EMOVA: A Semi-supervised End-to-End Moving-Window Attentive Framework for Aspect Mining

Aspect mining or extraction is one of the most challenging problems in aspect-level analysis on customer reviews; it aims to extract terms from a review describing aspects of a reviewed entity, e.g., a product or service. As aspect mining can be formulated as the sequence labeling problem, supervised deep sequence learning models have recently achieved the best performance. However, these supervised models require a large amount of labeled data which are usually very costly or unavailable. To this end, we propose a semi-supervised End-to-end MOVing-window Attentive framework (called EMOVA) that has three key features for aspect mining. (1) Two neural layers with Bidirectional Long Short-Term Memory (BiLSTM) are employed to learn representations of reviews. (2) Cross-View Training (CVT) is used to improve the representation learning over a small set of labeled reviews and a large set of unlabeled reviews from the same domain in a unified end-to-end architecture. (3) Since past nearby information in a text provides important semantic contexts for a prediction task in aspect mining, a moving-window attention component is proposed in EMOVA to enhance prediction accuracy. Experimental results over four review datasets from the SemEval workshops show that EMOVA outperforms the state-of-the-art models for aspect mining.