries that capture the conceptual core of a story.

Introduction

Summarizing a story requires more than identifying frequent keywords; it necessitates an understanding of the narrative's "gist." In the early 1980s, the dominant approach to this problem was "story grammars," which utilized rigid, top-down templates to categorize stories into fixed hierarchies. Wendy Lehnert argues that these grammars are too brittle to handle the diversity of human storytelling. Instead, she proposes a bottom-up method focused on narrative cohesion: selectively ignoring peripheral details to focus on the central conflict.

The Affect Unit Framework

The core of this method is a representational framework called Affect Units. These units are built from three primary states and four types of relational links.

The Building Blocks

Affect units are composed of:

  • Mental States (M): A character's internal goals or desires (e.g., "John wants a house").
  • Positive Events (+): Successes or achieved goals.
  • Negative Events (-): Failures or thwarted goals.
Basic building blocks of Affect Units showing "motivation" (an event causing a mental state) and "failure" (a desire leading to a negative event).

These states are connected by four types of links:

  1. Motivation (m): An event causes a desire.
  2. Actualization (a): A desire leads to an event.
  3. Termination: An event ends a state.
  4. Equivalence (e): Two states are essentially the same.

Graph Theory for Plots

By overlapping these units, the system constructs a graph where nodes represent "top-level affect units" such as competition, retaliation, or fleeting success. The node with the highest degree of connectivity is designated the pivotal unit, representing the core of the story. To produce text, the system utilizes a "generational frame" associated with this pivotal unit, augmenting it with information from adjacent nodes to add necessary detail.

A complete Affect Unit graph mapping the narrative of John and Mary competing for a house.

Key Insights and Results

The paper demonstrates that the "gist" of a story is structural rather than purely thematic. The most important element is the conceptual node that shares the most "connective tissue" with other parts of the narrative.

In complex or symmetrical stories, such as O. Henry’s "The Gift of the Magi," there may not be a single pivotal unit. In these instances, summaries are derived from "boundary units"—the points where the characters' subgoals intersect.

Lehnert provides a qualitative proof-of-concept using a scenario involving John and Mary. By identifying the "retaliation" unit as the pivotal node, the system can evolve a summary from a bare-bones baseline to a nuanced narrative:

  • Baseline: "When Mary prevented John from getting something he wanted, John set her house on fire."
  • Refined: "When Mary bought a house that John wanted, John unsuccessfully tried to set it on fire."

Limitations and Future Work

While the framework provides a robust theory for narrative structure, several challenges remain:

  • Recognition Gap: The paper does not detail how a computer would automatically recognize affect units from raw natural language text.
  • Generation Complexity: Turning abstract graph nodes into natural-sounding English remains a significant area for future development.
  • Scalability: It is unclear how this graph-based approach would scale to handle long-form literature with dozens of characters and complex subplots.
  • Formal Logic: The author notes that only 15 of 36 possible state/link configurations are "legal," suggesting the need for a more formal logic of affect.

Summary updated

BibTeX
@inproceedings{lehnert1980narrative,
  author = {Wendy G. Lehnert},
  title = {{Narrative Text Summarization}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1980},
}
Figure