Improving Pretrained Language Model Fine-Tuning With Noise Stability Regularization.

Area of Science:

• Natural Language Processing (NLP)
• Machine Learning
• Deep Learning

Background:

• Pretrained language models (PLMs) have advanced NLP.
• Fine-tuning PLMs can lead to overfitting and poor generalizability due to model complexity and limited data.

Purpose of the Study:

• To introduce a novel fine-tuning framework, Layerwise Noise Stability Regularization (LNSR), to mitigate overfitting in PLMs.
• To enhance the generalizability and domain generalization capabilities of language models.

Main Methods:

• LNSR perturbs neural network inputs with Gaussian or in-manifold noise in the representation space.
• The method regularizes the output of each layer within the language model.
• Theoretical and experimental analyses validate the proposed approach.

Main Results:

• LNSR outperforms state-of-the-art methods including L2-SP, Mixout, FreeLB, and SMART.
• The framework demonstrates effectiveness on text classification and more challenging question-answering tasks.
• Empirical results show improved domain generalization abilities for language models.

Conclusions:

• LNSR is an effective fine-tuning strategy for improving the generalizability of PLMs.
• The method offers a robust solution to overfitting in NLP tasks.
• LNSR shows promise for enhancing model performance across diverse downstream applications.