Introduction
Fine-tuning large pre-trained language models (PLMs) like BERT on limited datasets—a common scenario in few-shot learning—often leads to overfitting, where models memorize training examples rather than learning generalizable features. While Data Augmentation (DA) is a standard remedy, its effectiveness is often hampered by quality control issues. Static augmentation methods treat all synthetic samples as equally beneficial throughout training. However, this paper argues that the utility of an augmented sample is dynamic; a sample that aids an early-stage model might be detrimental to a more mature one. To address this, we present DynSelector, a framework that dynamically selects optimal augmentation samples based on the model's current learning state.
Methods
DynSelector employs a multi-faceted approach to manage data selection:
- Two-Stage Strategy: Training begins with "in-sample" augmented data (e.g., synonym replacement of gold labels) to provide initial regularization. Upon convergence, the model transitions to "out-of-sample" data (e.g., harvested from Wikipedia) to enhance generalizability.
- Curriculum Learning (CL) Filter: A moving confidence threshold is used to filter noisy labels. The model initially selects only high-confidence samples, gradually lowering the threshold as its expertise grows.
- Influence Function (IF) Selection: This core component calculates an "influence score" for each potential sample, estimating its impact on validation loss. Samples with negative scores (indicating they reduce loss) are retained.
- Computational Optimization: To make influence function calculations feasible, the Sherman-Morrison formula is used to update the inverse Hessian matrix incrementally, avoiding costly full inversions.
Key Insights: Data Effectiveness as a Moving Target
A critical finding of this research is that the "oracle effectiveness" of a specific data sample is not constant. As shown in Figure 1, the percentage of samples deemed effective fluctuates significantly across training epochs. This volatility underscores the necessity of dynamic selection over static, one-time filtering.
Furthermore, the success of DynSelector over first-order methods like BERT-GradSim suggests that second-order information (the Hessian matrix) provides a more nuanced and accurate representation of how data influences model generalization.
Experimental Results
The framework was evaluated across several benchmarks, including Stance Detection (SD), Chinese Sentiment Analysis (CNSA), and the GLUE benchmark.
Performance Comparison
DynSelector consistently outperformed strong baselines. In Stance Detection, it achieved a score of .667, a significant improvement over the standard BERT baseline of .555.
Ablation Studies and Few-Shot Robustness
Ablation tests revealed that removing the Influence Function (IF) component led to an 11-point drop in Stance Detection performance, highlighting its role as the primary driver of success. Additionally, the method demonstrated extreme robustness in low-resource settings, maintaining effectiveness even when the training set was reduced to a single instance.
Conclusion and Future Work
DynSelector demonstrates that "when" data is introduced is as vital as "what" data is used. By leveraging influence functions and curriculum learning, the framework provides a robust solution for few-shot text classification. Future research will explore simultaneous mixing of data sources and extending the framework to architectures beyond BERT. While computationally more intensive than standard training, the performance gains justify the overhead in resource-constrained scenarios.
Summary updated