TL;DR
This paper proposes a novel AI framework for story comprehension that integrates "Reasoning about Actions and Change" with "Default Reasoning" using argumentation theory. It aims to computationally model how humans integrate explicit story text with implicit common-sense world knowledge to build a coherent understanding, and has been empirically evaluated against human readers.
What Problem Does This Paper Address?
Understanding stories is a fundamental challenge for Artificial Intelligence, going beyond simple information extraction to require deep reasoning about causality, implicit assumptions, and evolving situations. Humans excel at this by seamlessly blending what's explicitly stated in a narrative with their vast store of common-sense knowledge about how the world works. This paper tackles the problem of building an AI system that can perform this kind of sophisticated story comprehension. Specifically, it addresses how to represent and reason with narratives, handle the "frame problem" (what doesn't change when an action occurs), the "qualification problem" (what conditions must hold for an action to occur), and the "ramification problem" (what are the indirect effects of an action), all while accounting for the defeasible nature of common-sense knowledge.
Methods
The core of the proposed solution is a Knowledge Representation and Reasoning (KRR) framework built upon argumentation theory from AI. This framework integrates two key AI areas:
- Reasoning about Actions and Change (RAC): This deals with how the state of the world changes over time due to actions.
- Default Reasoning: This allows for reasoning with incomplete or uncertain information, where conclusions are drawn based on what is typically true, but can be revised if new information contradicts them.
The authors formalize this by defining "unit arguments" (basic pieces of knowledge like properties, causes, and persistence rules) and then constructing more complex "arguments" that represent a sequence of inferences. A crucial aspect is the use of a preference-based argumentation semantics. This means that when conflicting inferences arise (which is common with default reasoning), the system uses a priority relation to determine which argument is stronger and thus preferred. This preference relation is extended to handle the frame, ramification, and qualification problems. The framework also incorporates a mechanism for revision, allowing the comprehension model to be updated as new information from the story emerges.
Datasets
The paper does not explicitly list or introduce formal datasets for training. Instead, it uses a running example story ("It was the night of Christmas Eve...") throughout the paper to illustrate its framework. For empirical evaluation, they describe a methodology involving:
- A set of stories and questions designed to test different aspects of story comprehension.
- Harnessing human-generated common-sense world knowledge through qualitative studies (think-aloud protocols) to inform the system's knowledge base.
Key Insights and Ideas
- Argumentation as the unifying mechanism: The paper argues that argumentation theory provides a natural and powerful way to formalize story comprehension by accounting for defeasible inferences, conflicts, and revisions.
- Bridging explicit text and implicit knowledge: The core challenge is integrating what's in the story with what the reader "knows" without being told. The framework models this by using world knowledge as premises for arguments that elaborate on the explicit narrative.
- Psychologically-inspired framework: The approach is heavily guided by findings from the psychology of reading and comprehension, aiming to mirror human cognitive processes.
- Handling defeasible knowledge: Common-sense knowledge is inherently uncertain. The preference-based argumentation allows the system to make default assumptions but also to revise them when the story provides contradictory evidence.
- Empirical validation is key: The authors emphasize the importance of evaluating their computational model against human readers to ensure its psychological plausibility and effectiveness.
Key Results
The paper presents a formal framework and a prototype system. While specific quantitative results on benchmark datasets are not detailed (as it's more of a foundational framework paper), the key qualitative findings are:
- The framework successfully models the integration of explicit narrative with implicit world knowledge.
- It can capture solutions to the frame, ramification, and qualification problems using argumentation.
- The system's behavior, when tested with their example story and evaluated against human responses, shows promise in capturing both the majority understanding and the variability observed in human readers. For instance, the system's reasoning about Papa Joe's location and motive aligns with common human interpretations and their variations.
Limitations and Future Work
The authors acknowledge several limitations and outline future directions:
- Natural Language Processing (NLP) bottleneck: The current framework assumes that the narrative has already been extracted from free-format text into a structured form. A significant future task is to develop robust NLP techniques for this extraction.
- Computational Economy and Coherence: While the framework is theoretically sound, efficiently computing comprehension models and ensuring global coherence remains a challenge. They plan to address this by applying "computational heuristics" inspired by psychology, such as selection, dropping, and generalization operators.
- Systematic Acquisition of World Knowledge: A major ongoing effort is to semi-automatically acquire common-sense world knowledge from the web, leveraging existing lexical databases and NLP techniques.
- Non-Sceptical Entailment: The current definition of a comprehension model is "sceptical" (only accepting conclusions that are definitively supported). Future work may explore more "admissible" forms of entailment to better model situations where humans might choose between multiple plausible interpretations.
- Further Empirical Studies: The authors are continuing to conduct psychological experiments with more stories to gather more world knowledge and further validate their system.
Summary updated