Effects of Human Adversarial and Affable Samples on BERT Generalization

Aparna Elangovan, Estrid He, Yuan Li, Karin Verspoor

Submission Type: Regular Long Paper

Submission Track: Efficient Methods for NLP

Submission Track 2: Machine Learning for NLP

Keywords: BERT, Generalization, Robustness in NLP

TL;DR: We investigate how various proportion of label preserving and label alternative samples impact robustness.

Abstract: BERT-based models have had strong performance on leaderboards, yet have been demonstrably worse in real-world settings requiring generalization. Limited quantities of training data is considered a key impediment to achieving generalizability in machine learning. In this paper, we examine the impact of training data quality, not quantity, on a model's generalizability. We consider two characteristics of training data: the portion of human-adversarial (h-adversarial) samples, i.e. sample pairs with seemingly minor differences but different ground-truth labels, and human-affable (h-affable) training samples, i.e. sample pairs with minor differences but the same ground-truth label. We find that for a fixed size of training samples, having 10-30% h-adversarial instances improves the precision, and therefore F_1, by up to 20 points in the tasks of text classification and relation extraction. Increasing h-adversarials beyond this range can result in performance plateaus or even degradation. In contrast, h-affables may not contribute to a model's generalizability and may even degrade generalization performance.

Submission Number: 419