Unsupervised Reinforcement Adaptation for Class-Imbalanced Text Classification

Proceedings of the 11th Joint Conference on Lexical and Computational Semantics, pages 311 - 322 July 14-15, 2022 ©2022 Association for Computational Linguistics Unsupervised Reinforcement Adaptation for Class-Imbalanced Text Classification Yuexin Wu University of Memphis ywu10@memphis.edu Xiaolei Huang University of Memphis xiaolei.huang@memphis.edu Abstract Class-imbalance naturally exists when train and test models in different domains. Unsu- pervised domain adaptation (UDA) augment model performance with only accessible anno- tations from the source domain and unlabeled data from the target domain. However, exist- ing state-of-the-art UDA models learn domain- invariant representations and evaluate primar- ily on class-balanced data across domains. In this work, we propose an unsupervised domain adaptation approach via reinforcement learning that jointly leverages feature variants and imbal- anced labels across domains. We experiment with the text classification task for its easily accessible datasets and compare the proposed method with five baselines. Experiments on three datasets prove that our proposed method can effectively learn robust domain-invariant representations and successfully adapt text clas- sifiers on imbalanced classes over domains. The code is available at https://github. com/woqingdoua/ImbalanceClass 1 Introduction Unsupervised domain adaptation (UDA) is to find a shared feature space that is predictive across target and source domains (Ramponi and Plank, 2020). The shared space, domain-independent feature set, allows transferring of trained text classifiers from the source domain to the target domain. Methods to find the space have two major directions, pivot feature (Blitzer et al., 2006; Daumé III, 2007; Ziser and Reichart, 2018; Ben-David et al., 2020a) and adversarial learning (Ganin and Lempitsky, 2015; Chen et al., 2020b; Du et al., 2020). The pivot- based method selects a subset of shared features, called pivots, which learn important cross-domain information to represent shared feature space. Ad- versarial learning approaches the shared feature space by reducing document features’ capability to distinguish source and target domains. The com- mon method to achieve this is Gradient Reversal Layer (GRL) (Ganin and Lempitsky, 2015) aiming to reduce domain-specific patterns. However, the UDA approaches primarily focus on feature shifts (𝐴���(𝐴���𝐴���𝐴���𝐴���𝐴���𝐴���𝐴���)! = 𝐴���(𝐴���𝐴���𝐴���𝐴���𝐴���𝐴���𝐴���)) while ignore possible class shifts (𝐴���(𝐴���𝐴���𝐴���𝐴���𝐴���𝐴���𝐴���)! = 𝐴���(𝐴���𝐴���𝐴���𝐴���𝐴���𝐴���𝐴���)) across do- mains. Class-imbalance naturally exists in data when label distributions across domains (Cui et al., 2017; Cheng et al., 2020) are different. Under the class- imbalanced scenario, the label distribution is imbal- anced across domains, and the label distributions in source and target domains are not the same. Given the widely used Amazon data (Ni et al., 2019) as an example, the Book reviews may have more positive reviews than negative reviews, and the Kitchen may have a lower ratio of negative reviews. However, evaluating unsupervised domain adaptation under the class-imbalanced scenario is under-examined than the ideal scenario of the class-balanced bench- mark. A wide evaluation benchmark of UDA for text classifiers is extracted from the Amazon re- view (Blitzer et al., 2006). The data has the same balanced-class distributions for both source and tar- get domains. Such a well-balanced label distribu- tion may make existing UDA models inapplicable to the real-world scenario, where class distributions can …

Content truncated to 500 words