An Intelligent Search Engine and GUI-based Efficient MEDLINE Search Tool Based on Deep Syntactic Parsing

Tomoko Ohta, Yusuke Miyao, Takashi Ninomiya, Yoshimasa Tsuruoka, Akane Yakushiji, Katsuya Masuda, Jumpei Takeuchi, Kazuhiro Yoshida, Tadayoshi Hara, Jin-Dong Kim, Yuka Tateisi, Jun'ichi TsujiiACL 2006 · OpenReview

Abstract: Tomoko Ohta, Yusuke Miyao, Takashi Ninomiya, Yoshimasa Tsuruoka, Akane Yakushiji, Katsuya Masuda, Jumpei Takeuchi, Kazuhiro Yoshida, Tadayoshi Hara, Jin-Dong Kim, Yuka Tateisi, Jun’ichi Tsujii. Proceedings of the COLING/ACL 2006 Interactive Presentation Sessions.

2006.

Introduction

Biomedical researchers currently face a significant challenge: an overwhelming volume of data. MEDLINE, the primary repository for life science research, contains millions of abstracts, making manual information retrieval nearly impossible. Traditionally, researchers have relied on keyword-based Information Retrieval (IR). However, keywords are inherently "blind" to the relationships between entities.

For instance, a researcher seeking instances where Protein A inhibits Protein B might use a keyword search for both proteins. This approach frequently returns thousands of irrelevant results where the proteins are merely mentioned together, or worse, where the relationship is reversed (Protein B inhibits Protein A). To advance automated discovery, computational systems must move beyond simple word matching to understand the semantic structure of sentences—identifying the specific subjects, verbs, and objects that define biological events.

Deep Syntactic Parsing with Enju

The core of this research is Enju, a high-speed, practical parser based on Head-Driven Phrase Structure Grammar (HPSG). Unlike "shallow" parsers that only identify parts of speech, Enju performs deep syntactic parsing to identify predicate-argument relations.

Enju parser output showing deep syntactic analysis and predicate-argument structures.

Enju transforms a sentence into a set of structured tuples: [predicate type, head word, argument label, argument head]. This structural understanding allows the system to recognize that "A was inhibited by B" and "B inhibited A" describe the same biological event. By focusing on the "who-did-what-to-whom," Enju provides the foundation for more intelligent text mining that can handle complex linguistic variations like relative clauses, passive voice, and coordinate structures.

Search Applications: MEDIE and Info-PubMed

Building on the Enju parser, the researchers developed two primary tools to facilitate biomedical discovery:

MEDIE (Semantic Search Engine)

MEDIE pre-annotates the entire MEDLINE database using Enju. It offers a "Semantic Search" interface where users can specify a Subject, Verb, and Object. The system translates these inputs into a formal Graph Command Language (GCL) query to locate matches within the parsed corpus. This allows for highly specific queries, such as "What diseases does this protein cause?", with significantly higher precision than keyword searches.

MEDIE semantic search results for 'dystrophin cause', highlighting subjects, verbs, and objects in the retrieved abstracts.

Info-PubMed (GUI Tool)

Info-PubMed is a visual tool specifically designed for exploring protein-protein interactions. It leverages parsed data to distinguish between simple "co-occurrence" (two proteins appearing in the same sentence) and "direct evidence" of an interaction based on the sentence's grammatical structure. This distinction is crucial for filtering out the "noise" common in bioinformatics and identifying actual biological events.

Datasets and Evaluation

The researchers utilized several key datasets to train and evaluate the system:

  • Penn Treebank (PTB): Used to train the general English grammar of the Enju parser.
  • GENIA Corpus: A specialized biomedical corpus of 2,000 MEDLINE abstracts used to tune the parser to the linguistic style of medical research.
  • Dictionary Resources: A massive dictionary of 4,467,855 entries derived from the GENA dictionary and the Unified Medical Language System (UMLS) was used for entity recognition.

In performance evaluations, the Enju parser achieved an F-score of 87.16% on the Penn Treebank and 86.87% on the GENIA corpus. These results demonstrate that the parser remains highly reliable even when transitioning from general English to specialized scientific text.

Conclusion and Future Directions

The development of Enju, MEDIE, and Info-PubMed demonstrates that deep syntactic parsing is computationally feasible at the scale of millions of documents. By prioritizing practicality and speed, the authors moved HPSG from a complex linguistic theory into a functional tool for large-scale web mining.

However, limitations remain. The system relies heavily on dictionary matching for entity recognition, which can struggle with newly discovered proteins or non-standard naming conventions. Future work aims to use these deep parsing results for even more advanced text mining, potentially moving toward automated knowledge base construction that maps entire networks of biological interactions across the literature.

Summary updated

BibTeX
@inproceedings{ohta2006intelligent,
  author = {Tomoko Ohta and Yusuke Miyao and Takashi Ninomiya and Yoshimasa Tsuruoka and Akane Yakushiji and Katsuya Masuda and Jumpei Takeuchi and Kazuhiro Yoshida and Tadayoshi Hara and Jin-Dong Kim and Yuka Tateisi and Jun'ichi Tsujii},
  title = {{An Intelligent Search Engine and GUI-based Efficient MEDLINE Search Tool Based on Deep Syntactic Parsing}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2006},
  url = {https://openreview.net/forum?id=yUdOSNzYvP8},
}
Figure