BanglaBioMed: A Biomedical Named-Entity Annotated Corpus for Bangla (Bengali)

Salim SazzedACL 2022

TL;DR

This paper introduces BanglaBioMed, the first biomedical named-entity annotated corpus for the Bengali language. By addressing the scarcity of resources for low-resource languages like Bengali, this work aims to significantly advance biomedical NLP research and applications in this domain.

What Problem Does This Paper Address?

Named Entity Recognition (NER) is a crucial task in biomedical and health science research, enabling downstream applications like entity linking and relation extraction. While English and other major languages boast extensive biomedical NER resources, Bengali, a low-resource language, has been severely lacking. This deficit hinders the development of NLP tools for understanding Bengali health-related texts, a growing need with the rise of telemedicine and digital health information in the language. The authors aim to fill this gap by creating a comprehensive, annotated corpus for Bengali biomedical entities.

Methods

The core of this work is the creation of the BanglaBioMed corpus. The authors followed a standard IOB2 format for annotation, a widely accepted method for sequence labeling tasks. The process involved:

  • Data Collection: Biomedical texts were collected from health articles published on the official website of the popular Bangladeshi daily newspaper, Prothom Alo.
  • Preprocessing: The collected texts were segmented into sentences using the Bengali full stop delimiter, and then tokenized into words and punctuation.
  • Manual Annotation: The corpus was manually annotated at the token level by an annotator with a university-level education. The annotation focused on four predefined biomedical entity types: Anatomy (AN), Chemical and Drugs (CD), Disease and Symptoms (DS), and Medical Procedure (MP).
  • Entity Tagging: The standard IOB2 format was employed:
  • B- for the beginning of an entity.
  • I- for tokens within an already initiated entity.
  • O for tokens not part of any entity of interest.
  • Handling Complex Entities: The paper acknowledges and illustrates complex entities, including overlapping and nested entities. For overlapping entities, the longer entity was designated as the "top-level" entity, and its sub-parts as "nested" entities. Co-referential or anaphoric references, as well as intensifiers or possessive adjectives, were excluded from annotation to maintain consistency.

Datasets

The paper introduces a newly created dataset:

  • BanglaBioMed: This is the primary contribution. It consists of over 12,000 tokens, with approximately 2,000 tokens annotated with four types of biomedical entities (AN, CD, DS, MP) and around 10,000 non-entity tokens. The corpus is made publicly available for researchers.

Key Insights and Ideas

  • Bridging the Low-Resource Gap: The most significant insight is the successful creation of a valuable resource for a severely under-resourced language in a critical domain. This opens doors for research that was previously infeasible.
  • Importance of Manual Annotation: The paper emphasizes the necessity of meticulous manual annotation by domain experts (or those with relevant education) to ensure the quality and accuracy of the corpus, especially for nuanced biomedical terms.
  • Dominance of Disease and Symptom Entities: The corpus statistics reveal that Disease and Symptom (DS) entities are the most prevalent, which is expected given the nature of health articles. This highlights the focus area for future biomedical NER systems in Bengali.
  • Complexity of Biomedical Text: The acknowledgment and handling of complex and nested entities underscore the inherent challenges in biomedical text annotation, requiring careful guideline development.

Key Results

The paper reports on the creation and characteristics of the BanglaBioMed corpus:

  • The corpus contains approximately 12,000 tokens in total.
  • Around 2,000 tokens are annotated as biomedical named entities.
  • The four entity types are Anatomy (AN), Chemical and Drugs (CD), Disease and Symptoms (DS), and Medical Procedure (MP).
  • The Disease and Symptom (DS) category is the most dominant, accounting for about 60% of the annotated entities.
  • The Chemical and Drugs (CD) category has the lowest presence.
  • Most entities are single words, with longer entities primarily belonging to the DS category.
  • The corpus is available in the standard IOB2 format, making it compatible with existing NER frameworks.

Limitations and Future Work

  • Corpus Size: While a significant contribution, the authors acknowledge that the corpus size is moderate. Future work will focus on enhancing its size.
  • Transformer-Based Models: The authors plan to investigate the use of transformer-based language models to create strong baselines for automatic NER tasks on this corpus.
  • Cross-Lingual Transfer: An interesting future direction is exploring how to leverage cross-lingual resources from languages like English to improve the performance of Bengali NER systems.
  • Annotation Consistency: While efforts were made for consistency, the inherent subjectivity in annotation can always be a point for further refinement.

Summary updated

BibTeX
@inproceedings{sazzed2022banglabiomed,
  author = {Salim Sazzed},
  title = {{BanglaBioMed: A Biomedical Named-Entity Annotated Corpus for Bangla (Bengali)}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2022},
}
Figure