Mathematical Term Disambiguation Using Semantic Representations

This technical report presents a study on the automatic disambiguation of mathematical terms based on contextualized semantic representations. The document outlines the challenges faced in mathematical literature, where terms can have multiple meanings depending on the context, leading to significant manual effort for disambiguation. It introduces a new dataset, MathD2, specifically constructed for this purpose using ProofWiki’s disambiguation pages. The report details two approaches for disambiguation: supervised classification utilizing embeddings of concatenated definitions and titles, and zero-shot prediction based on semantic textual similarity. Both methods demonstrate high accuracy and macro F1 scores exceeding 0.9 on the ground truth dataset, indicating their effectiveness. The study also discusses the need for an automatically constructed knowledge base of mathematical definitions to enhance the discovery and indexing of relevant mathematical statements. The findings contribute to the field of natural language processing by showcasing the application of contextualized representations in resolving ambiguities in mathematical definitions.