Adversarial Category Alignment Network for Cross-domain Sentiment Classification

Proceedings of NAACL-HLT 2019, pages 2496–2508 Minneapolis, Minnesota, June 2 - June 7, 2019. c⃝2019 Association for Computational Linguistics 2496 Adversarial Category Alignment Network for Cross-domain Sentiment Classification Xiaoye Qu1∗ Zhikang Zou1∗ Yu Cheng2 Yang Yang3 Pan Zhou1† 1Huazhong University of Science and Technology 2Microsoft AI & Research 3University of Electronic Science and Technology of China {xiaoye, panzhou}@hust.edu.cn | yu.cheng@microsoft.com {zhikangzou001, dlyyang}@gmail.com Abstract Cross-domain sentiment classification aims to predict sentiment polarity on a target domain utilizing a classifier learned from a source domain. Most existing adversarial learning methods focus on aligning the global marginal distribution by fooling a domain discrimina- tor, without taking category-specific decision boundaries into consideration, which can lead to the mismatch of category-level features. In this work, we propose an adversarial category alignment network (ACAN), which attempts to enhance category consistency between the source domain and the target domain. Specif- ically, we increase the discrepancy of two po- larity classifiers to provide diverse views, lo- cating ambiguous features near the decision boundaries. Then the generator learns to create better features away from the category bound- aries by minimizing this discrepancy. Exper- imental results on benchmark datasets show that the proposed method can achieve state- of-the-art performance and produce more dis- criminative features. 1 Introduction Sentiment classification aims to automatically identify the sentiment polarity (i.e., positive or negative) of the textual data. It has attracted a surge of attention due to its widespread ap- plications, ranging from movie reviews to prod- uct recommendations. Recently, deep learning- based methods have been proposed to learn good representations and achieved remarkable success. However, the performances of these works are highly dependent on manually annotated training data while annotation process is time-consuming and expensive. Thus, cross-domain sentiment classification, which aims to transfer knowledge learned on labeled data from related domains ∗Equal contribution † Corresponding author (called source domain) to a new domain (called target domain), becomes a promising direction. One key challenge of cross-domain sentiment classification is that the expression of emotional tendency usually varies across domains. For in- stance, considering reviews about two sorts of products: Kitchen and Electronics. One set of reviews would contain opinion words such as “de- licious” or “tasty”, and the other “rubbery” or “blurry”, to name but a few. Due to the small in- tersection of two domain words, it remains a sig- nificant challenge to bridge the two domains diver- gence effectively. Researchers have developed many algorithms for cross-domain sentiment classification in the past. Traditional pivot-based works (Blitzer et al., 2007; Yu and Jiang, 2016) attempt to infer the correlation between pivot words, i.e., the domain- shared sentiment words, and non-pivot words, i.e., the domain-specific sentiment words by utilizing multiple pivot prediction tasks. However, these methods share a major limitation that manual se- lection of pivots is required before adaptation. Recently, several approaches (Sun et al., 2016; Zellinger et al., 2017) focus on learning domain invariant features whose distribution is similar in source and target domain. They attempt to mini- mize the discrepancy between domain-specific …

Content truncated to 500 words