KenyaESG: A Sentence-level ESG Disclosure Classification Dataset for Nairobi Securities Exchange Corporate Reports

Unofficial AI-generated summary based on the public title and abstract. Not an official translation.

🔬研究者:A valuable resource for NLP and ESG research, especially for understudied African markets; the dataset enables reproducibility and cross-market comparisons.

🏢実務担当者:Can be used to build automated ESG screening tools for African portfolios or to benchmark disclosure quality against global standards.

🏛政策担当者:Highlights the need for structured ESG data in emerging markets and may inform disclosure regulation design.

KenyaESG is a sentence-level dataset for three independent binary ESG classification tasks (Environmental, Social, Governance) built from the annual, integrated, and sustainability reports of firms listed on the Nairobi Securities Exchange (NSE), 2010–2024, combined with reference sentences from Schimanski et al. (2024). For each pillar, the dataset provides a fine-tuning set of 3,900 sentences and a disjoint held-out human evaluation set of 100 sentences carrying two independent annotators' judgments (Cohen's kappa: 0.94 Environmental, 0.86 Social, 0.88 Governance). Training and evaluation sentences are disjoint to prevent leakage; together they reconstruct the full 4,000-sentence pool per pillar. Training labels for the Kenya-sourced sentences were assigned by a keyword-based filter and refined by a single-reviewer pass (not full manual annotation); the human-annotated evaluation sets are the appropriate benchmark for performance estimates. The dataset captures disclosure intensity rather than disclosure quality. The corpus is predominantly English but retains a small number of Swahili and OCR-degraded sentences, reflecting the original NSE reports. Company names and personal names have been masked in the text field ([COMPANY] and [PERSON]) in line with the anonymisation principles of the Kenya Data Protection Act, 2019 (No. 24 of 2019) and Regulation (EU) 2016/679 (GDPR). Masking is semi-automatic and not guaranteed to be exhaustive; a small number of false-positive masks on common words (e.g. "equity", "equality", "Standard") were subsequently restored, as documented in the correction log. The dataset must not be used to attempt re-identification of individuals or firms. Files: three training sets (KenyaESG_train_environmental.xlsx, KenyaESG_train_social.xlsx, KenyaESG_train_governance.xlsx), three evaluation sets (KenyaESG_evaluation_environmental.xlsx, KenyaESG_evaluation_social.xlsx, KenyaESG_evaluation_governance.xlsx), the dataset card (KenyaESG_dataset_card.pdf), and the correction log (KenyaESG_correction_log.txt). The companion fine-tuned classifiers are available on the Hugging Face Hub: KenyaESG-RoBERTa-env (DOI 10.57967/hf/9126), -soc (10.57967/hf/9127), and -gov (10.57967/hf/9128). Licensed under Apache-2.0.

• openalex https://doi.org/10.5281/zenodo.20608237first seen 2026-06-14 04:40:56 · last seen 2026-06-14 04:41:29