Semantic Rules from Text: Automating Common-Sense Knowledge Curation with LLMs

Common Sense knowledge (CSK) is critical in improving artificial intelligence (AI) systems by enhancing their under- standing, reasoning, and interaction with the human world, especially in planning and decision-making problems. To achieve the practical applicability of CSK, an extensive set of this latter should be available for the problem domain. Large Language Models (LLMs) have shown promise in quickly curating CSK for a particular domain. Modern LLMs engines, such as GPT, can trans- late CSK expressed in natural language to semantic rules using formal languages, e.g., Semantic Web Rule Language (SWRL) or Datalog rules. However, they fail to use standard vocabularies ISO-21838 while generating domain-specific semantic rules. This paper proposes a hybrid method that leverages LLMs to generate commonsense knowledge, which is then transformed into semantic rules based on standard vocabulary. By using these rules, we address the limitations of LLMs, ensuring the resulting ontologies are semantically rich and comply with established ontology engineering standards. This hybrid approach empowers LLMs to contribute to creating ontologies that are consistent, standardized, and semantically rich by adhering to the princi- ples of formal logics. Finally, this paper proposes a template-based prompt-engineering technique and pre-defined mapping that leverages LLMs to generate commonsense knowledge and transform it into semantic rules based on standard vocabulary. By comparing the quality of the result with other automated approaches like ChatGPT3.5 and GPT4.0, we show that the proposed semantic rules based approach guarantees consistency and adherence to specific upper-level ontologies for expressing CSK as semantic rules