Domain-Invariant Feature Distillation for Cross-Domain Sentiment Classification

Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing, pages 5559–5568, Hong Kong, China, November 3–7, 2019. c⃝2019 Association for Computational Linguistics 5559 Domain-Invariant Feature Distillation for Cross-Domain Sentiment Classification Mengting Hu1∗ Yike Wu1∗ Shiwan Zhao2† Honglei Guo2 Renhong Cheng1 Zhong Su2 1 Nankai University 2 IBM Research - China mthu@mail.nankai.edu.cn, wuyike@dbis.nankai.edu.cn, {zhaosw, guohl}@cn.ibm.com chengrh@nankai.edu.cn, suzhong@cn.ibm.com Abstract Cross-domain sentiment classification has drawn much attention in recent years. Most existing approaches focus on learning domain- invariant representations in both the source and target domains, while few of them pay attention to the domain-specific information. Despite the non-transferability of the domain- specific information, simultaneously learning domain-dependent representations can facili- tate the learning of domain-invariant represen- tations. In this paper, we focus on aspect- level cross-domain sentiment classification, and propose to distill the domain-invariant sentiment features with the help of an orthog- onal domain-dependent task, i.e. aspect de- tection, which is built on the aspects varying widely in different domains. We conduct ex- tensive experiments on three public datasets and the experimental results demonstrate the effectiveness of our method. 1 Introduction Sentiment classification based on deep learning methods has developed rapidly in recent years. While achieving outstanding performance, these methods always need large-scale datasets with sentiment polarity labels to train a robust senti- ment classifier. However, in most cases, large- scale labeled datasets are not available in prac- tice and manual annotation costs much. One of the solutions to this problem is cross-domain sen- timent classification, which aims to exploit the rich labeled data in one domain, i.e. source do- main, to help the sentiment analysis task in an- other domain lacking for or even without labeled data, i.e. target domain. The rationality of this so- lution is that the source domain and target domain ∗ Work performed while interning at IBM Research - China. †Corresponding author. The fried rice is amazing here. Surprisingly, Britney Spears is amazing. Source Domain: Target Domain: Figure 1: Example sentences from the source domain (restaurant) and the target domain (twitter) respec- tively. The sentiment expressions marked by solid lines are domain-invariant, while the aspect terms marked by dashed lines are domain-specific. share some domain-invariant knowledge that can be transferred across domains. Previous works on cross-domain sentiment classification mainly focus on learning the domain-invariant representations in both source and target domains, either based on manual fea- ture selection (Blitzer et al., 2006; Pan et al., 2010) or automatic representation learning (Glorot et al., 2011; Chen et al., 2012; Ganin and Lempit- sky, 2015; Li et al., 2017). The sentiment classi- fier, which makes decisions based on the domain- invariant features and receives the supervisory sig- nals from the source domain, can be also applied to the target domain. We can draw an empiri- cal conclusion: the better domain-invariant fea- tures the method obtains, the better performance it gains. However, few studies explore the usage of the domain-specific information, which is also helpful to the cross-domain …

Content truncated to 500 words