Enhancing Ethical Explanations of Large Language Models through Iterative Symbolic Refinement (2402.00745v1)

Abstract: An increasing amount of research in Natural Language Inference (NLI) focuses on the application and evaluation of LLMs and their reasoning capabilities. Despite their success, however, LLMs are still prone to factual errors and inconsistencies in their explanations, offering limited control and interpretability for inference in complex domains. In this paper, we focus on ethical NLI, investigating how hybrid neuro-symbolic techniques can enhance the logical validity and alignment of ethical explanations produced by LLMs. Specifically, we present an abductive-deductive framework named Logic-Explainer, which integrates LLMs with an external backward-chaining solver to refine step-wise natural language explanations and jointly verify their correctness, reduce incompleteness and minimise redundancy. An extensive empirical analysis demonstrates that Logic-Explainer can improve explanations generated via in-context learning methods and Chain-of-Thought (CoT) on challenging ethical NLI tasks, while, at the same time, producing formal proofs describing and supporting models' reasoning. As ethical NLI requires commonsense reasoning to identify underlying moral violations, our results suggest the effectiveness of neuro-symbolic methods for multi-step NLI more broadly, opening new opportunities to enhance the logical consistency, reliability, and alignment of LLMs.

Enhancing Ethical Explanations of LLMs through Iterative Symbolic Refinement

The paper presented by Quan et al. addresses a significant challenge in natural language inference (NLI) utilizing LLMs — specifically, the generation of logical, coherent, and ethically aligned explanations. Despite the promise shown by LLMs like BERT, RoBERTa, and PaLM in reasoning tasks, they suffer from limitations such as factual inaccuracies, hallucinations, and explanation inconsistencies. This limitation is pronounced in complex domains such as ethical NLI, where the reasoning process is heavily reliant on abstract moral principles and commonsense knowledge.

To mitigate these shortcomings, the authors propose Logic-Explainer, a neuro-symbolic framework integrating LLMs with an external backward-chaining solver. This approach aims to enhance the logical validity, completeness, and non-redundancy of ethical explanations in NLI tasks. The process involves generating explanations using in-context learning and Chain-of-Thought (CoT) prompting, which are then refined iteratively to ensure logical consistency and alignment with ethical principles.

The paper outlines a method that combines abductive and deductive reasoning within a structured framework. This hybrid methodology enhances the LLM's ability to generate stepwise natural language explanations that are logically coherent. The framework utilizes a backward-chaining solver for symbolic refinement, ensuring that the explanations not only verify correctness but also reduce redundancy and incompleteness.

Empirical analyses demonstrate the efficacy of Logic-Explainer in improving the accuracy of identifying underlying moral violations in ethical NLI tasks. Notably, the method outperforms standard in-context learning by 22% and CoT methods by 5% in accuracy. The framework achieves a significant improvement in the logical validity of explanations, raising it from 22.9% to 65.1% for easy tasks and 10.3% to 55.2% for challenging ones. Additionally, the iterative refinement drastically reduces explanation redundancy from 86.6% to 4.6% and 78.3% to 6.2%.

In advancing ethical explanations, Logic-Explainer addresses critical issues in LLM reasoning, enhancing the reliability and comprehensibility of machine-generated explanations of ethical statements. The framework holds profound theoretical implications for the field of AI, marking a step toward more robust and ethically aligned AI systems. Practically, it offers a method to extract and verify knowledge in high-stakes domains where ethical reasoning is paramount.

The release of the ExplainEthics dataset, developed through this framework, contributes a new resource for the research community to further investigate and advance in ethical NLI tasks. The integration of symbolic reasoning with neural methods in Logic-Explainer presents potential avenues for future work, including exploration into more abstract and nuanced ethical reasoning and its application in diverse cultural contexts.

As AI continues to integrate into various aspects of human decision-making, ensuring the alignment of machine reasoning with ethical standards is imperative. The proposed neuro-symbolic framework signifies a promising direction for achieving this alignment, suggesting a future where LLMs can engage more reliably and comprehensively with complex ethical dilemmas.