Modern ML systems ingest data aggregated from diverse sources, such as synthetic, human-annotated, and live customer traffic. Understanding which examples are important to the performance of a learning algorithm is crucial for efficient model training. Recently, a growing body of literature has given rise to various “influence scores,” which use training artifacts such as model confidence or checkpointed gradients to identify important subsets of data. However, these methods have primarily been developed in computer vision settings, and it remains unclear how well they generalize to language-based tasks using pretrained models. In this paper, we explore the applicability of influence scores in language classification tasks. We evaluate a diverse subset of these scores on the SNLI dataset by quantifying accuracy changes in response to pruning training data through random and influence-score-based sampling. We then stress-test one of the scores – “variance of gradients” (VoG) from Agarwal and Hooker (2022) – in an NLU model stack that was exposed to dynamic user speech patterns in a voice assistant type of setting. Our experiments demonstrate that in many cases, encoder-based language models can be fine-tuned on roughly 50% of the original data without degradation in performance metrics. Along the way, we summarize lessons learned from applying out-of-the-box implementations of influence scores, quantify the effects of noisy and class-imbalanced data, and offer recommendations on score-based sampling for better accuracy and training efficiency.

Manually labeled training data is expensive, noisy, and often scarce, such as when developing new features or localizing existing features for a new region. In cases where labeled data is limited but unlabeled data is abundant, semi-supervised learning methods such as consistency training can be used to improve model performance, by training models to output consistent predictions between original and augmented versions of unlabeled data. In this work, we explore different data augmentation methods for consistency training (CT) on Natural Language Understanding (NLU) domain classification (DC) in the limited labeled data regime. We explore three types of augmentation techniques (human paraphrasing, back-translation, and dropout) for unlabeled data and train DC models to jointly minimize both the supervised loss and the consistency loss on unlabeled data. Our results demonstrate that DC models trained with CT methods and dropout based augmentation on only 0.1% (2,998 instances) of labeled data with the remainder as unlabeled can achieve a top-1 relative accuracy reduction of 12.25% compared to fully supervised model trained with 100% of labeled data, outperforming fully supervised models trained on 10x that amount of labeled data. The dropout-based augmentation achieves similar performance compare to back-translation based augmentation with much less computational resources. This paves the way for applications of using large scale unlabeled data for semi-supervised learning in production NLU systems.