Exploiting Rich Syntactic Information for Relationship Extraction from Biomedical Articles

Yudong Liu, Zhongmin Shi, Anoop SarkarNAACL 2007

TL;DR

This paper introduces a novel method for extracting "PROTEIN-ORGANISM-LOCATION" (POL) relationships from biomedical text by leveraging rich syntactic features from parse trees. Their approach significantly outperforms simpler methods, achieving the best accuracy by combining Semantic Role Labeling (SRL) features with tree kernels.

What Problem Does This Paper Address?

Identifying functional relationships between biomedical entities, such as where a protein is located within an organism, is crucial for understanding biological processes and identifying potential drug targets. While previous work has focused on predicate-argument structures or shallow semantic features, this paper tackles the challenge of extracting these complex ternary relationships (PROTEIN-ORGANISM-LOCATION) by exploring the power of detailed syntactic information derived from sentence parse trees. They aim to bridge the gap between existing methods and a more comprehensive understanding of these interactions.

Methods

The core of this paper's contribution lies in its sophisticated use of syntactic information. They propose a ternary relation extraction method that is primarily based on rich syntactic features extracted from parse trees.

  • Ternary to Binary Reduction: The POL ternary relation is first simplified into two binary relations: PROTEIN-ORGANISM (PO) and PROTEIN-LOCATION (PL). The ORGANISM-LOCATION relation is ignored as it doesn't directly involve the PROTEIN.
  • SRL-Inspired Features: They adapt features commonly used in Semantic Role Labeling (SRL) for their relation extraction task. This includes features derived from the syntactic path between entities, such as words, part-of-speech tags, head words, and subcategorization information.
  • Tree Kernels: In addition to SRL features, they employ tree kernels to capture structural information from the entire parse tree. This allows them to consider the broader syntactic context.
  • Composite Kernel: To integrate these different sources of information, they use a composite kernel approach, combining a linear kernel for flat features with a tree kernel for structured features.
  • SVM Learner: An Support Vector Machine (SVM) is used as the learning algorithm to classify these extracted features.

Datasets

The experimental data set is derived from a small, expert-curated corpus. This corpus contains labeled PROTEIN, ORGANISM, and LOCATION named entities (NEs) and their POL relations. Specifically, it includes approximately 150,000 words, with 565 POL relation instances, 371 PO relation instances, and 431 PL relation instances. The data was parsed using the Charniak-Johnson parser.

Key Insights and Ideas

  • Syntactic Depth Matters: The paper strongly suggests that richer, more detailed syntactic information from parse trees is significantly more effective for relation extraction than simpler, word-based features.
  • SRL Features are Powerful: Adapting features from Semantic Role Labeling, which are designed to capture argument structures, proves to be highly beneficial for identifying biomedical relations.
  • Combining Views is Key: The best performance is achieved not by relying on a single type of syntactic information but by combining different perspectives: the detailed, path-based SRL features and the global, structural information from tree kernels.
  • PROTEIN as Predicate: Treating the PROTEIN entity as the "predicate" in their simplified binary relation framework is a sensible heuristic that helps anchor the extraction process.

Key Results

The experiments demonstrate a clear advantage of using rich syntactic features.

  • The best performing system, TRK+SRL, which combines SRL features with tree kernels over the entire parse tree, achieved an overall accuracy of 71.8% for ternary relation extraction.
  • This significantly outperforms a shallow word-based baseline system, which achieved only 56.3% accuracy.
  • The paper also reports results for individual binary relations (PL and PO), where the SRL-based approach generally shows strong performance.

Limitations and Future Work

  • Data Size: The authors acknowledge that their experimental data set is relatively small (~150k words).
  • Future Work: The authors plan to explore taking relevant labeled NEs from a generative model as input for their system in future work. They also mention that their ultimate goal is to facilitate biologists in generating annotated corpora, implying a need for more user-friendly or interactive systems.

Summary updated

BibTeX
@inproceedings{liu2007exploiting,
  author = {Yudong Liu and Zhongmin Shi and Anoop Sarkar},
  title = {{Exploiting Rich Syntactic Information for Relationship Extraction from Biomedical Articles}},
  booktitle = {Conference of the North American Chapter of the Association for Computational Linguistics},
  year = {2007},
}
Figure