Towards Non-Monotonic Entailment in Propositional Defeasible Standpoint Logic

Technical Summary: Towards Non-Monotonic Entailment in Propositional Defeasible Standpoint Logic

Problem Statement
Recent advancements in defeasible reasoning have successfully applied preferential semantics and KLM-style entailment to modal logics, specifically Propositional Defeasible Standpoint Logic (PDSL). PDSL allows for the representation of conflicting viewpoints and typical beliefs within a single framework using modalities like $❘∼∼❘_e \alpha$ (meaning "$\alpha$ holds in the most typical understandings of standpoint $e$"). However, existing work in this domain has primarily focused on satisfiability checking and Tarskian (monotonic) entailment ($\models_P$).

The authors identify a critical inferential weakness in the current monotonic approach. Under $\models_P$, a knowledge base often fails to support intuitive non-monotonic inferences. For instance, if a standpoint $P$ typically believes $\alpha$, and $\alpha$ defeasibly implies $\beta$, the monotonic framework cannot conclude that $P$ typically believes $\beta$. Similarly, sub-standpoints (e.g., Engineering) cannot inherit typical beliefs from a super-standpoint (e.g., Physics) if there is any potential for conflict, even if no specific conflict exists. This limits the utility of PDSL for prototypical reasoning where inheritance and typicality are expected.

Methodology
To address these limitations, the paper proposes lifting non-monotonic rational entailment relations from traditional propositional KLM reasoning to a fragment of PDSL. The methodology proceeds through four main stages:

1. Extension of Expressivity via Situated Conditionals:
   The authors introduce "situated standpoint conditionals" of the form $\alpha \rightsquigarrow_e \beta$, read as "within standpoint $e$, a belief in $\alpha$ usually implies a belief in $\beta$." This construct allows defeasible implications to be situated specifically within a standpoint's context, rather than globally. The authors demonstrate that this new syntax is expressive enough to re-characterize the entire PDSL language, effectively replacing standard defeasible modalities and implications with these situated conditionals and a universal standpoint modality ($\square_*$).

2. Translation to Propositional KLM:
   The paper defines a specific fragment of PDSL, denoted as $Cond(\tilde{\mathcal{L}}_S)$, consisting of positive defeasible conditionals ($\alpha \rightsquigarrow_e \beta$) and standpoint sharpenings ($e \lesssim d$). A translation function $T$ is defined to map formulas in this fragment to traditional propositional KLM conditional knowledge bases.

   • Standpoint expressions are translated into propositional atoms or Boolean combinations thereof.
   • A conditional $\alpha \rightsquigarrow_e \beta$ is translated to $(\alpha \land t(e)) \mid\sim \beta$.
   • A standpoint sharpening $e \lesssim d$ is translated to $t(e) \mid\sim t(d)$.
     Crucially, the authors prove that this translation is faithful: a PDSL knowledge base is satisfiable if and only if its propositional translation is satisfiable, and model satisfaction is preserved under the translation ($M \models \varphi \iff T(M) \models T(\varphi)$).

3. Lifting Entailment Relations:
   Using the translation $T$, the authors lift rational defeasible entailment relations from the propositional case to the PDSL case. For a given ranking strategy $r$ (which assigns a ranking function to a knowledge base), the entailment relation $| \approx_r$ in PDSL is defined such that $\mathcal{K} | \approx_r \varphi$ if and only if the preimage of the propositional ranked model $r(T(\mathcal{K}))$ satisfies $\varphi$.
   This approach allows the use of existing propositional algorithms (such as DefeasibleEntail) to perform entailment checking in PDSL.

4. Specific Instantiations:
   The framework is applied to two well-known rational entailment systems:

   • Rational Closure: Defined via the minimal ranking function. The paper shows that Rational Closure in PDSL avoids the inferential weaknesses of monotonic entailment, allowing for the inheritance of typical beliefs in the absence of direct conflict.
   • Lexicographic Closure: Introduced to solve the "drowning problem" (where an exception in one attribute blocks inheritance of unrelated typical beliefs). The authors demonstrate that Lexicographic Closure in PDSL successfully distinguishes between unrelated beliefs, allowing a sub-standpoint to inherit beliefs that are not involved in the specific exception.

Key Results

• Expressivity: The introduction of situated conditionals allows PDSL to express its full syntax using a structure analogous to propositional conditionals, facilitating the translation to KLM logic.
• Inferential Strength: The proposed non-monotonic entailment ($| \approx_r$) is strictly stronger than the previous monotonic entailment ($\models_P$). It successfully supports inheritance of typical beliefs for sub-standpoints and the derivation of consequences from typical beliefs, which $\models_P$ failed to do.
• Complexity Preservation: The complexity of entailment checking in the proposed PDSL fragment is identical to that of the propositional case.
  • Checking Rational Closure is $P^{NP}_{||}$-complete.
  • Checking Lexicographic Closure is $P^{NP}$-complete.
  • This is achieved because the translation $T$ is linear, and the underlying propositional algorithms can be used directly.
• Handling of Diamond Modalities: The paper introduces a "negation as failure" operator (not) to handle diamond modalities ($\Diamond_e \alpha$) within the non-monotonic framework. This allows for "brave reasoning" regarding possibility: if a possibility is not precluded by the knowledge base, it is considered entailed.

Significance and Claims
The paper claims to provide a systematic method for integrating non-monotonic reasoning into standpoint logics, addressing the "inferentially weak" nature of previous monotonic approaches. By establishing a faithful translation between a fragment of PDSL and propositional KLM logic, the authors enable the direct application of established rational entailment systems (Rational and Lexicographic Closure) to multi-perspective reasoning.

The significance lies in:

1. Resolving Inferential Weakness: Enabling PDSL to perform prototypical reasoning where sub-standpoints inherit typical beliefs and defeasible implications propagate within specific viewpoints.
2. Solving the Drowning Problem: Demonstrating that Lexicographic Closure can be applied to PDSL to prevent unrelated exceptions from blocking the inheritance of typical beliefs.
3. Computational Efficiency: Proving that these advanced non-monotonic inferences can be computed using existing propositional algorithms without increasing the computational complexity class.

The authors conclude that their work bridges the gap between the expressivity of standpoint logics and the robust inferential capabilities of KLM-style defeasible reasoning, offering a practical framework for reasoning with conflicting and typical viewpoints.