Active Learning with Large Language Models for Cross-lingual Pseudo-labeling in Computational Social Science

Large Language Models (LLMs) have become ubiquitous in the area of Natural Language Processing (NLP). This project aims to advance the applications of LLMs in the multidisciplinary field of Computational Social Science (CSS). Specifically, LLMs are often used to pseud-label (at scale) variables in data that are of interest to social scientists (e.g., a psychologist would look for emotion markers, an economist for statements about money, a human rights researcher for toxic language etc.) I will develop and test novel tooling for applying LLMs to efficiently pseudo-label CSS data in a cross-lingual setting. The project will proceed in three phases. First, a dataset and methodology for evaluating the language robustness of LLMs (isolated from all other confounding factors) will be developed. Second, using this methodology a set of general best practices will be determined for the scenario of applying LLMs to cross-lingual pseudo-labeling for CSS. Finally, both existing and novel active learning approaches will be investigated to minimize the manual annotation effort required to oversee the pseudo-labeling process. Using insights from this project a user-friendly freely publicly available tool for pseudo-labeling using LLMs will be published. The methods will be tested on two example tasks: toxicity detection and disinformation detection. Overall, the project results will make the work of CSS researchers more efficient in terms of both time and financial resources. Moreover, the cross-lingual nature of the tool will make it applicable to languages other than English, including small and low-resourced languages. Consequently, this project will contribute to make tooling accessible to a wide and diverse set of CSS researchers, increasing the outreach and inclusivity of CSS research and fostering international collaboration.