Proceedings of the 18th BioNLP Workshop and Shared Task

Dina Demner-Fushman, Kevin Bretonnel Cohen, Sophia Ananiadou, Junichi TsujiiACL 2019

TL;DR

This document presents the proceedings of the 18th BioNLP Workshop and Shared Task, highlighting advancements in biomedical natural language processing. It showcases research on diverse topics including clinical text analysis, information extraction, question answering, and the development of resources for multiple languages, with a strong emphasis on deep learning models like BERT.

What Problem Does This Paper Address?

The field of biomedical natural language processing (BioNLP) has seen a significant surge in research, driven by the increasing availability of public tools, corpora, and community challenges. This workshop and its associated shared task aim to foster progress in this domain by providing a platform for researchers to present their work, share novel approaches, and address challenges in understanding and processing complex biomedical and clinical text. Specifically, the BioNLP-ST has evolved into an open platform supporting various challenges, and this year's event focuses on research directions in inference, entailment, information retrieval, and question answering within the medical domain. The workshop also emphasizes the importance of research in languages beyond English and the application of cutting-edge deep learning techniques to these specialized text types.

Methods

The papers presented at BioNLP 2019 explore a wide range of methods, with a dominant theme being the application of deep learning architectures. Many submissions leverage pre-trained language models, particularly BERT and its variants (e.g., BioBERT), often fine-tuned on biomedical corpora. Techniques discussed include:

  • Transformer-based models: BERT, BioBERT, and other transformer architectures are widely used for various tasks.
  • Embeddings: Contextualized word embeddings and their application in tasks like named entity recognition and relation extraction are prevalent.
  • Transfer Learning: Utilizing pre-trained models and adapting them to specific biomedical tasks is a common strategy.
  • Multi-task Learning: Several approaches combine multiple related tasks to improve overall performance.
  • Domain Adaptation: Methods for adapting general NLP models to the specialized language of the biomedical domain are explored.
  • Resource Development: Creation of new corpora and benchmarks for specific languages and tasks (e.g., MoNERO for Romanian, ChiMed for Chinese).

Datasets

The proceedings reference a variety of datasets, both established and newly introduced for specific tasks and shared challenges. Notable mentions include:

  • MEDIQA 2019 Shared Task: This task focused on textual inference, question entailment, and question answering in the medical domain.
  • MoNERO: A biomedical gold standard corpus for the Romanian language.
  • ChiMed: A Chinese medical corpus for question answering.
  • Various Biomedical Benchmarking Datasets: Used for evaluating transfer learning approaches.
  • Drug Labels and LactMed: Used for extraction of specific information.
  • Clinical Notes: A common data source for tasks like diagnosis, phenotyping, and information extraction.

Key Insights and Ideas

  • BERT's Dominance: The widespread adoption and success of BERT and its biomedical adaptations (like BioBERT) underscore their power in capturing complex linguistic nuances in specialized domains.
  • Multilingual BioNLP: There's a growing effort to extend BioNLP research to languages beyond English, highlighting the need for language-specific resources and models.
  • Deep Learning for Diverse Tasks: Deep learning models are proving effective across a broad spectrum of BioNLP tasks, from named entity recognition and relation extraction to question answering and clinical decision support.
  • The Importance of Benchmarks: The continued development and use of shared tasks and benchmark datasets are crucial for driving progress and enabling fair comparison of different approaches.
  • Bridging Clinical and Biological Text: Research continues to address both clinical text (patient records, notes) and biological text (research papers, literature), recognizing the interconnectedness of these domains.

Key Results

While specific quantitative results are detailed within individual papers, the overall trend indicates significant improvements in BioNLP tasks due to the application of advanced deep learning methods. The workshop highlights research that contributes to accelerating foundational biomedical research and improving clinical practice through enhanced information retrieval, extraction, question answering, and clinical decision support. The success of various teams in the MEDIQA 2019 Shared Task demonstrates the effectiveness of different approaches to textual inference and question answering in the medical domain.

Limitations and Future Work

  • Reviewer Workload: The organizers acknowledge the significant workload placed on reviewers, which, due to time constraints, led to only papers recommended by at least two reviewers being accepted. This suggests a potential bottleneck in the review process for highly competitive workshops.
  • Focus on English: While efforts are being made for other languages, a significant portion of the research still focuses on English biomedical text.
  • Future Directions: The continued exploration of deep learning architectures, the development of resources for under-resourced languages, and the application of BioNLP to real-world clinical decision support systems are likely areas for future work. The evolution of BioNLP-ST into an open platform suggests a future of sustained community-driven challenges.

Summary updated

BibTeX
@inproceedings{demnerfushman2019proceedings,
  author = {Dina Demner-Fushman and Kevin Bretonnel Cohen and Sophia Ananiadou and Junichi Tsujii},
  title = {{Proceedings of the 18th BioNLP Workshop and Shared Task}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2019},
}
Figure