Forum Numerica - Federica GRANESE (Inria): Neural Topic Models for Evolving Text Streams

In an era where information flows faster than ever, understanding large textual corpora requires tools capable not only of discovering salient topics, i.e., coherent groups of words representing themes in a text, but also of tracking how these themes emerge, persist, transform, or disappear over time. While online topic models provide a natural framework for this task, they remain far less explored than their offline counterparts.
In this context, two challenges are particularly central. First, choosing the number of topics, which is often fixed a priori and poorly suited to evolving corpora. Second, tracking topics across timesteps, ensuring that recurring themes at different moments are recognized as the same rather than mistakenly treated as new ones.
In this talk, I will present SB-SETM, a model specifically designed to address both challenges.

Federica Granese holds an Inria Starting Faculty Position (ISFP) at the Inria Center at Université Côte d'Azur, where she is a member of the MARIANNE team.
She received her B.Sc. and M.Sc. in Computer Science from Sapienza University of Rome in 2017 and 2019, respectively, and her Ph.D. from Institut Polytechnique de Paris (École Polytechnique) in 2023.
Her research lies at the intersection of machine learning and natural language processing, with a particular focus on topic models. She is also interested in trustworthy machine learning, notably in monitoring and characterizing emergent behaviors and manipulation patterns in AI systems.