AI Lund lunch seminar: Cheap(er) human-in-the-loop labelling strategy for better datasets?

Topic: Cheap(er) human-in-the-loop labelling strategy for better datasets?

When: 10 December 12.00 to 13.00 CET

Where: Online - link by registration

Speaker: Dylan Pashley, PhD Student at the Department of Political Science Lund University

Moderator: Bibi Imre-Millei, PhD Student at the Department of Political Science Lund University

Spoken language: English 

Abstract

In the social sciences, the creation of high-quality labeled datasets is a persistent challenge: categories are often nuanced, context-dependent, and vulnerable to both model-driven and annotator-driven biases. Automated labeling alone frequently fails to capture these subtleties, while fully manual annotation is prohibitively costly at scale. 

This work addresses that gap by proposing a cost-efficient human-in-the-loop labeling strategy that preserves essential human judgment, increases awareness of potential bias, and enables the construction of reliable, interpretable datasets suitable for downstream scientific analysis. 

The proposed approach integrates a phased pipeline that combines human expertise, semi-automated labeling, and machine learning techniques to iteratively enhance dataset quality and model performance. By striking a balance between automation and human input, this strategy offers a practical and efficient solution for generating high-quality labeled datasets, reducing labeling costs, and supporting iterative model refinement.

Speaker Biography

Dylan Pashley is a Doctoral Candidate in Political Science at Lund University whose research focuses on leveraging large language models to improve data curation and streamline analytic workflows. His work applies these methods to large text corpora to study semantic drift, political framing, and the contestation of climate change within national parliaments.

Registration 

Participation is free of charge. 
Sign up at ai.lu.se/2025-12-10/registration and we send you an access link to the zoom platform.