Extracting Biomedical Event with Dual Decomposition Integrating Word Embeddings

Area of Science:

• Biomedical Natural Language Processing
• Computational Biology
• Information Extraction

Background:

• Biomedical event extraction from literature is crucial but challenging.
• Existing pipeline systems suffer from cascading errors.
• One-hot word encoding lacks semantic representation.

Purpose of the Study:

• To develop an improved system for biomedical event extraction.
• To address cascading errors and semantic information limitations in current methods.
• To enhance the accuracy of detecting triggers and arguments in biomedical texts.

Main Methods:

• Utilized joint inference with dual decomposition to mitigate cascading errors.
• Implemented novel word embeddings learned from large unlabeled texts.
• Integrated unsupervised word embeddings with dependency parse graph features.
• Developed a four-component system: trigger detection, argument detection, joint inference, and rule-based post-processing.

Main Results:

• Achieved superior performance compared to state-of-the-art systems on BioNLP datasets.
• Obtained an F-score of 59.77% on BioNLP'09 development set, surpassing the best system by 0.96%.
• Reached an F-score of 56.09% on BioNLP'11 test set, outperforming previous results by 0.89%.
• Attained an F-score of 53.19% on BioNLP'13 test set, a 2.22% improvement over the best published result.

Conclusions:

• The proposed system effectively extracts biomedical events by combining joint inference and advanced word embeddings.
• The novel approach significantly enhances accuracy and overcomes limitations of traditional pipeline methods.
• Demonstrated state-of-the-art performance on multiple benchmark datasets, validating the system's efficacy.