Cross-domain Sentiment Classification Based on Transfer Learning and Adversarial Network

摘要

Cross-domain sentiment classification have raised much attention in recent years. Due to the lack of large labeled corpus, it's very hard and usually fail to apply sentiment classification task in new domains even excellent deep learning models are used. To address the problem, we introduce a Shared Knowledge Learning and Transfer Model (SKLT) for the cross-domain sentiment classification task based on Transfer Learning and Adversarial Network. This SKLT model can extract the domain-independent shared knowledge through bi-GRU combined with adversarial network and redundant features penalty, and we transfer the knowledge extracted from multi source domains to the target domain with partial weight transfer. Experiments on multi domains of review dataset demonstrate that, the shared knowledge extracted from the SKLT model works well in the new domain task, and it can significantly outperform original methods with domain adaptation.