Concept2Text: an explainable multilingual rewriting of concepts into natural language

Automated and explainable data interpretation hinges on two critical steps: (i) identifying emerging properties from data and representing them into abstract concepts, and (ii) translating such concepts into natural language. While Large Language Models have recently demonstrated impressive capabilities in generating natural language, their trustworthiness remains difficult to ascertain. The deployment of an explainable pipeline enables its application in high-risk activities, such as decision making. Addressing this demanding requirement is facilitated by the fertile ground of knowledge representation and automated reasoning research. Building upon previous work that explored the first step, we focus on the second step, named Concept2Text. The design of an explainable translation naturally lends itself to a logic-based model, once again highlighting the contribution of declarative programming to achieving explainability in AI. This paper explores a Prolog/CLP-based rewriting system designed to interpret concepts expressed in terms of classes and relations derived from a generic ontology, generating text in natural language. Its key features encompass hierarchical tree rewritings, modular multilingual generation, support for equivalent variants across semantic, grammar, and lexical levels, and a transparent rule-based system. We present the architecture and illustrate a simple working example that allows the generation of hundreds of different and equivalent rewritings relative to the input concept.