Data Augmentation for Clinical Multilingual Information Extraction

Technical Summary: Data Augmentation for Clinical Multilingual Information Extraction

Problem Statement
Clinical Natural Language Processing (NLP) is critical for health informatics applications such as cohort identification, pharmacovigilance, and clinical decision support. However, the deployment of robust models across different institutions and languages is severely constrained by the scarcity of annotated clinical text. While Pre-trained Language Models (PLMs) have advanced the state-of-the-art, they require task-specific annotated datasets for fine-tuning, which are costly and resource-intensive to create manually. This bottleneck is particularly acute in multilingual contexts, where high-quality resources are often available only in English, leaving other languages with significant performance gaps. Existing data augmentation strategies often rely on computationally expensive generative models or require curated synonym lexicons, limiting their scalability and adaptability to new languages and domains.

Methodology
The authors propose a low-cost, computationally efficient data augmentation strategy based on neighbor word replacement using word embeddings. The method utilizes pre-trained Word2Vec (W2V) and FastText (FT) embeddings to identify semantically similar words for replacement within entity mentions.

The approach differs slightly for the two target tasks:

• Named Entity Recognition (NER): The algorithm operates sentence-by-sentence. It identifies entity mentions within a sentence and randomly selects one word from each mention (excluding structural words) to replace with its closest neighbor ($v_1$) based on cosine similarity, provided the similarity score exceeds a predefined threshold ($\tau$). This process is repeated $t$ times (augmentation factor) to generate variations. To ensure diversity, the algorithm cycles through words in a mention and, if all words are replaced before $t$ iterations are complete, it proceeds to the next closest neighbor ($v_2$, etc.).
• Entity Linking (EL): The algorithm groups entity mentions by their associated codes. For each code, it attempts to augment each mention $t$ times. Unlike NER, which replaces a word with a single closest neighbor, the EL augmentation utilizes all neighbor words with similarity scores above the threshold $\tau$ (up to the top five) to generate a wider variety of augmentations per round.

The study evaluated this method across five languages (Spanish, English, French, Italian, Portuguese) using three multilingual benchmarks: SympTEMIST (symptom recognition and normalization), CANTEMIST (tumor morphology), and MultiCardioNER (disease and drug recognition). The authors compared the performance of PLMs fine-tuned on non-augmented data versus data augmented with W2V and FT embeddings, utilizing various similarity thresholds ($0.75$ to $0.90$).

Key Contributions

1. A Lightweight Augmentation Algorithm: A simple, scalable method for multilingual clinical NER and EL that does not require curated synonym lexicons or large generative models, making it adaptable to resource-constrained settings.
2. Extensive Multilingual Evaluation: A comprehensive assessment across three major clinical benchmarks spanning five languages, providing a broad view of the method's efficacy in diverse linguistic contexts.
3. Embedding Model Analysis: A comparative analysis of Word2Vec and FastText embeddings, revealing distinct strengths for NER versus EL tasks.

Results

• NER Performance: The augmentation strategy generally improved NER performance, with F1-score gains ranging from 0.75 to 5.54 percentage points.
  • The most significant improvement was observed in the Italian MultiDrug task, where the F1-score increased by 5.54 points.
  • A new state-of-the-art result was achieved for English drug recognition in the MultiCardioNER benchmark, surpassing the previous best score by 0.22 points.
  • Word2Vec generally outperformed FastText for NER. The authors attribute this to W2V's ability to generate more diverse synonyms (e.g., replacing "kidney" with "liver" in a functional context) compared to FastText, which often produces morphological variations (e.g., changing singular to plural forms) that offer less training diversity.
• EL Performance: Gains for Entity Linking were more modest but generally positive in four out of five languages, with accuracy improvements reaching up to 0.77 points.
  • FastText outperformed Word2Vec for EL. The authors suggest that while W2V's diverse substitutions benefit NER, FastText's subword-level modeling better preserves concept-specific meanings, which is critical for accurate entity linking.
  • Lower similarity thresholds ($0.75, 0.80$) consistently hindered EL performance, while higher thresholds ($0.85, 0.90$) yielded marginal improvements.

Significance and Claims
The paper concludes that word-embedding-based neighbor replacement offers a simple and computationally efficient approach to increase training diversity for multilingual clinical NLP. The method is particularly effective for strengthening NER performance in low-resource settings, where the scarcity of labeled data is a primary bottleneck.

For Entity Linking, the authors claim the method provides modest improvements but highlight its limitations: the high semantic precision required for concept normalization means that simple word replacement can introduce inaccuracies if the substituted words do not map to the exact same clinical concept. Consequently, the authors suggest that future augmentation strategies for EL should emphasize concept preservation, potentially through controlled vocabularies or ontology-informed constraints.

Overall, the work supports the development of more scalable health informatics pipelines by enabling the extraction of information from clinical documents even when labeled data is limited, without the heavy computational overhead of generative models.