The task of Sentence-Level Sentiment Analysis (SLSA) aims to detect the general sentiment polarity of an entire sentence. Even though many Deep Neural Network (DNN) models have been proposed for SLSA, the task remains very challenging because the efficacy of these deep models depends on the i.i.d (Independent and Identically Distributed) assumption; but, in real scenarios, the distributions of training and target data are almost certainly different to some extent.
To alleviate this limitation resulting from distribution misalignment, this paper proposes a supervised approach based on the non-i.i.d paradigm of Gradual Machine Learning (GML) for SLSA. Beginning with the labeled training instances, the proposed approach gradually labels target instances in the order of increasing hardness by iterative knowledge conveyance. It leverages DNNs for feature extraction to supervise gradual knowledge conveyance. Specifically, it trains a sentence-level polarity classifier, which can detect polarity similarity between close neighbors in a deep embedding space, and separately a binary semantic network, which can extract implicit polarity relations between two arbitrary instances. Then, it fulfills knowledge conveyance by modeling the detected relations as binary features in a factor graph. We have empirically evaluated the performance of the proposed approach on real benchmark workloads by a comparative study. Our extensive experiments show that it achieves the state-of-the-art performance across all the test workloads.