Comparing and combining chunkers of biomedical text

Area of Science:

• Biomedical Natural Language Processing
• Computational Linguistics
• Information Extraction

Background:

• Text chunking is crucial for biomedical information extraction.
• No prior comparative studies exist for biomedical chunking systems.
• Existing systems lack comprehensive evaluation for usability and performance.

Purpose of the Study:

• To compare the usability and performance of six state-of-the-art chunkers in the biomedical domain.
• To evaluate chunking systems including sentence splitting, tokenization, and part-of-speech tagging.
• To combine chunker results to enhance overall chunking performance.

Main Methods:

• Evaluated six chunkers: GATE, Genia Tagger, Lingpipe, MetaMap, OpenNLP, and Yamcha.
• Integrated chunkers into the Unstructured Information Management Architecture (UIMA) framework.
• Utilized the GENIA Treebank corpus for training and testing, assessing noun-phrase and verb-phrase chunking.

Main Results:

• OpenNLP achieved the highest performance for both noun-phrase (89.7% F-score) and verb-phrase (95.7% F-score) chunking.
• Lingpipe and OpenNLP offered the best usability (ease of integration, speed, trainability).
• Combining chunker results via voting improved F-scores by up to 3.1% and allowed adjustable precision-recall.

Conclusions:

• OpenNLP is a top-performing and user-friendly chunker for biomedical text.
• Combining multiple chunkers enhances performance and offers flexible precision-recall trade-offs.
• This study provides the first comprehensive comparison and combination strategy for biomedical chunking pipelines.