# From POS tagging to dependency parsing for biomedical event extraction

**Authors:** Dat Quoc Nguyen, Karin Verspoor  
**arXiv:** [1808.03731](https://arxiv.org/abs/1808.03731)  
**Date:** 2018-08-11  
**Summary page:** https://www.summarizepaper.com/en/paper/1808.03731v2/

## Key points

- Authors Dat Quoc Nguyen and Karin Verspoor focus on relation and event extraction from biomedical research publications.
- They highlight the significance of syntactic information in this process and aim to determine effective approaches to syntactic processing in the biomedical domain.
- The researchers compare traditional feature-based models with neural network-based models for POS tagging and dependency parsing tasks using GENIA and CRAFT corpora.
- Their analysis shows that neural models generally outperform feature-based models on these benchmark datasets.
- The study includes a task-oriented evaluation to assess how parsing models impact downstream applications like biomedical event extraction.
- Results indicate that superior intrinsic parsing performance doesn't always lead to better extrinsic event extraction performance, showing the complexity of integrating syntactic processing into practical applications.
- Nguyen and Verspoor provide a detailed exploration of traditional and neural network-based models for POS tagging and dependency parsing in biomedicine, emphasizing the importance of parser selection for optimizing downstream task performance.
- Retrained models from their study are publicly available for further research at https://github.com/datquocnguyen/BioPosDep.

## Summary

In their study titled "From POS tagging to dependency parsing for biomedical event extraction," authors Dat Quoc Nguyen and Karin Verspoor address the crucial task of relation and event extraction from biomedical research publications. They emphasize the significance of syntactic information in this process and aim to determine the most effective approaches to syntactic processing in the biomedical domain. The researchers conduct an empirical investigation comparing traditional feature-based models with neural network-based models for part-of-speech (POS) tagging and dependency parsing tasks using two widely recognized biomedical corpora, GENIA and CRAFT. Their analysis reveals that, overall, neural models outperform feature-based models on these benchmark datasets. This finding is particularly noteworthy as there has been a lack of recent comparative studies focusing on neural models in the context of biomedical text analysis. Furthermore, the study includes a task-oriented evaluation to assess how these parsing models impact downstream applications such as biomedical event extraction. Surprisingly, the results indicate that superior intrinsic parsing performance does not always translate to better extrinsic event extraction performance, highlighting the complexity of integrating syntactic processing into practical applications. In conclusion, Nguyen and Verspoor present a detailed empirical exploration of traditional and neural network-based models for POS tagging and dependency parsing in biomedicine. Their work sheds light on the importance of parser selection in optimizing performance for downstream tasks like event extraction. The retrained models from their study are made publicly available for further research at https://github.com/datquocnguyen/BioPosDep. This research, accepted for publication in BMC Bioinformatics, contributes valuable insights into enhancing information extraction processes from biomedical literature through advanced syntactic processing techniques.