BLAR: Biomedical Local Acronym Resolver

William Hogan, Yoshiki Vazquez Baeza, Yannis Katsis, Tyler Baldwin, Ho-Cheol Kim, Chun-Nan HsuNAACL 2021

TL;DR

This paper introduces BLAR, a novel transformer-based model for resolving biomedical acronyms within their local context. By leveraging a two-step architecture and pre-trained language models, BLAR achieves state-of-the-art performance, significantly outperforming previous methods on the Ab3P benchmark.

What Problem Does This Paper Address?

The biomedical field is awash in text, and extracting knowledge from it is crucial. A major hurdle in this process is the pervasive use of acronyms, which are abbreviations for long, complex terms like diseases, bacteria, and chemicals. While global acronym resolution (using external dictionaries) has seen progress, local acronym resolution – identifying the expansion of an acronym within the same document – has lagged behind. Existing methods are often rule-based or rely on older machine learning techniques, failing to benefit from recent advancements like transformers. This paper aims to bridge this gap by applying modern transformer architectures to the local acronym resolution problem.

Methods

BLAR employs a two-step approach:

  1. Sentence Classification: The first step identifies sentences that are likely to contain an acronym definition (a pair of short-form and long-form entities). This is framed as a binary classification task.
  2. SF-LF Acronym Tagging: For sentences flagged in Step 1, this step identifies the exact spans of the short-form (SF) and long-form (LF) entities. This is treated as a token classification task, using a BILOU (Beginning, Inside, Last, Outside, Unit) tagging scheme.

Both steps leverage transfer learning by fine-tuning the pre-trained SciBERT model. A key aspect is that the tagging model is trained on the output of the sentence classification model, mitigating potential "exposure bias" from using perfectly labeled inputs. The authors also explored a "single-step" BLAR model that attempts to detect and resolve acronyms simultaneously, serving as an ablation study.

Datasets

  • BioADI Corpus: Used for training BLAR. It contains 1,668 true SF-LF pairs from 1,200 annotated PubMed abstracts.
  • Ab3P Corpus: Used for testing BLAR. It contains 1,221 true SF-LF pairs from 1,250 annotated PubMed abstracts.

The paper notes that both datasets provide span-level data for SF-LF pairs and are available for download.

Key Insights and Ideas

  • Transformers are powerful for local AR: The paper demonstrates that modern transformer architectures, specifically SciBERT, can be effectively adapted to the local acronym resolution task, achieving superior performance.
  • A two-step approach is beneficial: The proposed two-step architecture (sentence classification followed by tagging) significantly outperforms the single-step approach, highlighting the importance of first identifying relevant sentences.
  • Local AR has been stagnant: The authors point out that significant progress in local acronym resolution has been limited since around 2009, with recent work focusing more on disambiguation. BLAR revitalizes this sub-field.
  • Handling diverse acronym structures: BLAR shows promise in resolving acronyms where the number of letters doesn't match the number of words, and even complex acronyms that mix short-form entities with complete words (e.g., "HIT type II").

Key Results

  • BLAR (two-step) achieved an F1 score of 0.966 on the Ab3P corpus.
  • This significantly outperforms the next best performing approach, which achieved an F1 score of 0.899 (Kuo et al., 2009).
  • The single-step BLAR model achieved an F1 score of 0.953, demonstrating the benefit of the two-step architecture.
  • The paper provides a detailed table comparing BLAR's performance (Precision, Recall, F1) against several previous SOTA local AR models.

Limitations and Future Work

  • Modified Acronyms: A majority of BLAR's errors fall into the "modified acronyms" category, where the model might incorrectly expand or truncate the long-form. This often involves adding or missing a single word/token.
  • Complex Long-Forms: The model sometimes struggles with very complex long-forms, predicting a simpler expansion when a more intricate one is correct.
  • Future Work: The authors plan to:
  • Further improve the model based on their error analysis.
  • Explore extending BLAR to cover acronyms found in domains other than scientific literature.
  • Investigate hybrid models that combine deep learning with rule-based algorithms.

Summary updated

BibTeX
@inproceedings{hogan2021blar,
  author = {William Hogan and Yoshiki Vazquez Baeza and Yannis Katsis and Tyler Baldwin and Ho-Cheol Kim and Chun-Nan Hsu},
  title = {{BLAR: Biomedical Local Acronym Resolver}},
  booktitle = {Conference of the North American Chapter of the Association for Computational Linguistics},
  year = {2021},
}
Figure