Simultaneous Identification of Biomedical Named-Entity and Functional Relation Using Statistical Parsing Techniques

Zhongmin Shi, Anoop Sarkar, Fred PopowichNAACL 2007

TL;DR

This paper introduces a novel statistical parsing approach to simultaneously identify biomedical entities (like proteins and locations) and extract their functional relationships, specifically bacterial protein localization (BPL). By leveraging a semi-supervised method with automatically labeled "noisy" data, they achieve a dramatic improvement in relation extraction accuracy, reaching 83.2% F-score. This work is crucial for building comprehensive biological databases and advancing drug discovery by automating the extraction of vital protein information from vast scientific literature.

What Problem Does This Paper Address?

Extracting structured information from unstructured text is a fundamental challenge in many scientific fields, especially in biology. Specifically, understanding where a protein is located within a bacterium (its subcellular localization) is critical for deciphering its function and for developing targeted therapies. While Named-Entity Recognition (NER) can identify proteins, bacteria, and locations, linking them to form functional relationships like "Protein X is located in the cytoplasm of Bacterium Y" is a more complex task. Existing methods often rely on separate steps for entity recognition and relation extraction, which can lead to error propagation. Furthermore, creating large, manually annotated datasets for training these systems is time-consuming and expensive. This paper tackles the problem of efficiently and accurately extracting these "Bacterial Protein Localization" (BPL) relations from biomedical literature, aiming to build a more comprehensive and automated approach.

Methods

The core of this paper is a statistical parsing technique that simultaneously identifies biomedical named-entities (NEs) and extracts functional relations. Instead of a traditional pipeline approach (NER first, then relation extraction), their parser integrates both tasks.

Here's a breakdown of their approach:

  • Syntactic and Semantic Parsing: They use a lexicalized statistical parser (Charniak and Johnson, 2005) to generate parse trees. These trees are then enriched with two types of semantic annotations:
  • Relevant NEs: Specific tags (e.g., PROTEIN_R, BACTERIUM_R, LOCATION_R) are added to entities deemed relevant to a potential BPL relation within a sentence.
  • Path Annotations: Nodes along the path in the parse tree connecting relevant NEs are tagged (LNK, PTR) to represent the syntactic connection.
  • Ternary to Binary Relation Splitting: The BPL relation, which involves three entities (BACTERIUM, PROTEIN, LOCATION), is split into two binary relations: BP (BACTERIUM and PROTEIN) and PL (PROTEIN and LOCATION). This simplifies the parsing and extraction process.
  • Semi-Supervised Learning: Recognizing the scarcity of labeled data, they employ a semi-supervised approach. They leverage a large corpus of unlabeled MEDLINE articles. They assume that any sentence containing a BACTERIUM, PROTEIN, and LOCATION NE likely expresses a BPL relation. This "noisy" data, where the assumed relations might not always be correct, is used to train the parser. This approach allows the model to learn preferences for structural relations between entities that appear close together in a sentence.
  • Fusion of Binary Relations: The extracted BP and PL relations are then fused to predict the final BPL relation.

Datasets

  • Curated Corpus: A small, expert-curated corpus was used for initial training and evaluation. The paper mentions this corpus is used for "fully supervised" experiments and as part of the "semi-supervised" training.
  • MEDLINE Articles: A large collection of unlabeled MEDLINE articles was used as the source for automatically generating "noisy" training data for the semi-supervised approach. This is a key aspect of their method, as it bypasses the need for extensive manual annotation.

Key Insights and Ideas

  • Simultaneous Identification and Extraction: The most significant insight is that performing NER and relation extraction simultaneously within a single parsing framework can be more effective than a pipelined approach. This avoids compounding errors from separate systems.
  • The Power of "Noisy" Data: The paper demonstrates that even automatically generated, potentially incorrect (noisy) training data can be highly beneficial for training relation extraction models, especially when labeled data is scarce. This is a crucial finding for making progress in data-scarce domains.
  • Learning Structural Preferences: The semi-supervised method's success suggests that the parser learns to identify structural patterns in sentences that strongly indicate a relationship, even if the specific entities aren't perfectly labeled. This allows it to distinguish between competing relations within a sentence.
  • Bridging the Gap to Databases: By automating the extraction of BPL relations, this work directly contributes to populating biological databases like PSORTdb, making valuable information more accessible to researchers.

Key Results

The paper presents compelling results, particularly highlighting the effectiveness of their semi-supervised approach:

  • Baseline: A simple baseline system achieved an F-score of 41.0% for BPL extraction.
  • Fully Supervised Parsing: A fully supervised approach, trained only on the curated data, achieved an F-score of 48.4% for BPL extraction. This was limited by the small size of the training data.
  • Semi-Supervised Parsing (Noisy Data Only): Using only the automatically generated noisy data for training, the F-score for BPL extraction reached 74.5%.
  • Semi-Supervised Parsing (Noisy Data + Curated Data): Combining noisy data with the curated data for training yielded the best results, achieving a remarkable 83.2% F-score for BPL extraction. This represents a significant improvement over the fully supervised method.
  • NER Improvement: The semi-supervised method also dramatically improved recall for NER, with PROTEIN NER recall increasing from 25.0% to 81.3%.

Limitations and Future Work

  • Limitations:
    • Syntactic Parsing Quality: The paper acknowledges that the quality of the syntactic parsing annotation can impact performance, especially in the fully supervised setting where the lack of training data led to lower recall for PROTEIN NER.
  • Alignment Issues in Pipelined Systems: They note that aligning entities from a separate NER system with syntactic parses in a pipelined approach is non-trivial and was a challenge they avoided by integrating the tasks.
  • Future Work:
    • Discriminative Models: The authors plan to integrate their parser's output with discriminative models (like the one they implemented in Liu et al., 2007) to further refine relation extraction.
  • Graphical Models: They aim to train graphical models that can infer relations from multiple sentences and documents, moving beyond single-sentence analysis.
  • Explicit Bootstrapping: They hope to explore explicit bootstrapping techniques to improve the quality of the noisy training data.

Summary updated

BibTeX
@inproceedings{shi2007simultaneous,
  author = {Zhongmin Shi and Anoop Sarkar and Fred Popowich},
  title = {{Simultaneous Identification of Biomedical Named-Entity and Functional Relation Using Statistical Parsing Techniques}},
  booktitle = {Conference of the North American Chapter of the Association for Computational Linguistics},
  year = {2007},
}
Figure