Towards Explaining STEM Document Classification using Mathematical Entity Linking

Mathematical content is underutilized in AI explanations: Current explainable AI methods focus primarily on text analysis, but STEM documents contain both text and mathematical formulas. This research demonstrates that mathematical entities can be "a crucial part of a STEM document," suggesting that ignoring them leaves important decision-making factors unexplained.

Mathematical entities improve transparency in document classification: By linking mathematical formulas to subject classifications, the researchers found that mathematical content provides "high explainability" for why documents are categorized as they are. This means AI systems analyzing scientific papers could justify their decisions by pointing to specific equations and mathematical concepts, not just keywords.

Domain expertise remains valuable but can be augmented: While human domain experts currently classify STEM documents manually, semi-supervised machine learning can support them by predicting classifications for new documents. The research suggests this human-AI collaboration works better when the AI can explain its reasoning through both textual and mathematical evidence.

Current AI tools are inadequate for specialized content: The paper identifies that STEM documents are "more difficult to analyze" than legal or medical texts because of their mathematical complexity. This highlights a gap in responsible AI development—existing explainability methods haven't caught up to the needs of scientific and technical fields.