TL;DR
This paper introduces "algorithmic equivalence" as a way to measure narrative complexity, defining two stories as equivalent if they can be summarized by the same retrieval process. This shifts focus from static narrative structures to the computational demands of understanding them, proposing plot unit graphs and their structural features as a basis for this equivalence.
What Problem Does This Paper Address?
Traditional approaches to understanding narrative complexity often focus on structural representations like story grammars or causal chains. However, these representations can be detached from the actual cognitive processes involved in understanding and summarizing a story. The authors argue that narrative complexity is better characterized by the information processing requirements needed to comprehend and summarize a narrative. They aim to bridge this gap by developing a framework that links narrative structure to computational complexity, specifically through the lens of summarization algorithms.
Methods
The core of the paper's methodology lies in the plot unit approach and the concept of algorithmic equivalence.
- Plot Units: Narratives are broken down into "plot units," which are fixed configurations of "affect states." Affect states represent a character's subjective mental states (positive, negative, neutral) and are tracked chronologically in an "affect state map."
- Plot Unit Graphs: When a narrative is represented by plot units, a graph is constructed where nodes represent top-level plot unit instantiations. Arcs connect nodes that share a common affect state. This graph structure is designed to be well-suited for text summarization.
- Algorithmic Equivalence (a-equivalence): Two narratives are considered algorithmically equivalent if their plot unit graphs can be effectively summarized by the same retrieval process. This means that narratives with isomorphic (structurally identical) plot unit graphs are a-equivalent, but the concept is broader, allowing for non-isomorphic graphs to be a-equivalent if they can be processed by the same summarization algorithm.
- Universal Sets of Plot Units: To analyze plot unit graphs, the authors propose using "universal sets" of plot units. Different universal sets can account for variations in how individuals or computational models might parse and understand narratives, allowing for analysis across different "universal contexts" (e.g., focusing on narratives with few characters vs. many).
- Graph Analysis: The paper discusses analyzing plot unit graphs for critical nodes (e.g., nodes of maximal degree) and structural features (clusters, partitions) to identify concepts central to the narrative for summarization.
Datasets
The paper does not explicitly list or describe specific datasets used for training or evaluation. Instead, it mentions compiling a "library of plot unit graph representations" from various sources, including:
- Narratives from AI work in natural language processing.
- Cognitive psychology research on text comprehension.
- Published short stories by popular authors.
The authors state that they have analyzed "about 20 narratives" with respect to universal contexts A and B.
Key Insights and Ideas
- Complexity is Computational: The most significant insight is the reframing of narrative complexity from a structural property to a computational one. This means that understanding how a story is processed is key to understanding its complexity.
- Algorithmic Equivalence as a Measure: The concept of a-equivalence provides a novel way to categorize narratives based on the algorithms required for their summarization, offering a more nuanced view than simple structural isomorphism.
- Plot Unit Graphs as a Summarization Tool: The plot unit graph structure, with its emphasis on affect states and inter-unit connections, is presented as a powerful intermediate representation for identifying central narrative concepts for summarization.
- Context-Dependent Summarization: The idea of "universal sets" and "universal contexts" highlights that summarization strategies might need to adapt based on the complexity of the narrative components (e.g., number of characters) and potentially individual differences in cognitive processing.
- Potential for Distinct Complexity Plateaus: The authors suggest that algorithmic equivalence could lead to distinct "plateaus of summarization competence," implying that performance might not increase linearly with narrative length or representation size, but rather in steps as algorithms become applicable.
Key Results
The paper presents preliminary findings based on their analysis of about 20 narratives. They have identified three core "equivalence classes" of graph structures that appear to produce reasonable summaries:
- A. Simple Graphs with Unique Pivots: Graphs with a single node of maximal degree, often found in smaller narratives, where this pivotal node provides central concepts.
- B. Complex Graphs with Multiple Pivots: Larger graphs that may have two nodes of maximal degree, either partitioning the graph or providing critical concepts when adjacent.
- C. Separable Graphs: Large graphs (>50 nodes) that can be separated by deleting a single node. Summarization here involves finding shortest paths through deletion nodes.
These classes are presented as initial hypotheses, and the authors emphasize the need for more analysis to refine them.
Limitations and Future Work
- Manual Encoding: A significant limitation acknowledged is the need to "hand-code" affect state maps for each narrative. This process is described as "highly domain-dependent" and would require extensive knowledge encoding for each new narrative. They are developing encoding heuristics to aid this process.
- Limited Dataset: The analysis is based on a relatively small sample of "about 20 narratives."
- Context C Development: The paper notes that they have not yet formulated plot units for "context C" (narratives involving more than two characters), which is expected to be more complex.
- Future Work:
- Expanding the library of affect state maps and plot unit graphs.
- Developing a more systematic methodology for identifying graph equivalencies.
- Investigating the assignment of relative weights to nodes in complex graphs.
- Exploring the concept of "minimal eccentricity" for identifying central nodes in larger graphs.
- Further research into the psychological import of algorithmic equivalence and its relation to developmental studies of summarization.
- Investigating whether this approach is applicable to expository or instructional texts, or if they represent fundamentally different equivalence classes.
Summary updated