[Paper Review] BERT memorisation and pitfalls in low-resource scenarios.

Michael Tänzer, Sebastian Ruder+1 more|arXiv (Cornell University)|Apr 16, 2021

Topic Modeling24 references4 citations

TL;DR

This paper investigates BERT's memorisation and generalisation in low-resource and noisy settings, finding robustness to label noise but failure in few-shot and rare entity recognition. To address this, it proposes a novel BERT-based architecture with prototypical networks that significantly improves performance on low-resource named entity recognition tasks.

ABSTRACT

State-of-the-art pre-trained models have been shown to memorise facts and perform well with limited amounts of training data. To gain a better understanding of how these models learn, we study their generalisation and memorisation capabilities in noisy and low-resource scenarios. We find that the training of these models is almost unaffected by label noise and that it is possible to reach near-optimal performances even on extremely noisy datasets. Conversely, we also find that they completely fail when tested on low-resource tasks such as few-shot learning and rare entity recognition. To mitigate such limitations, we propose a novel architecture based on BERT and prototypical networks that improves performance in low-resource named entity recognition tasks.

Motivation & Objective

• To understand how pre-trained models like BERT generalise and memorise facts in low-resource and noisy training scenarios.
• To investigate the limitations of BERT in few-shot learning and rare entity recognition tasks.
• To address the failure of BERT in low-resource settings by proposing a novel architecture combining BERT with prototypical networks.
• To improve performance on named entity recognition when only limited or rare examples are available.

Proposed method

• The proposed model integrates BERT with a prototypical network to learn class prototypes from few-shot support sets.
• It uses BERT’s contextual representations as input to compute class prototypes via mean pooling over support samples.
• The model computes attention-based similarity between query embeddings and class prototypes for classification.
• A contrastive loss is applied during training to encourage separation between different class prototypes.
• The architecture is fine-tuned end-to-end on low-resource named entity recognition datasets with few-shot and rare entity examples.
• The method enables effective few-shot and zero-shot inference by leveraging prototypical embeddings derived from BERT representations.

Experimental results

Research questions

• RQ1 How does BERT perform under high label noise in low-resource settings?
• RQ2 Why does BERT fail in few-shot and rare entity recognition despite strong performance on standard benchmarks?
• RQ3 Can combining BERT with prototypical networks improve generalisation in low-resource named entity recognition?
• RQ4 To what extent does the proposed architecture mitigate memorisation issues in low-resource scenarios?

Key findings

• BERT exhibits strong robustness to label noise, achieving near-optimal performance even on extremely noisy datasets.
• Despite this robustness, BERT fails to generalise in few-shot and rare entity recognition tasks, indicating a critical limitation in low-resource settings.
• The proposed BERT-prototypical network architecture significantly improves performance on low-resource named entity recognition benchmarks.
• The model achieves state-of-the-art results in few-shot and rare entity recognition, demonstrating improved generalisation over standard fine-tuned BERT.