Can Data Diversity Enhance Learning Generalization?

Proceedings of the 29th International Conference on Computational Linguistics, pages 4933–4945 October 12–17, 2022. 4933 Can Data Diversity Enhance Learning Generalization? Yu Yu* and Shahram Khadivi [] and Jia Xu* * School of Engineering and Science, Steven Institute of Technology, NJ 07030, USA [] eBay Inc., Aachen 52064, Germany yyu50@stevens.edu, skhadivi@ebay.com, jxu70@stevens.edu Abstract This paper introduces our Diversity Advanced Actor-Critic reinforcement learning (A2C) framework (DAAC) to improve the generaliza- tion and accuracy of Natural Language Process- ing (NLP). We show that the diversification of training samples alleviates overfitting and im- proves model generalization and accuracy. We quantify diversity on a set of samples using the max dispersion, convex hull volume, and graph entropy based on sentence embeddings in high- dimensional metric space. We also introduce A2C to select such a diversified training sub- set efficiently. Our experiments achieve up to +23.8 accuracy increase (38.0% relatively) in sentiment analysis, -44.7 perplexity decrease (37.9% relatively) in language modeling, and consistent improvements in named entity recog- nition over various domains. In particular, our method outperforms both domain adaptation and generalization baselines without using any target domain knowledge. 1 Introduction We introduce the Diversity Advanced Actor-Critic reinforcement learning framework (DAAC) to im- prove the generalization and accuracy of Natural Language Processing (NLP). Training data plays a crucial role in Deep Learning (DL)-based NLP. Investigations (Hendrycks et al., 2020; Dodge et al., 2020; Ramponi and Plank, 2020) show that training data quality is imperative to the machine learning model’s performance. Good data are effective and easy to generalize, while bad data bring noise and overfits on out-of-domain test sets. Decades of work (van der Wees et al., 2017; Aharoni and Goldberg, 2020; Guo et al., 2020; Ax- elrod et al., 2011) have been devoted to finding the best data, with the most commonly used method we call domain adaptation (Qu et al., 2019; Liu et al., 2019) that uses target domain knowledge. Despite large improvements on specific domains, these adapted models lack robustness and are prone to rare events. For example, when models shift to- wards one target domain like Twitter, they may become erroneous for other domains like medical or travel, and real-world applications on news re- port data, for example, do not know which queries to receive before they launch, resulting in a criti- cal performance gap between laboratory findings and reality. Therefore, we move away from the independent and identically distributed assumption (i.i.d.) and focus on reducing generalization errors by studying the inter-dependencies among samples. In this paper, we ask three questions (RQs) and propose our answers. The first question (RQ 1) is: “does higher diversity among training sam- ples lead to better generalized NLP model learn- ing?”. Semantic diversity is a desirable aspect in human annotation, where training data sets should be large and diverse enough to learn the many ways the objects differ (Shankar et al., 2017; Wu et al., 2018; Merler et al., 2019; Schumann et al., 2021), like intrinsic facial diversity in face recognition data, since every face …

Content truncated to 500 words