Biomedical Event Extraction using Abstract Meaning Representation

Sudha Rao, Daniel Marcu, Kevin Knight, Hal Daumé IiiBioNLP 2017 · OpenReview

Abstract

We propose a novel, Abstract Meaning Representation (AMR) based approach to identifying molecular events and interactions in biomedical text. Our key contributions include: (1) an empirical validation of the hypothesis that a biological event corresponds to a subgraph within an AMR graph, (2) a neural network-based pipeline that identifies these event subgraphs, and (3) a distant supervision approach to augment training data. Evaluating our system on the Genia 2013 Event Extraction dataset shows that semantic parsing effectively normalizes linguistic variations, offering a promising alternative to traditional syntactic methods.

1. Introduction

Building curated knowledge bases of molecular interactions is essential for biomedical research, yet the volume of literature makes manual curation impossible. Automated extraction is challenged by the flexibility of natural language; identical biological events can be described using diverse verbs, voices, or nested structures. While existing methods rely on syntactic parses or pattern matching, they often fail to generalize across different sentence structures. This paper addresses these limitations by utilizing Abstract Meaning Representation (AMR), a semantic formalism that captures "who did what to whom" consistently, regardless of linguistic variation.

2. Methodology

2.1 The Subgraph Hypothesis

The core of our approach is the Subgraph Hypothesis: biological events correspond to specific subgraphs within a sentence's AMR graph. We empirically validated this by mapping annotated events from standard datasets to AMR structures. This representation allows the model to treat event extraction as a structural identification task rather than a surface-level text processing task.

An AMR graph illustrating how biological events (E1, E2) map to specific subgraphs.

2.2 Neural Pipeline for Event Identification

We developed a two-stage pipeline using Long Short-Term Memory (LSTM) networks to identify these subgraphs:

  1. Theme Identification: A multi-class classifier that analyzes the shortest path in the AMR graph between an interaction node (e.g., inhibit-01) and a protein node to determine the event type (e.g., Phosphorylation).
  2. Cause Identification: A binary classifier that determines if an entity or another event is the "cause" of the interaction, enabling the extraction of complex, nested regulatory events.
The two-stage classification process for identifying event themes and causes based on AMR paths.

2.3 AMR-based Distant Supervision

To combat the scarcity of labeled biomedical data, we employed distant supervision using the BioPax database. We identified sentences in PubMed Central that mention known interacting proteins and applied a novel AMR-path heuristic. By requiring the shortest AMR path between proteins to contain a relevant interaction term, we filtered out noise, ensuring the extracted sentences truly described the biological relationship.

3. Experimental Results

3.1 Performance and Impact

The model achieved an overall F1-score of 46.66% on the Genia 2013 development set, placing it within 5% of the state-of-the-art EVEX system. The introduction of 11,792 distantly supervised sentences significantly boosted recall, particularly for "Gene expression" events, which saw a recall increase from 66.33% to 76.98%.

3.2 Semantic Normalization

Our results confirm that AMR acts as a powerful "normalizer." By focusing on concepts rather than syntax, it reduces the feature space and allows the model to learn more robust patterns. We calculated that with perfect AMR parsing, this subgraph approach could reach an upper-bound F1-score of approximately 74%.

4. Conclusion

This research demonstrates that a purely semantic, AMR-based approach is a viable and promising direction for biomedical event extraction. While currently bottlenecked by the accuracy of general-purpose AMR parsers, the system's performance is expected to improve as parsing technology evolves. Future work will focus on more sophisticated sentence selection for distant supervision and leveraging larger biological databases to further expand the training corpus.

Summary updated

Abstract

We propose a novel, Abstract Meaning Representation (AMR) based approach to identifying molecular events/interactions in biomedical text. Our key contributions are: (1) an empirical validation of our hypothesis that an event is a subgraph of the AMR graph, (2) a neural network-based model that identifies such an event subgraph given an AMR, and (3) a distant supervision based approach to gather additional training data. We evaluate our approach on the 2013 Genia Event Extraction dataset and show promising results.
BibTeX
@inproceedings{rao2017biomedical,
  author = {Sudha Rao and Daniel Marcu and Kevin Knight and Hal Daumé III},
  title = {{Biomedical Event Extraction using Abstract Meaning Representation}},
  booktitle = {BioNLP},
  year = {2017},
  url = {https://openreview.net/forum?id=HTED-rup0ZZ},
}
Figure