Hierarchical Adversarial Training for Multi-domain Adaptive Sentiment Analysis

Zhao Xu, Lorenzo von Ritter, Giuseppe Serra

Abstract: Extracting useful insights with sentiment analysis is of increasing importance due to the growing availability of user-generated content. Sentiment analysis usually involves multiple different domains, and the labeled data is often difficult to obtain. In this paper we propose a hierarchical adversarial neural network (HANN) for adaptiveXu, Zhao sentiment analysis. Unlike most existing deep learning based methods, the proposed method HANN is able to share information between multiple domains bidirectionally, not justVon Ritter, Lorenzo transfers information from source domain to target domain in one direction only. In particular, the HANN method is inspired by the ideas of hierarchical Bayesian modeling and generative adversarial networks. We introduce each domain a distinct encoder to model the domain-specific distribution of the latent features. The learning procedures onSerra, Giuseppe different domains are coupled by a discriminator network to propagate the information, which can be viewed as adversarial networks in a supervised context by forcing the discriminator to identify domain labels. The proposed method HANN not only captures the distinct properties of each domain, but also shares common information across multiple domains. We demonstrate the superior performance of our method on real data including the Amazon review dataset and the Sanders Twitter sentiment dataset.