TL;DR
This paper introduces "argumentative inference," a novel way for AI systems to reason with contradictory information in knowledge bases. Instead of discarding conflicting data, it identifies arguments for and against a conclusion, allowing for nuanced reasoning even when faced with inconsistencies. This approach is particularly useful for building more robust and realistic AI systems that can handle real-world data.
What Problem Does This Paper Address?
Modern AI systems often rely on large knowledge bases to make decisions and draw conclusions. However, these knowledge bases can easily become inconsistent due to various reasons, such as combining information from multiple sources, or dealing with rules that have exceptions. Classical logic breaks down in the face of inconsistency, leading to the "principle of explosion," where anything can be deduced from a contradiction. This makes it impossible for AI systems to function reliably with inconsistent data.
Traditional approaches to handling inconsistency either try to restore consistency by revising the knowledge base (which can be complex and lead to loss of information) or employ methods that are either too conservative (discarding too much information) or too permissive (leading to contradictory conclusions). This paper addresses this challenge by proposing a method that "copes" with inconsistency, allowing for meaningful inferences even when contradictions are present.
Methods
The core of the paper's contribution is the argumentative inference relation, denoted by $\Sigma \vdash_A \phi$. This relation defines a formula $\phi$ as a consequence of a knowledge base $\Sigma$ if:
- There exists a consistent argument for $\phi$ within $\Sigma$.
- There is no consistent argument for the negation of $\phi$ within $\Sigma$.
An "argument" is defined as a consistent subset of the knowledge base that entails a formula. This approach is inspired by Toulmin's model of argumentation and is related to defeasible reasoning.
The paper then extends this concept to:
- Flat knowledge bases: Where all formulas have the same reliability.
- Prioritized knowledge bases: Where formulas have different levels of certainty or priority. This is modeled using a layered structure ($\Sigma = B_1 \cup \dots \cup B_n$), where $B_i$ represents formulas at a certain priority level.
- Paraconsistent-like treatment: For prioritized knowledge bases, a formula can carry two weights: its degree of certainty and the degree of certainty of its negation. This allows for a more nuanced understanding of conflict.
The paper also compares its proposed method with existing inconsistency-tolerant consequence relations, such as:
- Free-consequence: Based on removing elements involved in any inconsistency.
- MC-consequence (Maximal Consistent Subset): Based on inferring from all maximal consistent subsets.
- Lex-consequence: Based on a lexicographical ordering of maximal consistent subsets.
- Existential-consequence: Based on inferring from at least one maximal consistent subset.
A comparative hierarchy of these inference relations is presented, showing how argumentative inference fits within this landscape.
Datasets
Not discussed in the paper. The paper focuses on theoretical frameworks and uses illustrative examples rather than empirical evaluations on specific datasets.
Key Insights and Ideas
- Argumentation as a core mechanism: The paper highlights that reasoning with inconsistency can be effectively modeled by focusing on the presence of supporting arguments and the absence of counter-arguments, rather than solely on logical entailment.
- "Ecological" approach to inconsistency: The argumentative inference is described as "ecological" because it tries to be faithful to the actual content of the knowledge base, avoiding the wholesale discarding of information that can happen in other revision-based approaches.
- Distinguishing "local" vs. "global" inconsistency: The paraconsistent-like treatment allows for reasoning about specific propositions that are involved in conflicts, rather than treating the entire knowledge base as globally inconsistent. This leads to more informative responses.
- Argument weights for nuanced reasoning: In prioritized knowledge bases, attaching weights to arguments allows for distinguishing between conclusions derived from "free" propositions (not involved in contradictions) and those derived from propositions that have both supporting and conflicting arguments.
- Avoiding the "drowning effect": The proposed methods, particularly in layered knowledge bases, aim to salvage information that might be lost in simpler possibilistic logic approaches due to conflicts.
Key Results
The paper presents several propositions that establish relationships between the argumentative inference and other consequence relations. For instance:
- If a knowledge base is consistent, argumentative inference coincides with classical inference.
- Argumentative inference is non-monotonic, meaning adding new information can invalidate previous conclusions.
- Argumentative inference is not closed under conjunction, meaning if $\phi$ and $\psi$ are argumentative consequences, $\phi \land \psi$ is not necessarily one. This is seen as a feature, allowing for reasoning about mutually consistent properties derived from conflicting reasons.
- The paper demonstrates through examples how argumentative inference can handle inconsistencies and provide more informative results than some other methods, particularly in prioritized knowledge bases.
Limitations and Future Work
- Computational Complexity: While the paper mentions that argumentative inference is amenable to "efficient standard implementation methods based on classical resolution," the complexity of finding all arguments and checking for their absence can still be significant, especially for large knowledge bases.
- "Ecological" but potentially globally inconsistent: The paper acknowledges that while the argumentative inference avoids outright contradictory responses (like $\phi$ and $\neg\phi$), the set of deduced sentences can still be globally inconsistent. The authors argue this is debatable as the supporting arguments for these conflicting conclusions might be distinct, representing different "points of view."
- Future Work:
- Positioning the paraconsistent inference with respect to other inference modes to assess its benefits.
- Applying the results to default reasoning and comparing it with existing frameworks.
- Further exploration of combining knowledge bases from multiple sources using the proposed argumentative framework.
- Developing more efficient implementation methods.
Summary updated