Narrative Complexity Based on Summarization Algorithms

Wendy G. Lehnert

Abstract

Traditional measures of narrative complexity often rely on surface-level metrics such as word count or structural depth. This paper proposes an alternative paradigm: narrative complexity as a function of the algorithms required to summarize a story. By representing narratives as graphs of "plot units"—primitive social interactions derived from character affect states—we can categorize stories into equivalence classes based on their computational processing requirements. This approach, termed algorithmic equivalence, provides a topological framework for understanding the "gist" of a narrative and offers insights into the cognitive processes involved in text comprehension.

Introduction

In the early 1980s, the field of Artificial Intelligence sought to formalize the concept of narrative complexity. While story grammars and goal hierarchies provided internal structural representations, they often overlooked the process of information retrieval. Wendy Lehnert argued that if two narratives of vastly different lengths can be summarized using the same retrieval strategy, they share a fundamental "algorithmic complexity." This paper details a methodology for extracting the conceptual core of a story by analyzing the connectivity of its underlying plot unit graph.

The Plot Unit Framework

The core of this research is the Plot Unit representational framework, which transforms raw narrative text into a structured graph through a three-stage process.

1. Affect State Mapping

Narratives are first hand-coded into a chronological sequence of "affect states" for each character. These states are simplified into three categories: positive (+), negative (-), and neutral mental states (M). This mapping captures the emotional trajectory and intentions of the characters.

Mapping character intentions and reactions to discrete affect states.

2. Plot Unit Instantiation

These individual affect states are then grouped into "plot units"—fixed configurations that represent primitive social interactions. For example, a "denied request" occurs when one character's mental state (M) is met with another's negative reaction (-). These units serve as the building blocks of the narrative's social logic.

Examples of primitive plot units representing social interactions.

3. Graph Generation

Finally, the Plot Unit Graph Generator (PUGG) creates a network where each node is a top-level plot unit. An edge is drawn between two nodes if they share a common internal affect state, creating a topological map of the story's interconnected events.

Algorithmic Equivalence and Complexity

The methodology introduces Algorithmic Equivalence (a-equivalence): two narratives are considered equivalent if their plot unit graphs can be condensed using the same summarization algorithm. This shifts the focus from narrative length to graph topology.

Key strategies identified include:

  • The Pivot Strategy: In simple, highly connected graphs, the node with the highest degree (most connections) typically represents the story's "gist."
  • Graph Partitioning: For more complex narratives, the summary often resides in "boundary units"—nodes that bridge distinct clusters of events.
  • Separable Graphs: Very long narratives (such as synopses of the life of Jesus) often form separable graphs where "deletion nodes" (articulation points) must be identified to break the story into manageable sub-plots for summarization.

Results and Insights

Analysis of a library of approximately 20 narratives revealed that narrative complexity is fundamentally topological. A long story with a single central "pivot" is computationally simpler to summarize than a short story with a dense, multi-pivoted graph. Furthermore, structural centrality in these graphs correlates strongly with human perceptions of importance, suggesting that these computational models mirror cognitive summarization processes.

Limitations and Future Directions

The primary bottleneck of this approach is the manual effort required for affect state mapping, which remains difficult to automate. Additionally, the current taxonomy is based on a small sample size and is specifically tailored to human-centric narratives driven by emotion. Future work should investigate whether these algorithmic classes align with human developmental stages in reading comprehension and explore the applicability of this framework to non-narrative texts.

Summary updated

BibTeX
@inproceedings{lehnert1983narrative,
  author = {Wendy G. Lehnert},
  title = {{Narrative Complexity Based on Summarization Algorithms}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {1983},
}
Figure