Proceedings of the 19th SIGBioMed Workshop on Biomedical Language Processing

Dina Demner-Fushman, Kevin Bretonnel Cohen, Sophia Ananiadou, Junichi TsujiiACL 2020

TL;DR

This document comprises the proceedings of the 19th SIGBioMed Workshop on Biomedical Language Processing, held virtually on July 9, 2020. It showcases cutting-edge research in applying natural language processing (NLP) techniques to the biomedical domain, with a significant portion of the work influenced by the COVID-19 pandemic. The workshop highlights advancements in information retrieval, clinical language processing, language understanding, and named entity recognition, demonstrating the growing importance of NLP in accelerating biomedical discovery and healthcare.

What Problem Does This Paper Address?

The SIGBioMed workshop addresses the critical need to advance natural language processing (NLP) techniques specifically for the biomedical domain. This includes challenges like extracting complex information from vast amounts of scientific literature and clinical notes, improving the accuracy and efficiency of biomedical search, and developing models that can understand and process specialized medical language. The context of the COVID-19 pandemic further amplified the urgency, highlighting the need for rapid analysis of research and clinical data to combat global health crises. The workshop also grapples with the inherent difficulties in biomedical NLP, such as the scarcity of labeled data, the complexity of medical terminology, and the need for reproducible research in a rapidly evolving field.

Methods

The papers presented at the workshop explore a diverse range of NLP methods, including:

  • Information Retrieval (IR): Techniques like linguistic knowledge enhancement, dependency structure encoding with word embeddings, and sequence-to-set semantic tagging for complex query reformulation are employed to improve search capabilities in biomedical literature.
  • Clinical Language Processing: This area focuses on tasks such as visual dialog for radiology, clinical temporal relation extraction, automatic clinical coding, and text classification on electronic health records. Methods include BERT-based models, reinforcement learning, and the development of silver-standard datasets.
  • Language Understanding: Research in this area includes building datasets for biomedical machine reading comprehension (e.g., BioMRC), developing models for evidence inference, and applying NLP to understand complex medical conditions like Alzheimer's disease through text analysis.
  • Named Entity Recognition (NER) and Knowledge Representation: This involves analyzing errors in NER, developing multi-task learning approaches for relation extraction, and creating benchmarks for knowledge graph embeddings in the biomedical domain.

Datasets

Several datasets are utilized and discussed in the proceedings:

  • CORD-19: A corpus of work on the SARS-CoV-2 virus and COVID-19 disease, noted for its inclusion of "gray literature" and flexibility in licensing.
  • MIMIC-III and MIMIC-CXR: Widely used datasets for clinical NLP tasks, including clinical coding and radiology image-text analysis. The workshop also discusses the development of silver-standard datasets for these tasks.
  • Evidence Inference Dataset: Expanded upon in this work to include more data and models.
  • Biomedical Literature: Various corpora of scientific publications are used for tasks like relation extraction, gender bias quantification, and knowledge graph embedding.
  • President Ronald Reagan's Speeches: Used for detecting early stages of Alzheimer's disease.
  • Veterinary Clinical Notes: Utilized for disease syndrome classification.

Key Insights and Ideas

  • COVID-19's Impact: The pandemic significantly accelerated the relevance and application of NLP in biomedical research, particularly in analyzing large-scale datasets like CORD-19.
  • Reproducibility in Biomedical NLP: There's a growing emphasis on studies of reproducibility and replicability, acknowledging a "reproducibility crisis" that is not sparing the biomedical NLP field.
  • Advancements in Clinical NLP: The use of large clinical datasets like MIMIC-III and MIMIC-CXR is driving progress in tasks like temporal relation extraction, clinical coding, and patient outcome prediction.
  • The Power of Pre-trained Models: BERT and other pre-trained language models are proving highly effective across various biomedical NLP tasks, from relation extraction to question answering.
  • Bridging the Gap to Precision Medicine: Research is increasingly focused on leveraging NLP to extract actionable knowledge from complex biomedical data, paving the way for more personalized and effective healthcare.
  • Innovative Presentation Formats: Due to the virtual nature of the workshop, new, more engaging presentation formats (panel-like sessions) were explored to foster interaction among researchers.

Key Results

  • The workshop accepted 73 submissions, with 22 papers (approximately 30%) selected for presentation in panel-like sessions.
  • Specific quantitative results are detailed within individual papers, covering improvements in accuracy for various NLP tasks, the creation of new benchmarks, and the evaluation of different model configurations. For example, some papers report significant gains in relation extraction accuracy or improved performance on clinical question answering tasks.

Limitations and Future Work

  • Virtual Conference Challenges: The cancellation of the physical workshop due to the pandemic meant that not all high-quality submissions could be presented as posters, limiting the breadth of in-person interaction.
  • Data Scarcity: While efforts are being made to create larger and more diverse datasets (e.g., CORD-19, expanded Evidence Inference), the scarcity of labeled data remains a challenge for many biomedical NLP tasks.
  • Reproducibility: The ongoing focus on reproducibility suggests that ensuring consistent and verifiable results across different research groups is a key area for future work.
  • Adapting to Evolving Needs: The rapid pace of biomedical research and the emergence of new health challenges (like pandemics) necessitate continuous adaptation and development of NLP tools.
  • Improving User Experience: Future work aims to develop systems that better approximate user experience, particularly in complex tasks like medical entity recognition.

Summary updated

BibTeX
@inproceedings{demnerfushman2020proceedings,
  author = {Dina Demner-Fushman and Kevin Bretonnel Cohen and Sophia Ananiadou and Junichi Tsujii},
  title = {{Proceedings of the 19th SIGBioMed Workshop on Biomedical Language Processing}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2020},
}
Figure