Classifying Multi-Labels of SDGs on Scientific Articles Using Keyword Extraction with Semantic Similarities

Multi-label classification of scientific publications against the United Nations Sustainable Development Goals (SDGs) is an increasingly important bibliometric task, yet existing approaches force a trade-off between accuracy, computational cost, and interpretability. Conventional models that rely on static word embeddings under-perform on this task because of the heavy semantic overlap between goals, while large language models reach competitive accuracy only at the cost of billions of parameters and opaque decisions. This paper proposes a hybrid pipeline that combines Sentence-BERT keyword retrieval from an Elsevier SDG dictionary, a finetuned SciBERT multi-label classifier, and a per-class threshold calibration step that adapts the operating point of each goal to its support distribution. The pipeline is evaluated on 5,765 publications from Universitas Airlangga, with five-fold multilabel-stratified cross-validation. The proposed method reaches a macro-averaged F1 of 0.746 (precision of 0.748, recall of 0.751). Under matched conditions on the same corpus and folds, it outperforms a re-implemented TF-IDF and logisticregression baseline (macro-F1 of 0.631) by 11.5 absolute F1 points and a re-implemented Word2Vec and logistic-regression baseline (macro-F1 of 0.477) by 26.9 absolute F1 points. A published zero-shot large language model baseline (Zephyr 7B, 0.717 on a ZORA-derived corpus) is discussed as a reference point but is not replicated here due to the compute it requires. Because every prediction is paired with the matched keywords from the dictionary, the system also produces a human-readable justification for each assignment, which addresses the transparency gap of LLM-based classifiers and makes the pipeline usable in institutional reporting pipelines.