A Topic Tone Framework for Tracking Linguistic Framing in Climate Science Abstracts: Evidence from arXiv (1995–2026)

Abstract

Climate change has become central to global policy and public debate, but it's less clear whether this shift shows up in the more guarded language of scientific writing itself. Because standard sentiment analysis tools don't work well on formal academic prose, we instead build a topic tone framework that pairs Latent Dirichlet Allocation with custom lexicon-based scores for uncertainty, certainty, and risk intensity capturing how confidently and how urgently claims are made, rather than emotional tone. Using climate-related abstracts from the arXiv metadata corpus (1995–2026), we model topics, score tone, and run regression and change point analyses to track trends over time. We finds a gradual but statistically meaningful rise in risk-oriented language, apparently accelerating after 2015, while uncertainty fluctuates rather than steadily declining possibly reflecting the growing complexity of climate modeling. Topics also carry distinct linguistic signatures: policy related work shows more risk language, while modeling heavy work hedges more. None of this points to scientific writing becoming "alarmist," but it does suggest a real, incremental shift in how climate risk is discussed one this reproducible framework could help track in other fields too.