Data Boost: Text Data Augmentation Through Reinforcement Learning Guided Conditional Generation

Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing, pages 9031–9041, November 16–20, 2020. c⃝2020 Association for Computational Linguistics 9031 Data Boost: Text Data Augmentation Through Reinforcement Learning Guided Conditional Generation Ruibo Liu1, Guangxuan Xu2, Chenyan Jia3, Weicheng Ma4, Lili Wang5, and Soroush Vosoughi6 1,2,4,5,6Department of Computer Science, Dartmouth College 3Moody College of Communication, University of Texas at Austin 1,4,5{ruibo.liu.gr,weicheng.ma.gr,lili.wang.gr}@dartmouth.edu 2guangxuan.xu.ug@dartmouth.edu 3chenyanjia@utexas.edu 6soroush.vosoughi@dartmouth.edu Abstract Data augmentation is proven to be effective in many NLU tasks, especially for those suf- fering from data scarcity. In this paper, we present a powerful and easy to deploy text augmentation framework, Data Boost, which augments data through reinforcement learning guided conditional generation. We evaluate Data Boost on three diverse text classification tasks under five different classifier architec- tures. The result shows that Data Boost can boost the performance of classifiers especially in low-resource data scenarios. For instance, Data Boost improves F1 for the three tasks by 8.7% on average when given only 10% of the whole data for training. We also compare Data Boost with six prior text augmentation meth- ods. Through human evaluations (N=178), we confirm that Data Boost augmentation has comparable quality as the original data with re- spect to readability and class consistency. 1 Introduction Data augmentation is a widely-used technique in classification tasks. In the field of computer vi- sion (CV), data is augmented by flipping, cropping, tilting, and altering RGB channels of the original images (Krizhevsky et al., 2012; Chatfield et al., 2014; Szegedy et al., 2015); however, similar in- tuitive and simple strategies do not obtain equal success in NLP tasks. Existing methods tend to produce augmentation with low readability or un- satisfying semantic consistency (Yang et al., 2020). Table 1 shows some output samples of popu- lar text augmentation methods. Naive methods imitate pixel manipulation in CV, augmenting sen- tences by adding spelling errors (Xie et al., 2017), or randomly deleting and swapping tokens (Wei and Zou, 2019). The output of such augmentation methods are often illegible since the word order is disrupted (e.g., “is The baby very!”); even worse, Original So Cute! The baby is very lovely! Naive Aug. Delete + Swap So Cute! is The baby very! Word2Vec Aug. Insert + Replace So Cute adorable! The baby is very fabulous! Back Translate Aug. Eng. → Fr. → Eng. Cute! The baby is very cute! Data Boost Look at this adorable baby! He is so cute! Table 1: A simple demo of existing text data augmenta- tion methods on positive sentiment label. crucial feature words (e.g., the word lovely which is a signal-carrying word for sentiment detection) could be mistakenly removed through random dele- tion. A more advanced method is synonym in- sertion or replacement (Zhang et al., 2015; Wang and Yang, 2015), which uses Word2Vec (Mikolov et al., 2013) to replace words with their synonyms. Such a method respects the original sentence struc- ture but fails to consider the context. It some- times replaces words with synonyms that are awk- ward in …

Content truncated to 500 words