Andrea Volpini, WordLift, CEO
Beatrice Gamba, WordLift, Head of Innovation

In this talk, I will present a hybrid approach that merges ontology-based reinforcement learning with a content assessment algorithm to create AI systems capable of self-refinement and tailored content generation. Drawing inspiration from models like DeepSeek R1 and guided by frameworks such as SEOntology, our methodology encourages language models to develop a structured chain-of-thought aligned with established SEO practices and audience needs.

By integrating a formal ontology and knowledge graph into the training pipeline, we constrain the model’s reasoning process with clear, semantic rules—ensuring every decision is both consistent and explainable. This symbolic guidance, combined with iterative trial-and-error learning, drives the model to that breakthrough “aha moment” where it spontaneously re-evaluates and corrects its path. Moreover, the content assessment algorithm provides objective, continuous evaluation of output quality, ensuring that the learned representations capture the cognitive core essential for effective, audience-specific content creation.

I will share the theoretical foundations, the design of our training pipeline, and experimental results that demonstrate how this neuro-symbolic approach can improve enterprise AI applications—from optimizing marketing strategies to automating digital content creation.