Testing the Assumptions of Active Learning for Translation Tasks with Few Samples

TL;DR

This paper investigates why active learning strategies underperform in translation tasks with few samples, revealing that core assumptions about informativeness and diversity do not correlate with test performance.

Contribution

The study challenges existing assumptions of active learning by showing factors like sample ordering and pre-training interactions are more influential in low-sample scenarios.

Findings

Informativeness and diversity are not correlated with test performance in few-sample settings.

Sample ordering and pre-training interactions significantly impact active learning effectiveness.

Current AL strategies need to incorporate factors beyond informativeness and diversity for better performance.

Abstract

Active learning (AL) is a training paradigm for selecting unlabeled samples for annotation to improve model performance on a test set, which is useful when only a limited number of samples can be annotated. These algorithms often work by optimizing for the informativeness and diversity of the training data to be annotated. Recent work found that AL strategies fail to outperform random sampling on various language generation tasks when using 100-500 samples. To understand AL's poor performance when only using few samples, we investigate whether the core assumptions underlying AL strategies hold. We find that neither the informativeness nor diversity of the training data, which AL strategies optimize for, are correlated with test set performance. Instead, factors like the ordering of the training samples and interactions with pre-training data have a larger impact on performance. This suggests that future AL methods must take these factors into account in order to work with very few samples.