From POS tagging to dependency parsing for biomedical event extraction

Area of Science:

• Biomedical Natural Language Processing
• Computational Linguistics
• Bioinformatics

Background:

• Accurate relation and event extraction from biomedical literature is crucial for knowledge synthesis.
• Syntactic information plays a vital role in these information extraction tasks.
• Evaluating different syntactic processing approaches in the biomedical domain is essential.

Purpose of the Study:

• To empirically compare traditional feature-based and neural network-based models for part-of-speech (POS) tagging and dependency parsing.
• To analyze the performance of these models on benchmark biomedical corpora (GENIA and CRAFT).
• To investigate the impact of parser choice on downstream biomedical event extraction tasks.

Main Methods:

• Comparative analysis of state-of-the-art feature-based and neural network models.
• Evaluation on GENIA and CRAFT corpora for POS tagging and dependency parsing.
• Task-oriented evaluation for biomedical event extraction.

Main Results:

• Neural network models generally outperformed traditional feature-based models on both GENIA and CRAFT corpora.
• A detailed analysis of neural models on biomedical data was provided, filling a gap in recent research.
• Task-oriented evaluation revealed that superior intrinsic parsing performance does not consistently lead to improved extrinsic event extraction performance.

Conclusions:

• A comprehensive empirical study comparing traditional and neural models for biomedical POS tagging and dependency parsing was conducted.
• The study highlights the performance differences between model types in the biomedical context.
• The influence of parser selection on a downstream biomedical event extraction task was investigated.