Lexical Normalization of User-Generated Medical Text

Anne Dirkson, Suzan Verberne, Wessel KraaijSMM4H @ ACL 2019 · DOI · OpenReview

Abstract: In the medical domain, user-generated social media text is increasingly used as a valuable complementary knowledge source to scientific medical literature. The extraction of this knowledge is complicated by colloquial language use and misspellings. Yet, lexical normalization of such data has not been addressed properly. This paper presents an unsupervised, data-driven spelling correction module for medical social media. Our method outperforms state-of-the-art spelling correction and can detect mistakes with an $F_{0.5}$ of 0.888. Additionally, we present a novel corpus for spelling mistake detection and correction on a medical patient forum.

Introduction

Mining medical insights from social media—such as tracking adverse drug reactions or public health trends—is a high-value task hindered by linguistic "noise." User-generated text is often riddled with misspellings, colloquialisms, and domain-specific abbreviations. Existing lexical normalization tools, typically built for clinical records or generic social media like Twitter, often fail to bridge this gap. Clinical tools miss informal "layman" language, while Twitter-based tools lack specialized medical vocabulary (e.g., complex drug names like Gleevec). To address this, we present an unsupervised, data-driven spelling correction module specifically designed for medical patient forums.

Methodology

The proposed pipeline is purely data-driven and unsupervised, avoiding reliance on static medical dictionaries that quickly become obsolete. The core components include:

  1. Spelling Mistake Detection: The system identifies errors based on corpus frequency and word similarity, determining if a rare word has a much more common, similar-looking counterpart.
  2. Relative Weighted Edit Distance (RWE): This core innovation improves upon standard Levenshtein distance by weighting character changes based on common typo patterns (using the Norvig corpus) and employing a relative measure to handle varying word lengths and deletions/additions.
  3. Conservative Correction: Prioritizing precision ($F_{0.5}$), the system is tuned to avoid "correcting" rare but valid medical terms into common English words. In medical contexts, a false correction is often more damaging than a missed typo.
Decision logic for the unsupervised spelling correction module.

Data

The study utilized several key data sources to evaluate performance and generalizability:

  • GIST Forum: Approximately 1.25M tokens from a Gastrointestinal Stromal Tumor patient forum.
  • Reddit: Approximately 4.5M tokens from cancer-related subreddits used to test generalizability.
  • Annotated Medical Corpus: A novel benchmark consisting of 500 manually annotated posts from the GIST forum—the first of its kind for medical forum data.
  • External Validation: Six public Twitter datasets (e.g., SMM4H 2019) were used to assess improvements in downstream machine learning tasks like sentiment analysis.

Results and Discussion

The RWE method achieved 62.3% accuracy, a significant leap over the 20.8% achieved by previous state-of-the-art social media baselines. Mistake detection reached an $F_{0.5}$ of 0.888 with a precision of 0.90, proving the system can effectively identify typos without flagging valid medical jargon.

Reduction in Out-of-Vocabulary (OOV) terms across GIST and Reddit datasets compared to baseline.

The "domain gap" is evident: generic tools perform poorly because medical communities have unique linguistic signatures. By deriving candidates from the corpus itself, the system avoids "Out-of-Vocabulary" (OOV) traps and learns specific vocabularies (e.g., mutation types or hospital names) without human intervention. While the reduction in OOV terms (0.50% in GIST, 0.27% on Reddit) may appear small, it represents a significant portion of the actual spelling errors present in the text.

Conclusion

This work demonstrates that unsupervised, data-driven normalization is superior for niche medical domains where dictionaries are often incomplete. Future work will focus on addressing "real-word" errors (e.g., typing "dose" instead of "does") and evaluating the impact of normalization on Named Entity Recognition (NER) tasks for identifying specific drugs and symptoms.

Summary updated

Abstract

In the medical domain, user-generated social media text is increasingly used as a valuable complementary knowledge source to scientific medical literature. The extraction of this knowledge is complicated by colloquial language use and misspellings. Yet, lexical normalization of such data has not been addressed properly. This paper presents an unsupervised, data-driven spelling correction module for medical social media. Our method outperforms state-of-the-art spelling correction and can detect mistakes with an F0.5 of 0.888. Additionally, we present a novel corpus for spelling mistake detection and correction on a medical patient forum.
BibTeX
@inproceedings{dirkson2019lexical,
  author = {Anne Dirkson and Suzan Verberne and Wessel Kraaij},
  title = {{Lexical Normalization of User-Generated Medical Text}},
  booktitle = {SMM4H @ ACL},
  year = {2019},
  doi = {10.18653/v1/w19-3202},
}
Figure