Introduction
In the evolution of natural language processing (NLP), early systems such as BORIS and PAM achieved significant milestones by tracking character goals and plan execution. However, these systems remained fundamentally limited by their lack of a "moral compass." While they could identify that an action was taken to achieve a specific objective, they were unable to comprehend why a reader might find that action reprehensible, ironic, or satisfying. This paper presents THUNDER (THematic UNDerstanding from Ethical Reasoning), a framework that argues story understanding is not merely a logical exercise but a subjective one, deeply rooted in the reader's ethical expectations.
The Problem: Beyond Logical Planning
The authors posit that traditional AI models of the late 1980s failed to capture the essence of narrative because they ignored the reader's subjective attitude. Without a model of ethics, a system cannot recognize belief conflicts—instances where story events violate the reader's sense of right and wrong. These conflicts are essential for grasping higher-level thematic elements such as retribution, irony, and poetic justice. For an AI to truly "understand" a story, it must move beyond tracking what characters do and begin modeling how a reader judges those actions.
The THUNDER Framework
THUNDER is an explanation-based model that operates across a three-tiered hierarchy: acts and events, goals and plans, and finally, themes. The system bridges the gap between simple event tracking and thematic extraction through two primary mechanisms:
1. Character Assessments
THUNDER categorizes characters based on the moral quality of their actions and the outcomes they experience:
- Virtuous Positive: Characters who achieve their goals through adherence to moral principles.
- Sympathetic Positive: Characters who suffer failures through no fault of their own.
- Negative: Characters who intentionally cause goal failures for others.
2. Belief Conflict Patterns (BCPs)
BCPs are abstract knowledge structures representing interpretations that are "problematic" or noteworthy to a reader. A BCP might represent a situation where a "bad" character achieves success or where a power relationship is exploited. When a story event violates a moral expectation, a BCP is triggered, signaling a conflict that requires thematic resolution.
Thematic Resolution and Irony
When THUNDER encounters a belief conflict, it searches for a resolution that restores a sense of thematic order. For example, if a "negative" character who previously caused harm later suffers a failure, the system does not just see a failed plan; it interprets the event as "retribution." This allows the system to abstract a universal theme, such as: "It is wrong to take advantage of others because it may result in your own failure."
Irony is modeled as a specific type of resolution where the tension between the reader's beliefs and the story's events is highlighted by a reversal of fortune. By representing the understander's role and their specific belief system, THUNDER can identify situational irony that purely goal-oriented models would overlook.
Evaluation: The Gelignite Story
The system was evaluated using a corpus of short ironic stories, most notably "The Gelignite Story." In this narrative, two hunters tie dynamite to a rabbit; the rabbit subsequently runs under their truck, destroying it.
While a purely logical system might suggest the hunters simply made a tactical error (e.g., they should have restrained the rabbit), THUNDER recognizes the hunters' actions as morally negative. When their truck is destroyed, the system identifies a resolution to the belief conflict, categorizing the outcome as poetic justice and extracting the underlying theme of the narrative.
Technical Performance
Implemented in approximately 4.8K lines of T (a Lisp dialect), THUNDER utilized seven distinct Memory Organization Packets (MOPs) to represent complex social concepts like M-Sado-Pleasures and M-Revenge. On an Apollo DN3000 workstation, the system processed "The Gelignite Story"—from raw text parsing to thematic extraction—in 327 seconds of CPU time.
Key Insights: Ethics as a Search Constraint
One of the paper's most significant contributions is the idea that moral judgment serves as a computational search constraint. By applying an ethical lens, the system can bypass "logical" but narratively irrelevant explanations. Instead of calculating every possible optimization for a character's plan, the system focuses on thematic resolutions. This suggests that subjectivity is not a hindrance to AI but a necessary component for efficient and human-like narrative processing.
Limitations and Future Work
Despite its successes, THUNDER’s "morality" is essentially a reflection of its programmer's beliefs. The authors acknowledge that different readers bring different ethical frameworks to a text, which would result in different thematic interpretations. Furthermore, the system relies on hand-crafted BCPs and MOPs, presenting a scalability challenge for broader applications.
Future research directions include implementing structural reasoning models, such as Kohlberg’s stages of moral development, to explore how varying levels of "moral maturity" might change a system's interpretation of the same story. This would move AI closer to simulating the diverse range of human perspectives found in literary criticism and social discourse.
Summary updated