Abstract
Dana Movshovitz-Attias, William W. Cohen. BioNLP: Proceedings of the 2012 Workshop on Biomedical Natural Language Processing.
2012.
Introduction
The exponential growth of biomedical literature necessitates automated information extraction (IE) systems to manage and synthesize scientific knowledge. Semi-supervised bootstrapping systems, such as the Never-Ending Language Learner (NELL), provide a powerful framework for open IE. These systems start with a small set of "seed" examples for specific categories and iteratively expand their knowledge base by identifying new instances through pattern matching in large text corpora.
However, the biomedical domain presents a unique challenge: extreme terminology ambiguity. Many gene names, particularly in model organisms like Drosophila, are common English words (e.g., "white", "arm", "dad", "28"). In standard bootstrapping, using such ambiguous terms as seeds leads to semantic drift, where the system begins extracting non-biomedical entities (e.g., colors or family members), rapidly degrading the precision of the knowledge base.
Methodology: BioNELL and Rank-and-Learn
To address these challenges, the authors developed BioNELL, an adaptation of NELL specifically for the biomedical domain. BioNELL incorporates three major components:
- Ontology Merging: Six major biomedical ontologies (including the Gene Ontology and Human Disease Ontology) were automatically merged into a hierarchical structure to provide initial categories.
- PMI Collocation Ranking: To mitigate semantic drift, the system evaluates potential seeds using Pointwise Mutual Information (PMI). By measuring how often a term appears near its category name in a massive, domain-independent Web corpus (ClueWeb09), BioNELL can identify and filter out polysemous terms.
- Rank-and-Learn Bootstrapping: Unlike standard bootstrapping, which promotes all newly extracted facts to seeds, BioNELL ranks facts at every iteration. Only those with high PMI scores are promoted to seeds for further learning. Ambiguous but technically correct facts are stored in the knowledge base but "quarantined" from the learning loop to prevent drift.
Experimental Results
BioNELL was evaluated using a 200,000-article PubMed corpus. The system's performance was compared against standard NELL and a "BioNELL+Random" baseline that uses the Rank-and-Learn algorithm but starts with random seeds.
Lexicon Precision
In the task of learning a lexicon of fruit-fly genes, BioNELL achieved a precision of 83%, a significant improvement over the 29% precision of standard NELL. BioNELL also consistently outperformed NELL in other categories, such as Chemical Compounds (66% vs 15% precision) and Drugs (52% vs 40% precision).
Named-Entity Recognition (NER)
The utility of the learned lexicons was further tested in a downstream NER task. A lexicon generated by BioNELL achieved 90% precision in identifying genes in free text. In contrast, using a complete, official gene dictionary resulted in a precision of only 9%, as the official nomenclature includes many highly ambiguous terms that cause false positives in natural language.
Conclusion
BioNELL demonstrates that general-purpose web data can effectively "police" domain-specific learning. By ranking and quarantining ambiguous seeds, the Rank-and-Learn methodology maintains high precision in the face of polysemous terminology. While recall remains a challenge—likely due to corpus size and iteration limits—the approach provides a robust foundation for building high-quality biomedical knowledge bases from scientific text.
Summary updated