Chemical entity recognition in patents by combining dictionary-based and statistical approaches

Area of Science:

• Biomedical Informatics
• Natural Language Processing
• Computational Chemistry

Background:

• Accurate identification of chemical entities in patents is crucial for scientific literature analysis.
• Existing methods require robust systems for both chemical mention recognition and passage classification.

Purpose of the Study:

• To develop and evaluate an ensemble system for chemical entity recognition in patents.
• To address the Chemical Entity Mention in Patents (CEMP) and Chemical Passage Detection (CPD) tasks in the BioCreative 2015 CHEMDNER-patent track.

Main Methods:

• An ensemble approach combining dictionary-based and statistical methods.
• Utilized Peregrine open-source indexing engine for lexical resource assessment.
• Integrated tmChem (conditional random field classifier) with additional features: part-of-speech tags, lemmas, and word-vector clusters.

Main Results:

• Achieved an F-score of 85.21% for CEMP and 91.53% accuracy for CPD on training data.
• On the test set, ranked sixth out of 21 teams for CEMP (86.82% F-score) and second out of nine teams for CPD (94.23% accuracy).

Conclusions:

• The developed ensemble system demonstrates strong performance in chemical entity recognition within patent documents.
• The combination of dictionary and statistical approaches, enhanced with linguistic features, effectively addresses complex NLP tasks in cheminformatics.