GrantExtractor: Accurate Grant Support Information Extraction from Biomedical Fulltext Based on Bi-LSTM-CRF

Area of Science:

• Biomedical Informatics
• Computational Biology
• Information Extraction

Background:

• Grant support (GS) information in MEDLINE, comprising funding agencies and contract numbers, is crucial for tracking research funding outcomes.
• Accurate and automated extraction of GS information from biomedical literature presents a significant challenge.
• Effective tracking of funding outcomes is essential for funding organizations.

Purpose of the Study:

• To develop and present GrantExtractor, a pipeline system for accurate and automated extraction of grant support information from full-text biomedical literature.
• To integrate advanced machine learning techniques for robust information extraction.

Main Methods:

• Utilized a sentence classifier to identify sentences containing funding information.
• Employed a bi-directional LSTM and CRF layer (BiLSTM-CRF) combined with pattern matching for entity extraction (grant numbers and agencies).
• Implemented a multi-class model to filter noisy numbers and a matching algorithm to pair grant numbers with their corresponding agencies.

Main Results:

• GrantExtractor demonstrated superior performance compared to baseline methods on benchmark datasets.
• Achieved first place in Task 5C of the 2017 BioASQ challenge, with a Micro-recall of 0.9526 for 22,610 articles.
• Attained a Micro F-measure score of 0.90 for extracting grant pairs.

Conclusions:

• GrantExtractor provides an effective solution for accurately extracting grant support information from biomedical literature.
• The system's performance highlights the potential of integrated machine learning techniques for biomedical text mining.
• The success in the BioASQ challenge validates GrantExtractor's capability in real-world applications.