Analysis of Space Research through Neural Topic Modeling leveraging Large Language Models
Published in SPAICE 2026, 2026
Recommended citation: Ruben Catarino Belo, Marta Guimaraes, Claudia Soares, "Analysis of Space Research through Neural Topic Modeling leveraging Large Language Models." SPAICE 2026, 2026. https://openreview.net/forum?id=GMZrV2CgYb
Abstract
The rapid growth of space-related research, spanning disciplines from Earth observation to robotics and machine learning, presents challenges for efficiently navigating and contextualizing the literature. Traditional categorization approaches struggle to capture emerging themes and cross-disciplinary connections. In this work, we present a data-driven analysis of space science publications from the well-known preprint server arXiv, using neural topic modeling enhanced with large language models to uncover latent topics and track their evolution over time. Our methodology identifies major research themes, highlights the emergence of multimodal and self-supervised approaches, and captures methodological trends within high-prevalence topics. We further assess the robustness of topic discovery by comparing neural topic models, revealing both consistent and model-specific patterns. This analysis provides a resource for researchers and newcomers alike to navigate the rapidly expanding literature, supporting informed exploration and fostering interdisciplinary understanding.
BibTeX
@inproceedings{spaice2026GMZrV2CgYb,
author = "Belo, Ruben Catarino and Guimaraes, Marta and Soares, Claudia",
title = "Analysis of Space Research through Neural Topic Modeling leveraging Large Language Models",
booktitle = "SPAICE 2026",
year = "2026",
month = "oct",
day = "21",
url = "https://openreview.net/forum?id=GMZrV2CgYb",
abstract = "The rapid growth of space-related research, spanning disciplines from Earth observation to robotics and machine learning, presents challenges for efficiently navigating and contextualizing the literature. Traditional categorization approaches struggle to capture emerging themes and cross-disciplinary connections. In this work, we present a data-driven analysis of space science publications from the well-known preprint server arXiv, using neural topic modeling enhanced with large language models to uncover latent topics and track their evolution over time. Our methodology identifies major research themes, highlights the emergence of multimodal and self-supervised approaches, and captures methodological trends within high-prevalence topics. We further assess the robustness of topic discovery by comparing neural topic models, revealing both consistent and model-specific patterns. This analysis provides a resource for researchers and newcomers alike to navigate the rapidly expanding literature, supporting informed exploration and fostering interdisciplinary understanding.",
note = "Accepted for oral presentation"
}