Domain-specific or Uncertainty-aware models: Does it really make a difference for biomedical text classification?

Abstract: The success of pretrained LLMs (PLMs) across a spate of use-cases has led to significant investment from the NLP community towards building domain-specific foundational models. On the other hand, in mission critical settings such as biomedical applications, other aspects also factor in-chief of which is a model's ability to produce reasonable estimates of its own uncertainty. In the present study, we discuss these two desiderata through the lens of how they shape the entropy of a model's output probability distribution. We find that domain specificity and uncertainty awareness can often be successfully combined, but the exact task at hand weighs in much more strongly.

• The paper demonstrates that domain-specific models consistently outperform general-domain counterparts in classification accuracy, emphasizing the value of specialized pretraining.
• The study employs Bayesian Neural Networks with DropConnect to integrate uncertainty awareness, enhancing prediction reliability through improved Brier Score and calibration metrics.
• The findings reveal that the best model configuration is task-dependent, highlighting the need for careful calibration between domain-specificity and uncertainty-awareness in biomedical NLP.

The Significance of Domain-Specificity and Uncertainty-Awareness in Biomedical NLP Models

The paper entitled "Domain-specific or Uncertainty-aware models: Does it really make a difference for biomedical text classification?" explores the intersection of two critical aspects in the deployment of NLP models for biomedical applications: domain-specificity and uncertainty-awareness. The authors questioned whether these aspects could be harmoniously combined to improve the efficacy, and particularly the reliability, of biomedical text classification models.

Introduction and Motivation

Deep learning models optimized for high prediction accuracy are often constrained by their domain limitations and susceptibility to biases. While domain-specific models have been developed to counteract these issues, particularly in specialized fields like biomedicine, they often neglect the element of uncertainty, which is pivotal in mission-critical applications. This paper primarily investigates the compatibility of domain-specific pretraining with uncertainty-aware modeling to offer insights into improving model robustness and credibility in biomedical contexts.

Methodology

The authors employed six standard biomedical datasets, three in English and three in French, covering various medical tasks from predicting patient conditions based on medical abstracts (MedABS) to determining drug prescription intent from user speech transcriptions (PxSLU). The datasets varied in their class imbalance ratios and textual lengths, allowing for a comprehensive evaluation of model performance under different conditions.

Four types of models were compared across these datasets:

1. General-domain, uncertainty-unaware (denoted $-\mathcal{D}-\mathcal{U}$)
2. General-domain, uncertainty-aware (denoted $-\mathcal{D}+\mathcal{U}$)
3. Domain-specific, uncertainty-unaware (denoted $+\mathcal{D}-\mathcal{U}$)
4. Domain-specific, uncertainty-aware (denoted $+\mathcal{D}+\mathcal{U}$)

For general-domain models, the authors used BERT and CamemBERT for English and French datasets, respectively. Domain-specific models were derived from BioBERT for English and CamemBERT-bio for French, which are specifically pretrained on datasets germane to the biomedical field. Bayesian Neural Networks (BNNs) with DropConnect as the primary Bayesian method facilitated the uncertainty-aware architectures.

Results and Discussion

Classification Performance

The results showcased that domain-specific models ($+\mathcal{D}$) consistently outperformed their general counterparts ($-\mathcal{D}$) in terms of Macro-F1 and accuracy across all datasets. This aligns with previous findings that domain-specific pretraining yields better semantic understanding and relevance for specialized tasks. Notably, the $+\mathcal{D}-\mathcal{U}$ configuration generally rendered the highest classification performance, indicating that domain-specificity remains a primary driver of prediction accuracy.

Uncertainty Metrics

Uncertainty-aware models exhibited better scores for uncertainty quantification metrics such as Brier Score (BS), Expected Calibration Error (ECE), and Negative Log-Likelihood (NLL). While domain-specific uncertainty-aware models often registered higher-than-average scores for these metrics, intriguing observations emerged from the entropy evaluation:

The entropy scores revealed that domain-specific models, both uncertainty-aware and unaware, led to lower entropy, indicating higher confidence in their predictions. However, when these models were incorrect, entropy varied more significantly, especially in uncertainty-aware configurations.

Implications and Future Work

The balance between domain-specificity and uncertainty-awareness is task-dependent. SHAP attributions clarified that dataset-specific characteristics heavily influenced the model performance, occasionally resulting in general-domain and uncertainty-unaware models performing adequately for some tasks. This variability indicates that biomedical practitioners should calibrate their model choice not only based on whether it is domain-specific or uncertainty-aware but also on the specifics of the task at hand.

The authors ignored to find one-size-fits-all solutions and highlighted the importance of blended strategies that consider task intricacies for optimal model performance. Future work could focus on fine-tuning the interactions between domain-specific and uncertainty-aware elements and extending this approach to other domain-specific applications beyond biomedicine.

Conclusion

This paper provides a nuanced understanding of the contributions of domain-specificity and uncertainty-awareness to model performance in the biomedical NLP domain. While domain-specificity predominates in achieving higher classification performance, uncertainty-awareness contributes significantly to model reliability. No universally superior configuration emerges due to the substantial influence of task-specific factors. The interplay between domain-specific pretraining and uncertainty-aware design merits careful consideration in the development of dependable biomedical NLP models.