Pages 119–133

Authors

Suphitcha Thawornlimpaphong and Nasi Tantitharanukul

Abstract

This research develops and evaluates research classification tools for mapping research publications to the Sustainable Development Goals (SDGs) using SDG data collection and machine learning-based text classification. The proposed approach integrates Scopus textual metadata, SciVal SDG reference labels, an SDG key phrase repository, rule-based classification, and sdgBERT-based classification. Scopus was used as the source of textual metadata, including titles, abstracts, author keywords, and index keywords, while SciVal was used as the reference source for SDG labels. The Scopus dataset contained 55,702 records. After duplicate removal, 51,636 unique records remained, of which 31,660 records were successfully matched with SciVal, representing a match rate of 61.31%. The rule-based method classified 8,770 matched records, corresponding to 27.70% coverage. After excluding empty-empty cases, rule-based classification achieved 47.08% adjusted agreement and 38.22% exact match with SciVal. In contrast, sdgBERT at a threshold of 0.94 achieved 99.56% coverage, but its Micro F1, Macro F1, and adjusted agreement were 0.2061, 0.2411, and 11.71%, respectively. The results show a clear trade-off between the two approaches. Rule-based classification provides interpretable and more conservative outputs when explicit keyword evidence exists, but its coverage is limited. sdgBERT provides broader semantic coverage but shows low agreement with SciVal reference labels. Therefore, the developed tools should be used as preliminary decision-support mechanisms for SDG research classification rather than as fully automated final classification systems. Future work should develop and empirically test a hybrid approach that combines rule-based evidence, sdgBERT candidate labels, threshold-based configuration, and expert validation.