Interactive Explainer for Overlapping Labels in Multi-Label SDG Classification

Abstract

Multi-label tagging of research and policy documents against the 17 UN Sustainable Development Goals (SDGs)remains an open problem: even modern Large Language Models(LLMs) paired with knowledge graph-based post-correction leave a substantial residual error. We argue that part of this gap may reflect not a model deficiency but an artefact of how the 17 SDGs were constructed, through political negotiation among 193 UN member states rather than along empirical category boundaries. We present an interactive document explainer that surfaces the resulting class entanglement through three heterogeneous lenses: lexical overlap in the document text, label-correlation structure across the corpus, and label co-occurrence among nearest research neighbours. Pinning any SDG displays its entanglement
profile across all three lenses simultaneously, enabling a per-document audit of the classes the classifier could not separate.We illustrate the explainer on representative papers from the Aurora corpus, where the three lenses converge on the same conclusion: the SDGs the classifier missed and those it found correctly cluster together across all three lenses.