Current Topics in NLP and society (Master S3 Class)

Basic information

• Credits: 3
• Format: 2 hours weekly
• 2026-2027 instructors: Marie Candito
• 2026-2027 schedule: Thursdays, 11:15-13:15, room 309
• Moodle page

Description

Machine learning and NLP systems have penetrated society by making decisions impacting our choices in everyday life: choices of what to buy, which movie to see, and more recently with generative models, choices of what to write or how to understand a text.
The course will mainly focus on biases in NLP systems and language models, as potential sources of unfairness towards groups of people. Biases occur when systems behave differently depending on the demographic characteristics of people, whether mentioned in input to the systems and/or provided as output. They propagate from societal structures to training corpora to pretrained models, and back to society when these models are used in downstream tasks.
We will study current research on (i) bias measures in training corpora, models’ internal representations and models’ output, (ii) measures of bias amplification between training corpora and trained models, (iii) bias mitigation techniques and (iv) strategies to improve transparency of models.

Prerequisites

Familiarity with NLP and machine learning basics. Familiarity with pretrained language models.
Programming skills in python and possibly in PyTorch

Learning outcomes

On successful completion of this course, students should be able to:

• understand and implement current strategies for measuring and mitigating biases in NLP
• read and critically review research articles on these topics