In natural language processing, resolving word ambiguity is a classic challenge: the word "bank" means something entirely different in a financial context than in a geographical one. In scientific literature, this problem is severely amplified. Mathematical notation is highly constrained by a small alphabet of characters, leading to symbol overloading and contextual polymorphism. A single character like $\sigma$ might represent standard deviation in a statistics paper, stress in a mechanical engineering paper, or a selection operator in a database query. Even within a single document, an author might use $x$ as a local variable in one section and a global constant in another.
Standard language models frequently treat mathematical symbols as static tokens or literal noise, failing to recognize that their underlying meanings shift dynamically. To bridge the gap between abstract symbolic formulas and their intended semantic meaning, a growing body of research focuses on contextually grounding mathematical identifiers within their surrounding text, document structure, and scientific domain.
Grounding Symbols via Variable Typing
The foundational step in resolving symbolic ambiguity is transforming free-form mathematical characters into structured, machine-readable concepts. This process is known as variable typing or identifier grounding. Instead of treating a symbol as an isolated character, models learn to map it to a specific mathematical definition based on its immediate textual context.
Yiannos Stathopoulos et al. pioneered this approach in Variable Typing: Assigning Meaning to Variables in Mathematical Text. By framing the problem as a classification task, the authors demonstrated that deep learning models (such as BiLSTMs) could ingest the natural language surrounding an equation and assign a structured concept (e.g., "parabolic subgroup") to a specific variable. This structured grounding allows downstream applications, like mathematical information retrieval systems, to distinguish between identical symbols representing distinct ideas.
Building on this concept, the community has developed large-scale benchmarks to standardize how models link symbols to their descriptions. Amir Pouran Ben Veyseh et al. introduced the Symlink dataset in SemEval 2022 Task 12: Symlink - Linking Mathematical Symbols to their Descriptions, challenging models to extract explicit relations between symbols and their textual definitions across thousands of scientific LaTeX documents.
However, real-world scientific text introduces practical alignment challenges. As Maria Alexeeva and Rebecca Sharp note in MathAlign: Linking Formula Identifiers to their Contextual Natural Language Descriptions, authors frequently use simplified versions of variables in their prose—omitting subscripts, superscripts, or indices that appear in the formal equations. The authors designed alignment pipelines specifically to bridge this notation gap, ensuring that informal textual mentions are correctly mapped to their complex symbolic counterparts.
Navigating Intra-Document Volatility
Symbolic ambiguity does not just occur across different academic fields; it frequently occurs within the boundaries of a single paper. This phenomenon, known as intra-document polysemy or semantic volatility, occurs when an author reuses the same identifier for different purposes across different sections.
Takuto Asakura et al. quantified this challenge in Building Dataset for Grounding of Formulae - Annotating Coreference Relations Among Math Identifiers. Their analysis revealed that a single identifier type represents, on average, more than two distinct meanings within the same paper. Because symbols undergo rapid "scope switches," automated systems cannot rely on a global document-level dictionary; they must track the active lifespan of a symbol's meaning dynamically.
To resolve these localized shifts, Yusuke Miyao investigated the precise features required for intra-document disambiguation in What Is Needed for Intra-document Disambiguation of Math Identifiers?. The author discovered that a symbol's exact spatial position and its local formula structure are often more critical than the surrounding natural language words. Because math identifiers are brief, their physical placement within the document layout serves as a vital signal for tracking context.
This insight directly informs user-facing reading assistants. For example, Andrew Head et al. developed ScholarPhi, detailed in Augmenting Scientific Papers with Just-in-Time, Position-Sensitive Definitions of Terms and Symbols. By tracking position-sensitive definitions, the system surfaces the exact, locally relevant meaning of an overloaded symbol directly over the equation, preventing readers from applying an outdated definition from a previous section.
Structural and Domain-Level Context
To truly resolve symbolic ambiguity, models must look beyond immediate textual definitions and incorporate broader structural and macro-level domain context.
On a macro scale, the meaning of an ambiguous symbol is heavily constrained by the academic discipline. Michihiro Yasunaga et al. addressed this in TopicEq: A Joint Topic and Mathematical Equation Model for Scientific Texts. by designing a generative model that shares a latent topic space between text and equations. By learning domain-level topical distributions, TopicEq uses the overall theme of the paper to narrow down the highly polysemous meanings of mathematical expressions.
On a structural level, ambiguity also affects operators and functional roles. Dennis Müller 0001 and Cezary Kaliszyk focused on the challenge of autoformalization in Disambiguating Symbolic Expressions in Informal Documents. They demonstrated that resolving whether a symbol like + denotes integer addition, string concatenation, or a custom algebraic operation requires integrating document-level context, as isolated expressions lack the semantic specification required for accurate translation.
Furthermore, models must understand the functional role a symbol plays within an equation's syntax. Jiaru Zou et al. introduced the STEM-PoM benchmark in STEM-POM: Evaluating Language Models Math-Symbol Reasoning in Document Parsing to evaluate whether large language models can distinguish between variables, constants, and operators. Their findings emphasize that misinterpreting a symbol's functional role—such as treating a variable as a constant—leads to severe downstream reasoning errors.
Finally, this structural dependency extends to visual layouts. In tasks like mathematical Optical Character Recognition (OCR), Pegah Ahadian et al. demonstrated in MNIST-Fraction: Enhancing Math Education with AI-Driven Fraction Detection and Analysis that identifying a digit is insufficient on its own. A model must determine the digit's precise spatial relationship to other elements to resolve whether it functions as a numerator, a denominator, or a standalone integer, proving that mathematical meaning is fundamentally tied to spatial and contextual architecture.
Key Foundational Work
Variable Typing: Assigning Meaning to Variables in Mathematical Text (2018) — contributed 2 concept elements to this direction.
Mathematics Content Understanding for Cyberlearning via Formula Evolution Map (2018) — contributed 2 concept elements to this direction.
Towards Formula Translation using Recursive Neural Networks (2018) — contributed 1 concept elements to this direction.