Formula-Text Cross-Retrieval: A Benchmarking Study of Dense Embedding Methods for Mathematical Information Retrieval

AbstractMathematical information retrieval requires understanding the complex relationship between natural language and formulae. This paper presents a benchmarking study on Formula-Text Cross-Retrieval, comparing a sparse baseline (BM25), off-the-shelf dense embeddings (OpenAI, BGE), and a fine-tuned dual-encoder model. Our model, trained with a contrastive objective on the ARQAR dataset, significantly outperforms all baselines, achieving state-of-the-art results. Ablation studies confirm the importance of linearization, a shared-weight architecture, and the Multiple Negatives Ranking loss. The work provides a strong foundation for mathematical NLP applications.