TL;DR
This paper introduces a novel approach to handling inconsistencies in knowledge bases by refining the definition of logical arguments. By "compiling" arguments to remove irrelevant information, they create a more robust and productive logic that avoids missing intuitive conclusions and satisfies key rationality postulates.
What Problem Does This Paper Address?
In artificial intelligence, particularly in areas like knowledge representation and reasoning, dealing with inconsistent information is a significant challenge. When a knowledge base (KB) contains contradictory statements, standard logical systems can break down, leading to nonsensical conclusions or an inability to derive any useful information.
Existing approaches to handle this inconsistency fall into two main categories: coherence-based logics and argument-based logics. While these methods aim to extract meaningful conclusions from inconsistent KBs, they often suffer from two main drawbacks:
- Missing Intuitive Consequences: They can fail to infer conclusions that are logically sound and intuitively expected, especially when the inconsistency doesn't directly affect the desired conclusion.
- Violation of Rationality Postulates: Some of these logics, particularly those that are more productive, can produce inconsistent sets of conclusions or fail to be closed under deduction (meaning if a set of formulas is inferred, all their logical consequences should also be inferable).
The authors argue that these limitations stem from the very definition of a "logical argument" used in these systems, specifically how arguments are constructed and what constitutes their "support" (the premises used to derive a conclusion).
Methods
The core contribution of this paper is a new method for defining and processing logical arguments, which they call compilation. This process aims to address the identified limitations by:
- Generalizing Argument Definition: They start with a more general notion of an "unrestricted argument" that relaxes the strict consistency requirement for its support. This allows for more potential arguments to be formed, addressing the issue of missing justifications.
- Compiling Arguments: The key innovation is the compilation process. For each unrestricted argument, they apply a "forgetting" operator (based on prior work by Lang et al.) to its support. This operator systematically removes variables that are irrelevant for deriving the argument's conclusion. This process can result in zero, one, or multiple new, more concise arguments.
- Decomposition and Conciseness: Compilation can decompose complex arguments into simpler ones and ensures that the remaining support is concise, meaning it contains only necessary information. This also helps restore consistency to the argument's support.
- Abstract Argumentation Framework: These compiled arguments are then used within Dung's abstract argumentation framework. The framework defines an attack relation between arguments (based on "direct defeat") and uses stable semantics to identify extensions (sets of non-conflicting arguments).
- Novel Consequence Relation: The paper defines a new non-monotonic consequence relation based on the stable extensions. A formula is considered a consequence if it is supported by at least one argument in every stable extension.
Datasets
Not discussed in the paper. The paper focuses on theoretical aspects and uses example knowledge bases for illustration rather than standard datasets.
Key Insights and Ideas
- The definition of an argument matters: The paper convincingly argues that the way logical arguments are constructed (specifically, the consistency and minimality requirements on their supports) is a root cause of limitations in existing inconsistency-handling logics.
- Compilation as a unifying solution: The proposed "compilation" technique acts as a powerful pre-processing step. It addresses multiple issues simultaneously: it allows for more arguments to be formed (by relaxing initial consistency), makes arguments more concise, decomposes complex ones, and ultimately restores desirable properties like consistency and minimality to the argument supports.
- Restoring rationality: By compiling arguments and using them within Dung's framework with stable semantics, the authors demonstrate that their new logic satisfies crucial rationality postulates (consistency and closure under deduction) that were violated by some prior approaches.
- More productive, yet safe: The new logic is shown to be more productive than some existing methods (inferring more conclusions) while remaining cautious enough to avoid inconsistencies, unlike some other productive logics.
Key Results
- The paper formally defines a compilation process for logical arguments that removes irrelevant variables from their supports.
- It proves that compiled arguments are restricted arguments (consistent, valid, and relevant supports) built from a derived knowledge base (Σ*).
- The proposed argumentation system, built using compiled arguments and stable semantics, is shown to have stable extensions and to satisfy key rationality postulates:
- The set of inferred consequences is consistent.
- The set of inferred consequences is closed under deduction.
- The new logic is demonstrated to be more productive than certain existing argument-based and coherence-based logics, inferring intuitive conclusions that others miss.
Limitations and Future Work
- Computational Complexity: While the paper defines the compilation process and the new logic, it acknowledges that the computational complexity of computing compiled arguments and the resulting inferences is a significant area for future work. They mention that complexity of forgetting and stable semantics have been studied, but efficient algorithms for their specific framework are needed.
- Algorithm Development: The authors explicitly state their future work will focus on defining efficient algorithms for computing compiled arguments and the inferences drawn by their new logic.
- Practical Applications: The paper is theoretical; future work would likely involve exploring how this new logic can be applied to real-world AI problems involving inconsistent data.
Summary updated