Less Is Moral: A CHARMing Framework for Moral Foundations Detection in Endorsement Behaviour

A new framework, CHARM, has been developed to improve the detection of moral foundations in online endorsement behavior. This framework aims to overcome limitations of existing systems, such as poor cross-domain generalization and reliance on costly Large Language Models (LLMs), by integrating rationale alignment and hate-speech awareness into a lightweight fine-tuned LLM.

Why it matters

The ability to accurately detect moral foundations in online communications is critical for understanding public sentiment, managing brand reputation, and mitigating the spread of harmful narratives. This framework offers a more robust and efficient method for such analysis, potentially influencing strategic communication and risk management approaches in the digital sphere.

Key insights

• Existing moral foundation detection systems exhibit poor cross-domain generalization.
• Current systems often lack robust rationale grounding for their predictions.
• Many existing moral detection methods are reliant on expensive prompting-based LLMs.
• The CHARM framework introduces a 'MA-C- and Hate-speech-Aware Rationale-aligned Moral foundation detection' system.
• CHARM is built upon a lightweight fine-tuned LLM.
• The framework integrates complementary moral grounding, rationale alignment, and polarity-aware hate speech signals.
• CHARM aims to provide more robust and faithful moral prediction compared to prior dictionary-, fine-tune-, or prompt-based detectors.

Source

arXiv — Computers and Society — https://arxiv.org/abs/2609.03330