Research Software Classification

Research software is difficult to discover and compare because repositories often lack consistent metadata and scientific categories.

Context

My PhD research examines which repository and publication attributes support classification, how community taxonomies align with scientific knowledge graphs, and which machine-learning methods produce interpretable results.

My role

I work on the research questions, data preparation, model evaluation, semantic alignment analysis, and public writing around the project. Earlier work from my master’s thesis was implemented into SOMEF, an open-source framework for extracting software metadata.

Approach

• Link software repositories with publication and metadata signals.
• Compare textual attributes such as README files, abstracts, and category descriptions.
• Evaluate machine-learning and embedding methods for classification.
• Inspect taxonomies through semantic alignment with scientific knowledge graphs.
• Keep the methodology interpretable enough for research software discovery use cases.

Outcomes

An earlier implementation reached 92% F1 during training and 76% on the test set, compared with a 36% prior baseline. The method was integrated into SOMEF, an open-source framework for extracting software metadata.

Related pages

• Research overview
• Automated Research Software Classification paper
• Methodology thesis

Lessons

Research software classification is not only an NLP task. The quality of the result depends on metadata availability, taxonomy design, scientific context, and the ability to explain why a repository belongs to a category.