Moka-ADA: adversarial domain adaptation with model-oriented knowledge adaptation for cross-domain sentiment analysis

Cross-domain sentiment analysis (CDSA) aims to overcome domain discrepancy to judge the sentiment polarity of the target domain lacking labeled data. Recent research has focused on using domain adaptation approaches to address such domain migration problems. Among them, adversarial learning performs domain distribution alignment via domain confusion to transfer domain-invariant knowledge. However, this method that transforms feature representations to be domain-invariant tends to align only the marginal distribution, and may inevitably distort the original feature representations containing discriminative knowledge, thus making the conditional distribution inconsistent. To alleviate this problem, we propose adversarial domain adaptation with model-oriented knowledge adaptation (Moka-ADA) for the CDSA task. We adopt the adversarial discriminative domain adaptation (ADDA) framework to learn domain-invariant knowledge for marginal distribution alignment, based on which knowledge adaptation is conducted between the source and target models for conditional distribution alignment. Specifically, we design a dual structure with similarity constraints on intermediate feature representations and final classification probabilities, so that the target model in training learns discriminative knowledge from the trained source model. Experimental results on a publicly available sentiment analysis dataset show that our method achieves new state-of-the-art performance.