In the biological sciences, a name is rarely a unique identifier. When a researcher writes about "p53" or "Sonic Hedgehog," they are using a linguistic shorthand that masks a deep layer of ambiguity. To a computer trying to build a knowledge graph, that same string of characters could refer to a segment of DNA (a gene), its messenger RNA transcript, or the resulting functional protein. More complicating still, nearly identical versions of these entities exist across thousands of different organisms, from humans to fruit flies.
The research direction of Biomedical Entity Normalization (BEN)—often called Gene Normalization (GN)—seeks to bridge the gap between this messy, natural language and structured biological databases like UniProt or Entrez. The goal is to "ground" every mention in a paper to a single, unambiguous database ID.
The Three Layers of Ambiguity
The difficulty of this task stems from three distinct types of nomenclature overlap that researchers have identified across decades of work.
1. Common Language Overlap Many gene symbols are identical to common English words. As Joachim Baran and Martin Gerner point out in their work on pubmed2ensembl, automated systems frequently trip over genes named "WAS," "FOR," or "IF." Without sophisticated filtering, a simple string-matching algorithm will see a preposition and incorrectly link it to a genomic record.
2. Functional Polysemy Even within a single species, a term like "insulin" is polysemous. It can describe the gene sequence, the hormone protein, or the broader signaling pathway. Sampo Pyysalo and colleagues have argued that this lack of specificity in natural language hinders deductive reasoning, requiring formal ontology design patterns to help machines distinguish between a gene and its products.
3. The Species Problem This is the "boss level" of normalization. Because of evolution, different species share "orthologous" genes—entities that perform similar functions and often share the exact same name. David Kartchner and his team found that entity linking models perform significantly worse on genes and proteins than on diseases because they fail to integrate species-specific context. If a paper discusses a protein interaction, the system must first determine if the authors are talking about a human, a mouse, or a yeast cell before it can assign the correct ID.
Mechanisms for Disambiguation
To solve these overlaps, researchers have moved beyond simple dictionaries toward context-aware architectures.
Contextual Fingerprinting One of the most effective ways to resolve an ambiguous name is to look at its "neighbors." Jorma Boberg and Tapio Salakoski pioneered techniques that weight context words based on their distance from the target term, helping models distinguish between a gene and a protein sense. Similarly, Xinglong Wang and Sophia Ananiadou reframed disambiguation as a binary relation task, training models to look for "cue words" like "human" or "murine" to anchor the entity to a specific organism.
Integrating Biological Metadata Purely linguistic models are often supplemented with "wet lab" data. Jörg Hakenberg and Ulf Leser demonstrated that incorporating chromosomal locations and cytogenetic bands as features can boost disambiguation precision by over 60%. If a paper mentions a gene name and a specific chromosome, the system can cross-reference this with genomic databases to rule out look-alikes on other chromosomes.
Collective Inference Modern approaches, such as the KECI framework by Heng Ji and ChengXiang Zhai, use Graph Convolutional Networks to perform "collective inference." Instead of deciding the identity of each entity in isolation, the model looks at the global network of relationships in a document. If three other proteins in the sentence are known to exist in S. cerevisiae, the model can use that global context to infer that the fourth, ambiguous protein is likely the yeast variant as well.
The Role of Taxonomy Intelligence
The consensus across the field, from the BioCreative III challenges to recent LLM benchmarking, is that gene normalization is inseparable from species identification. Tools like LINNAEUS, developed by Martin Gerner and Casey Bergman, provide the taxonomic foundation necessary for this work. By first mapping mentions of "rat," "mouse," or "patient" to NCBI Taxonomy IDs, these systems create a "filter" that narrows down the millions of possible gene candidates to a manageable, species-specific subset.
As we move toward large-scale event extraction—such as the EVEX project which processed the entirety of PubMed—the ability to resolve these cross-species ambiguities remains the primary bottleneck for turning scientific prose into actionable biological data.