TalTech Research Stay

A scholarship awarded by the School of Information Technology at Tallinn University of Technology (TalTech) enabled me to conduct a research stay in the context of Prof. Anu Masso’s Social Data Migration project. The work focused on cross-border data flow and on making extensive scientific literature systematically analyzable in a research area that had remained comparatively underexplored.

In cooperation with social scientists, I developed an LLM-supported NLP pipeline for this purpose. It automatically captured relevant scientific literature, aggregated metadata and available text data, normalized the documents and prepared them for thematic exploration.

The scraped texts were first split into analyzable chunks, which were checked using metadata, text length and noise indicators. Very short, redundant or technically insufficient segments were flagged before they could distort the modeling step. The pipeline then generated embeddings, cached them reproducibly with NumPy and used BERTopic, c-TF-IDF, n-grams, keyphrase candidates, UMAP and HDBSCAN for unsupervised topic structuring.

To make the topics methodologically more robust than a single plausible run, I executed repeated model runs with varying seeds and parameters. The pipeline compared topic-quality metrics, reliability scores and multi-run results in an evaluation dashboard. This made it possible to identify which topics remained stable and which clusters reacted sensitively to parameter changes.

The goal was to enable literature-based topic exploration for the emerging field of Social Data Migration. The analysis aimed to reveal recurring patterns, concepts and theoretical reference points and to prepare an analogy to social regularities. The results were prepared for publication at the end of the project.