UKRAINE'S GREEN RECOVERY IN THE INTERNATIONAL MEDIA: THE POSSIBILITIES AND LIMITS OF FINBERT-BASED SENTIMENT ANALYSIS

Abstract

This paper examines the tone of international media coverage of Ukraine's green recovery and asks how far that measurement depends on choices made in constructing it. The corpus consists of 456 English-language headlines retrieved from Google News RSS across eight queries between November 2019 and August 2026, classified with FinBERT (ProsusAI/finbert). The overall tone is moderately positive: 35.31% of headlines are positive, 57.02% neutral and 7.68% negative, giving a sentiment index of +0.276 (95% CI [+0.222, +0.331]). Two dimensions of genuine variation are identified. Institutional publishers are three times more likely than media outlets to produce a positive headline (+0.581 against +0.245; Fisher exact p = 0.00062). Tone also differs sharply by retrieval frame, from +0.545 for reconstruction-and-energy coverage to exactly 0.000 for climate finance, where positive and negative headlines are in perfect balance (chi-square = 58.84; p < 0.0001). The apparent recovery of tone over the course of the war, by contrast, does not survive scrutiny: the aggregate rise from +0.016 in 2022 to +0.363 in 2025-2026 is driven by a shift in which queries return content in which period, and disappears when the retrieval frame is held constant (p = 0.121). Two further demonstrations point the same way. Removing the publisher name that Google News appends to every headline changes 14.69% of classifications and moves the index from +0.217 to +0.276; and classifying topic by keyword match rather than by retrieval query turns a significant effect into a non-significant one. Applying financial language models to reconstruction discourse therefore requires full reporting of preprocessing, explicit control for retrieval composition and validation against a hand-coded subsample.