TL;DR
This paper proposes a new approach to formalizing commonsense reasoning by defining "truth in intended models" rather than minimizing extensions of predicates. It introduces a three-level framework (meta, object, referential) where natural language constraints guide the selection of plausible interpretations, offering a more robust solution to commonsense reasoning problems than traditional circumscription methods.
What Problem Does This Paper Address?
Formalizing commonsense reasoning has been a long-standing challenge in Artificial Intelligence. Traditional approaches, like McCarthy's circumscription, aim to capture commonsense by minimizing "abnormal" situations. However, these methods struggle with certain types of inferences, particularly those involving temporal reasoning (as highlighted by the Hanks and McDermott paradox), and often require complex, problem-specific axioms. The authors argue that commonsense reasoning isn't always about minimizing abnormality but rather about selecting the most plausible interpretation from a range of possibilities, a task that natural language can significantly aid. They aim to provide a formal framework that leverages linguistic information to achieve this selection process.
Methods
The core of the proposed method is the concept of "intended models" and a three-level reasoning framework:
- Object Level: Represents the specific situation or problem being reasoned about (e.g., facts about Tweety).
- Meta Level: Specifies the general principles for constructing intended models (e.g., "minimize abnormality").
- Referential Level: Provides situation-independent interpretations of symbols, primarily derived from natural language (dictionaries, grammars). This level constrains possible interpretations of predicates.
Instead of minimizing extensions, the authors formalize commonsense reasoning as "truth in intended models." They define a class of intended models based on preferences in interpretations. Natural language constraints are used to exclude interpretations that contradict common sense. This approach differs from circumscription by forcing logical inferences to conform to the references of predicates, rather than relying solely on minimizing abnormality. They introduce the idea of "preferred interpretations," where linguistic information (like dictionary definitions) establishes a partial order of preference among possible interpretations of terms.
Datasets
Not discussed in the paper. The paper focuses on theoretical formalization and uses examples from existing literature and hypothetical scenarios.
Key Insights and Ideas
- Commonsense as Plausible Interpretation Selection: The paper argues that commonsense reasoning is often about selecting the most plausible interpretation from a set of possibilities, rather than solely minimizing abnormality.
- The Power of the Referential Level: Natural language, through dictionaries and grammars, serves as a crucial "referential level" that provides situation-independent constraints on symbol interpretation, guiding the selection of plausible models.
- Three-Level Framework: The proposed meta, object, and referential levels offer a more structured way to handle commonsense reasoning, separating general reasoning principles, specific problem states, and linguistic knowledge.
- Linguistic Constraints for Non-Monotonicity: Changes in preferences on the referential level (e.g., new dictionary entries or grammar rules) can naturally explain the non-monotonicity observed in commonsense reasoning, without needing complex meta-level rules.
- Addressing the Hanks and McDermott Paradox: The framework offers a solution to the paradox by using linguistic constraints (specifically, the interpretation of "unless") to disambiguate between competing models, demonstrating its superiority over pure circumscription for certain problems.
Key Results
The paper presents a formal framework and demonstrates its capabilities through examples:
- Tweety the Canary Example: The framework successfully deduces that Tweety can fly by using dictionary definitions as a referential level, and it also handles cases where Tweety might not fly (e.g., if it's a penguin or has no wings) by showing how changes in the object theory lead to different partial models and conclusions.
- Hanks and McDermott Paradox Solution: The proposed method, by interpreting "unless" based on linguistic constraints, correctly selects the intended model where the person dies, resolving the paradox that troubled circumscription.
- Generalization: The authors suggest their approach is more general than circumscription, capable of handling a wider range of commonsense inferences, including those not based on minimization.
Limitations and Future Work
- Formalization of Natural Language: While the paper emphasizes the role of natural language, the actual process of parsing and transforming linguistic information into logical constraints is acknowledged as an ongoing area of research (mentioning tools like PEG and PLNLP).
- Completeness: The authors admit that it's an open question how much of commonsense reasoning can be captured by their framework and how to measure "common sense" itself.
- Frame Problem: The paper does not directly address the frame problem but suggests their framework could contribute if a homomorphism exists between the qualification and frame problems.
- Algebraic and Logical Properties: The paper lays the groundwork for a theory of referential models based on preferences, but the algebraic and logical properties of these models are noted as areas for further investigation.
Summary updated