I2SC Lecture Series

At the Intersection of NLP and Survey Methodology: Potentials, Challenges, and Provocations

Recent advances in generative AI, particularly Large Language Models (LLMs), have renewed interest in Natural Language Processing (NLP) across many disciplines. One such field is survey methodology, which increasingly applies NLP methods (e.g., simulating survey respondents with LLMs) and, in turn, offers valuable insights back to NLP (e.g., improving annotation task design and evaluation). In this talk, I argue for stronger synergies between NLP and survey methodology, outlining both the opportunities and the obstacles at their intersection. I will focus on two use cases where we see both synergies and tensions: (1) reusing survey questions developed for humans to probe political biases in LLMs, and (2) assessing which people LLMs can simulate and provide personalized assistance to. Beyond these use cases, I highlight broader challenges that interdisciplinary researchers and practitioners must grapple with — from evaluating the social competencies of LLMs to responsibly using them for social measurement.

Society-centered AI: An Integrative Perspective on Algorithmic Fairness

In this talk, I will share my never-ending learning journey on algorithmic fairness. I will give an overview of fairness in algorithmic decision-making, reviewing the progress and wrong assumptions made along the way, which have led to new and fascinating research questions. Most of these questions remain open to this day and become even more challenging in the era of generative AI. Thus, this talk will provide only a few answers but many open challenges to motivate the need for a paradigm shift from owner-centered to society-centered AI. With society-centered AI, I aim to bring the values, goals, and needs of all relevant stakeholders into AI development as first-class citizens to ensure that these new technologies are at the service of society.

Machine Learning and Linked Open Data for Literary History

This talk presents an innovative approach to literary history that integrates machine learning (ML) with linked open data (LOD), demonstrating how computational methods can expand the scope of traditional humanities research. The approach was developed in the Mining and Modeling Text project at the Trier Center for Digital Humanities and leverages three complementary data sources: bibliographic metadata from the Bibliographie du genre romanesque français, a corpus of 200 XML-TEI encoded French novels (1750–1800), and secondary scholarship on the Enlightenment novel. Using ML techniques such as topic modeling and named entity recognition, we extract structured information and represent it as LOD triples in a public Wikibase instance. The resulting semantic knowledge graph enables complex querying across heterogeneous data, uncovering large-scale patterns and trends in 18th-century French literature. Beyond the case study, the talk highlights the broader potential of combining ML pipelines with LOD infrastructures to build open, federated, and multilingual knowledge resources for Computational Literary Studies and Digital Humanities more generally.

From Humans to Machines and Back: Fairness, Causality, and the Role of Social Science

Fairness in machine learning is not only a technical property of algorithms but a deeply social question: who benefits, who is harmed, and whose perspectives are embedded in our systems. While computational methods can surface biases and propose adjustments, addressing fairness requires more than optimization. It demands an understanding of social structures, power dynamics, and human behavior—areas where social science offers critical background knowledge, especially when applying causal approaches that aim to explain and intervene.

As we move into the era of large language models (LLMs) and increasingly agentic AI systems, these challenges grow more complex. Fairness concerns now span not just data and predictions, but dialogue, reasoning, and decision-making processes in systems that simulate human-like behavior. Traditional methods fall short, and novel approaches—bridging mechanistic insights from computer science with behavioral perspectives from the social sciences—are urgently needed.

At the same time, the relationship is not one-way. Studying LLMs and agentic AI can illuminate how human-like reasoning emerges (or fails to emerge) in artificial systems, offering new tools and sparking fresh questions for the social sciences. In this way, fairness research becomes a two-way street: moving from human to ML and back, with each side informing the other.

January 23, 2026

Statistical innovation for population estimation: integrating administrative records, geospatial data, and real-time streams

Reliable subnational population estimates are essential for effective governance, service provision, and humanitarian response, yet many low- and middle-income countries lack recent or comprehensive census data. This talk explores emerging strategies for estimating population counts when traditional sources are unavailable. I will focus on the integration of diverse and incomplete data sources—including administrative records (e.g., health, education), satellite imagery, and geospatial covariates—combined through Bayesian modeling frameworks to generate high-resolution estimates. In the second part of the talk, I will extend these methods to the problem of nowcasting population distributions in crisis contexts, including forced displacement and conflict. Here, I will explore the utility of real-time or near real-time data streams—such as mobile phone metadata, satellite-derived indicators, and social media signals—for capturing rapid demographic shifts at fine spatial and temporal scales. Throughout the talk, I will emphasize not only the technical and methodological advances enabled by the digital data but also the ethical considerations around data use, privacy, and representation.