TL;DR
This paper introduces a new corpus and methodology for extracting biological events that span multiple levels of organization, from molecules to whole organisms. By extending existing event extraction techniques to this broader scope, the authors demonstrate the feasibility of automatically understanding complex biological processes described in scientific literature.
What Problem Does This Paper Address?
Understanding biological systems requires tracing cause-and-effect relationships across various scales, from molecular interactions to cellular, tissue, organ, and even whole-organism effects. However, existing biomedical information extraction (IE) and text mining efforts have largely focused on molecular-level entities (like proteins and genes) and events (like binding and phosphorylation). This narrow focus limits their ability to capture the full picture of biological processes. The paper addresses this gap by developing methods and resources to extract events that encompass multiple levels of biological organization, moving beyond the purely molecular view.
Methods
The authors extend the established structured event extraction paradigm, which represents events as n-ary associations of participants in specific roles, linked by a "trigger" span of text. Their approach involves:
- Ontological Foundations: Leveraging existing ontologies (like the Gene Ontology and Common Anatomy Reference Ontology) to define new entity and event types that cover molecular, cellular, tissue, organ, and organism levels.
- Multi-Level Event Extraction (MLEE) Corpus: Creating a new, manually annotated corpus of over 8,000 entities and 6,000 events from abstracts related to angiogenesis. This corpus is unique in its multi-level scope.
- Adaptation of Existing Tools: Adapting state-of-the-art named entity recognition (NER) and event extraction tools (NERsuite and EventMine, respectively) to handle the new multi-level entity and event types. This involved fine-tuning feature sets and potentially introducing new rules.
- Structured Representation: Moving from simple typed spans to a structured representation that can capture complex associations and the direction of causality between participants.
Datasets
- MLEE Corpus: A newly introduced corpus of 262 PubMed abstracts on angiogenesis. This corpus is annotated with multi-level entities and events, serving as both training and evaluation data. It contains approximately 8,291 entities and 6,677 events.
- BioNLP Shared Task 2011 Corpora (GE, EPI, ID): Used for evaluating the "stacking" model in event extraction, where predictions from a model trained on these molecular-focused corpora are incorporated.
Key Insights and Ideas
- Bridging the Gap: The core insight is that existing event extraction frameworks can be successfully adapted to a broader range of biological organization levels, not just molecular ones.
- Ontology-Driven Expansion: The use of established biomedical ontologies provides a principled way to define new entity and event types, ensuring some level of compatibility and semantic grounding.
- Anatomical Events are Extractable: Surprisingly, the study found that events involving anatomical entities were extracted more reliably than molecular-level ones, suggesting that the proposed representation and methods are well-suited for this broader scope.
- High Event Density: The MLEE corpus exhibits a high density of event annotations, which the authors attribute to the richer set of entity and event types allowing for more comprehensive representation of statements.
Key Results
- Entity Mention Detection: Achieved an overall F-score of approximately 80% (with dictionary support), comparable to or exceeding performance on established molecular-level tasks.
- Event Extraction: The best model achieved an F-score of over 50%, which is promising for a first experiment on this novel task and competitive with results from molecular-focused BioNLP Shared Tasks.
- Anatomical Event Performance: Events involving anatomical entities showed higher precision and recall compared to molecular events, indicating the effectiveness of the approach for higher-level biological processes.
- Corpus Contribution: The MLEE corpus, despite being smaller in word count than some molecular-focused corpora, has a high density of annotations, making it a valuable resource for multi-level event extraction research.
Limitations and Future Work
- Domain Independence: The authors acknowledge that the current annotation and extraction methods are somewhat domain-specific. Future work will focus on extending these to improve domain independence, allowing for large-scale annotation of the entire biomedical literature.
- Annotation Consistency: While efforts were made to ensure consistency, the paper mentions referring to supplementary material for evaluation of corpus annotation consistency, implying it's an ongoing area of focus.
- Further Refinement: The authors suggest that further work will focus on improving the domain independence of the annotations and extraction methods.
Summary updated