Social, Legal, Ethical, Empathetic, and Cultural Rules: Compilation and Reasoning

Nicolas Troquard, Martina De Sanctis, Paola Inverardi, Patrizio Pelliccione, Gian Luca ScocciaAAAI 2024 · DOI

TL;DR

This paper presents a systematic approach to translate natural language "SLEEC" rules (Social, Legal, Ethical, Empathetic, Cultural) into classical logic. This translation enables automated reasoning for AI systems, ensuring they adhere to complex human-defined norms and behaviors.

What Problem Does This Paper Address?

As AI and autonomous systems become more integrated into our lives, ensuring their behavior aligns with human contexts and ethical principles is paramount. Townsend et al. (2022) proposed a methodology to elicit "SLEEC" rules, which are high-level normative principles expressed in natural language. However, for these rules to be effectively used by AI systems, they need to be translated into a formal, machine-interpretable language that supports automated reasoning. This paper addresses the challenge of bridging this gap by providing a method to compile these natural language SLEEC rules into classical logic, making them amenable to computational processing and verification.

Methods

The core contribution of this paper is a method for compiling SLEEC rules into classical propositional logic. The authors begin by analyzing the linguistic structure of SLEEC rules, particularly the "UNLESS... IN WHICH CASE..." construction. They propose a formal definition for this construction, translating it into logical implications.

The compilation process involves:

  1. Disambiguation: Identifying and assigning logical variables to atomic statements within the SLEEC rules.
  2. Logical Translation: Converting the natural language rules, including conditional statements and exceptions, into a structured logical form. This involves defining the semantics of the "UNLESS" connective and the "UNLESS... IN WHICH CASE..." construct.
  3. Formalization: Representing the compiled rules as conjunctions of clauses, suitable for input into standard logical reasoners.

The paper also explores the computational complexity of reasoning with these compiled rules, demonstrating that it is generally CONP-complete but can be PTIME-complete under specific restrictions. Finally, they show how these compiled rules can be implemented using logic programming frameworks like Answer Set Programming (ASP) and Prolog.

Datasets

Not discussed in the paper. The paper uses a running example from Townsend et al. (2022) involving a robot opening curtains based on user requests and conditions.

Key Insights and Ideas

  • Formalizing "Unless": The paper provides a precise logical interpretation of the "UNLESS" connective and the more complex "UNLESS... IN WHICH CASE..." construction, which is crucial for accurately capturing exceptions and conditions in normative rules. They argue against a simple XOR interpretation for "unless" and propose a definition equivalent to ¬ψ → φ for φ UNLESS ψ.
  • SLEEC Rule Pattern: They identify a general pattern for SLEEC rules (IF C0 THEN O0, UNLESS C1 IN WHICH CASE O1, ...) that allows for a systematic compilation into classical logic.
  • Computational Complexity: While reasoning with general SLEEC rules is computationally hard (CONP-complete), the paper shows that specific restrictions (e.g., negative conditions and positive obligations) can make it tractable (PTIME-complete). This highlights the importance of the structure of the rules.
  • Practical Implementation: The compiled logical forms are directly usable with off-the-shelf SAT solvers and logic programming frameworks (ASP, Prolog), demonstrating a practical path to integrating normative reasoning into AI systems.

Key Results

The paper successfully demonstrates a method to compile SLEEC rules into classical propositional logic. For the example rule concerning a robot opening curtains, the compiled form is (a^¬¬d → o)^(a^«d>«h → n^s)^(a^«d>h → o). This compilation allows for automated reasoning, such as determining obligations based on given states. For instance, if a user asks to open curtains, is dressed, and is highly distressed, the system correctly infers an obligation to open the curtains. The paper also shows that the compiled form of a SLEEC rule is quadratic in the size of the original rule.

Limitations and Future Work

  • Ambiguity in Natural Language: The authors acknowledge that natural language is inherently ambiguous, and the process of disambiguating statements and assigning logical variables requires careful engineering and domain expertise. The modeller must actively avoid or resolve such ambiguities.
  • Complexity of General Rules: While specific cases are tractable, the general case of reasoning with SLEEC rules remains computationally hard. This suggests that for very complex rule sets, efficient approximation or specialized reasoning techniques might be necessary.
  • Future Work: The paper suggests that the compiled logical forms can be seamlessly transferred to other components of a robot platform, such as planning modules. Further exploration of how these compiled rules can be integrated into more sophisticated AI architectures and decision-making processes would be a natural next step. The authors also mention the potential for using complex propositional formulas for conditions (Ci) and obligations (Oi) without impacting computational complexity, which could be further explored.

Summary updated

BibTeX
@inproceedings{troquard2024social,
  author = {Nicolas Troquard and Martina De Sanctis and Paola Inverardi and Patrizio Pelliccione and Gian Luca Scoccia},
  title = {{Social, Legal, Ethical, Empathetic, and Cultural Rules: Compilation and Reasoning}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2024},
  doi = {10.1609/aaai.v38i20.30245},
}
Figure