Mitchell Centre Seminar Series

About

Telmo Menezes CAMS (Centre d'analyse et de mathématique sociales) Interpretable by Design: Semantic Hypergraphs for Human-Centered Text Analysis Semantic Hypergraphs (SH) offer a knowledge representation formalism that naturally extends the idea of the Knowledge Graph with hyperedges, recursivity and a minimalist type system, becoming expressive enough to capture the full richness of natural language, while lending it structure and regularity. This in turn makes it possible to define simple but powerful extraction patterns, and also to appraise text corpora in a non-linear, interconnected way, with methods close to the realm of Network Science. In our lab we have been developing Hyperquest, a tool with a visual interface that is already fully functional, although not yet publicly released. I will use it throughout the talk to illustrate the concepts of SH, as well as to demonstrate the exploratory possibilities offered by this approach on various text corpora. A few examples of SH application will be given, such as the socio-semantic analysis of the Environmental Kuznets Curve (EKC) scientific field, and the mapping of worldviews in Reddit communities, combining user comments and the contents of shared news articles. In the context of the rapid rise of generative AI, I will discuss our goal of using computational tools to amplify, instead of replace, human intelligence. I will address questions of interpretability and groundedness, and describe neurosymbolic approaches that combine the strengths of symbolic representations such as SH with those of LLMs and related methods.