This research project addresses the challenge of structuring moral reasoning in large language models (LLMs) by integrating robust and explicit moral knowledge representations. While LLMs have demonstrated significant success across natural language processing tasks, concerns persist regarding their ability to capture, represent, and reason about moral values in a consistent and controllable manner. Current LLMs, predominately trained on vast and indiscriminate textual resources, tend to absorb and propagate biases and undesirable moral behaviors, raising both scientific and societal risks—especially as these models are increasingly deployed in decision-making contexts.

The proposed project, MORA, aims to enhance the moral capabilities of LLMs through the development and insertion of structured moral knowledge using knowledge graphs. Leveraging the Moral Foundations Theory as a principled framework, the research will focus on generating and expanding knowledge graphs that structurally encode core moral dimensions. The project will then design advanced knowledge injection techniques—such as retrieval-augmented generation and sophisticated prompting methods—to transfer moral knowledge from these graphs into LLMs, ensuring reliability and the mitigation of erratic moral outputs.

To rigorously assess the impact, the project will also deliver a standardized, extensible moral reasoning benchmark and supporting infrastructure to evaluate the moral inference abilities of LLMs across tasks, domains, and languages. The expected outcomes include open-access knowledge graph resources, methodological advances for moral knowledge integration in LLMs, and public benchmark datasets and evaluation tools—contributing both to the scientific field and to regulatory objectives for safe and trustworthy artificial intelligence.