Abstract
In the 1980s, plot units were proposed as a conceptual knowledge structure for representing and summarizing narrative stories. Our research explores whether current NLP technology can be used to automatically produce plot unit representations for narrative text. We create a system called AESOP that exploits a variety of existing resources to identify affect states and applies "projection rules" to map the affect states onto the characters in a story. We also use corpus-based techniques to generate a new type of affect knowledge base: verbs that impart positive or negative states onto their patients (e.g., being eaten is an undesirable state, but being fed is a desirable state). We harvest these "patient polarity verbs" from a Web corpus using two techniques: co-occurrence with Evil/Kind Agent patterns, and bootstrapping over conjunctions of verbs. We evaluate the plot unit representations produced by our system on a small collection of Aesop's fables.
Introduction
Plot units represent stories through the emotional states (affect) of characters and the causal links between them. For example, a character's "annoyance" might motivate a "plan," which leads to an "action." While powerful for narrative analysis and summarization, plot units traditionally required extensive manual knowledge engineering. This research introduces AESOP, a system designed to automate the generation of these structures from raw text.
The AESOP System Architecture
AESOP processes narrative text through a four-stage pipeline:
- Affect State Recognition: The system identifies positive (+), negative (-), and mental (M) states using resources like FrameNet, the MPQA Lexicon, and OpinionFinder.
- Character Identification: Focusing on fables with two main characters, AESOP uses rule-based heuristics and WordNet to resolve pronouns and identify protagonists.
- Affect State Projection: Rather than assigning sentiment globally, AESOP uses syntactic "projection rules." If a verb has a negative polarity for the patient (e.g., "killed"), that state is projected onto the direct object, while a positive state (success) may be projected onto the agent.
- Link Creation: Heuristics are applied to create causal links (e.g., motivation) and cross-character links (e.g., shared requests).
Patient Polarity Verbs (PPVs)
A core innovation of this work is the development of a Patient Polarity Verb (PPV) lexicon. Standard sentiment analysis often focuses on the speaker's opinion, but narratives depend on how actions affect the "patient" (the character being acted upon). To build this lexicon, the researchers used two corpus-based techniques:
- Evil/Kind Agent Patterns: Verbs were harvested from the Google Web 1T N-gram corpus based on their co-occurrence with agents like "villain" vs. "hero."
- Bootstrapping: The "Basilisk" algorithm was used to learn new PPVs by analyzing verbs joined by conjunctions (e.g., "abducted and [X]").
Evaluation and Results
The system was evaluated using a collection of 34 Aesop's fables, with a manually annotated "gold standard" for comparison.
As shown in the experimental results, AESOP achieved an overall F-score of 45% for affect state recognition. A key finding was the significant impact of the novel PPV lexicon; adding "Evil Agent" PPVs improved the F-score from 39% to 44%, largely by doubling the recall of negative states.
Further analysis revealed that while AESOP's heuristics for linking states are highly accurate when provided with perfect data (F-scores of 0.82 to 0.92), performance drops when using system-generated states. This indicates that errors in initial affect recognition propagate through the pipeline.
Conclusion and Future Work
AESOP represents a significant step toward the automated conceptual analysis of stories. The research highlights that identifying "Patient Polarity" is crucial for narrative conflict, and that negative states are generally easier to track than abstract positive ones. Future work will focus on improving the recognition of mental states (goals and plans) and scaling the system to handle more complex narratives with larger character casts and nuanced subplots.
Summary updated