LINNAEUS: A species name identification system for biomedical literature

Martin Gerner, Goran Nenadic, Casey M. BergmanBMC Bioinformatics 2010 · DOI

Abstract

Identifying species mentions in biomedical literature is a foundational task for text mining, yet it remains challenging due to linguistic ambiguity and the prevalence of common names. LINNAEUS is an open-source system designed to recognize species names and map them to NCBI taxonomy identifiers. By utilizing a high-speed dictionary-based approach combined with robust disambiguation heuristics, LINNAEUS achieves high precision and recall across large-scale corpora.

Introduction

The biomedical literature is expanding at an unprecedented rate, with MEDLINE alone containing over 18 million entries. A critical step in extracting knowledge from these documents is identifying the organism under study. However, automated species recognition faces several hurdles. Ambiguity is rampant; for instance, the abbreviation "C. elegans" can refer to 41 different species in the NCBI taxonomy. Furthermore, many existing tools overlook common names like "human" or "mouse," which dominate literature, or fail to distinguish species names from common English homonyms (e.g., "Spot"). LINNAEUS addresses these challenges to provide a reliable basis for downstream tasks like gene name recognition and taxonomically intelligent document retrieval.

Methods

LINNAEUS employs a dictionary-based approach centered on a Deterministic Finite-State Automaton (DFA). This engine allows for simultaneous searching of hundreds of thousands of species terms in linear time relative to the text length, making it exceptionally fast. The system's dictionary is built from the NCBI taxonomy, supplemented with common names and manually curated synonyms.

Overview of the LINNAEUS system architecture, showing dictionary generation (A) and the document tagging process (B).

The processing pipeline involves several stages: initial tagging, disambiguation, acronym detection, and filtering. Disambiguation is handled through heuristics, such as looking for explicit full scientific names within the same document to resolve ambiguous abbreviations. If ambiguity persists, the system uses probability scoring based on species frequency in MEDLINE—for example, prioritizing Homo sapiens over rarer species when a term is ambiguous.

Results and Evaluation

The system was evaluated against a manually annotated "gold standard" corpus of 100 full-text documents from PubMed Central (PMC). LINNAEUS achieved 94% recall and 97% precision at the mention level. At the document level, it correctly identified at least one species in 98% of documents with 90% precision.

Comparison of document and species counts across various tag sets and corpora, highlighting the scale of LINNAEUS output.

A key finding was the importance of full-text articles for disambiguation. Full texts often provide the "key" (the full scientific name) needed to unlock ambiguous abbreviations used later in the paper. Additionally, the study revealed that common names account for approximately 72% of species mentions in their manual corpus, underscoring the necessity of including them in the dictionary. As shown in the comparative data, LINNAEUS output on MEDLINE identified over 30 million tags across nearly 10 million documents.

Discussion and Conclusion

LINNAEUS demonstrates that a well-optimized dictionary-based approach can outperform complex rule-based systems in both speed and accuracy. While highly effective, limitations remain, such as sensitivity to OCR errors in older documents and the current lack of specific cell line identification. Future iterations may expand to higher-order taxonomic units like genera or families. Nevertheless, LINNAEUS provides a critical open-source infrastructure for the bioinformatics community, bridging the gap between natural language text and structured genomic databases.

Summary updated

BibTeX
@article{gerner2010linnaeus,
  author = {Martin Gerner and Goran Nenadic and Casey M. Bergman},
  title = {{LINNAEUS: A species name identification system for biomedical literature}},
  journal = {BMC Bioinformatics},
  year = {2010},
  doi = {10.1186/1471-2105-11-85},
}
Figure