Abstract
Mathematical symbols in scientific literature are often ambiguous, with the same identifier representing vastly different concepts across contexts. This paper introduces the task of variable typing, which automatically assigns specific mathematical concepts to variables based on their surrounding prose. We present a large-scale annotated dataset and demonstrate that deep learning models, particularly BiLSTMs, can effectively classify these relationships. Our results show that integrating these types into Mathematical Information Retrieval (MIR) systems nearly doubles search precision.
Introduction
In scientific papers, a variable like $E$ might represent "energy" in one context and an "elliptic curve" in another. For computational systems to accurately retrieve or reason about mathematical content, they must bridge the gap between symbolic formulas and textual descriptions. Previous attempts at "description extraction" produced free-form text that was difficult to normalize. We propose variable typing: a classification task where variables are mapped to a structured dictionary of mathematical types.
Methods
Variable typing is framed as a binary classification problem: determining if a specific variable $V$ in a sentence should be linked to a candidate type $T$ within the same sentence.
The authors evaluated several architectures, including a Nearest Type baseline, SVMs with hand-crafted features, and Convolutional Neural Networks. The top-performing model was a BiLSTM (bidirectional Long Short-Term Memory network) using word and character-level embeddings. It employs a "predictive gating" mechanism for out-of-vocabulary words and leverages "supertype" information to understand hierarchical relationships (e.g., recognizing that a "Riemannian manifold" is a specific kind of "manifold").
Datasets
The study utilizes a new benchmark derived from the arXiv-based Mathematical REtrieval Corpus (MREC). The dataset contains over 33,000 annotated edges between variables and types.
To support this, the authors expanded a seed list of 10,601 types into a dictionary of 1.23 million technical terms using a double suffix trie algorithm, capturing complex compositional terms like "locally compact Hausdorff space."
Key Insights: Solving Alpha-Equivalence
A core challenge in mathematical search is $\alpha$-equivalence: the structural identity of formulas like $x + y$ and $a + b$. Variable typing provides the metadata necessary to distinguish these. If $x$ and $y$ are typed as "vectors," the system can differentiate the formula from one where $a$ and $b$ are "matrices."
Results
The BiLSTM model achieved an F1-score of 78.98%, with a high precision of 83.11%. When integrated into a Mathematical Information Retrieval (MIR) system, the "typed index" significantly outperformed standard methods. The Mean Average Precision (MAP) increased from 0.083 (untyped) to 0.139 (typed), demonstrating the practical utility of the task for improving search relevance.
Limitations and Future Work
Currently, the system is limited to "simple expressions" (single identifiers) and operates only within single-sentence contexts. Future research will focus on handling complex mathematical objects, such as nested functions, and expanding the context window to capture definitions provided in preceding paragraphs or pages.
Summary updated