Speaker: Annapaola Ginammi
28 Gennaio 2026 | 15:00
DEIB, Aula 2A (Ed. 20)
Contatti: Prof.ssa Viola Schiaffonati
Sommario
On January 28, 2026, at 3:00 pm the seminar on "Three paradoxes of gender bias in data-driven language models: Philosophical challenges for Natural Language Processing" will take place in DEIB 2A Room (Building 20).A big challenge in present Natural Language Processing (NLP) is preventing machine learning (ML) applications from replicating and exacerbating societal biases such as gender bias. NLP research usually addresses this issue from a purely technical perspective, but, as I will argue in this talk, bias mitigation in data-driven language models faces deep challenges of a philosophical nature. I will present these challenges in the form of three paradoxes.
The first paradox concerns the theoretical foundations of data-driven language models, namely, the idea that a word’s meaning is based on its use. Even models that accurately model word use are taken to be ‘biased’, so the bias must consist in word use itself. But if word meaning is taken to lie in word use, then it follows that bias is constituent of word meaning. However, NLP bias mitigation presupposes a gap between bias and word meaning. There is thus a conceptual tension between the theoretical foundations of data-driven language models on the one hand, and bias mitigation in those models on the other.
The second paradox concerns the normative component of bias mitigation in data-driven language models, and is an analogue of the well-known “bias-paradox” from feminist philosophy of science. Language models inevitably encode values, since values shape language itself, for example through grammar rules and classification practices. NLP bias mitigation presupposes that some of these values encoded in data-driven language models are “bad”, but eliminating values from these models altogether, to obtain a neutral, impartial model is impossible: we can at most aim to replace them by different values. This raises the question of how we can distinguish between “good” and “bad” values, and how we can justify such a distinction.
The third paradox concerns the representation of social notions such as gender in data-driven language models. These models are often said to be biased insofar as they encode stereotypical associations between gender categories and other notions, such as occupations. Yet epistemically, such associations are precisely what probabilistic models are designed to capture: statistically salient patterns in large bodies of language data. The paradox is notmerely that, for notions such as gender, epistemic success and normative success come apart. Rather, it is that bias mitigation in data-driven language models may change the degree of stereotyping, but not the kind of representation involved. Thus, mitigation strategies threaten to undermine both epistemic and normative success simultaneously: epistemic success is weakened because the model departs from the data, while normative success remains at most partial because the representational form itself remains stereotype-based.
Biografia
Annapaola Ginammi is a philosopher who earned her PhD in 2018 from the Scuola Normale Superiore with a thesis on infinite idealizations in physics.She held postdoctoral positions on the history of logic and computational methods for philosophy at the University of Amsterdam and on analogical reasoning in science at Politecnico di Milano. Her recent research focuses on the philosophical and ethical challenges of bias detection and mitigation in data-driven language models.