Aspect-Based Sentiment Analysis Through EDU-Level Attentions

A sentence may express sentiments on multiple aspects. When these aspects are associated with different sentiment polarities, a model’s accuracy is often adversely affected. We observe that in such hard sentences mostly expressed through clauses, or formally known as elementary discourse units (EDUs), and one EDU tends to single aspect unitary towards aspect. In this paper, we propose consider boundaries modeling, attentions at both word levels. Specifically, highlight sentiment-bearing words word-level sparse attention. Then level, force the model attend right for aspect, by using EDU-level attention orthogonal regularization. Experiments three benchmark datasets show our simple EDU-Attention outperforms state-of-the-art baselines. Because can be automatically segmented high accuracy, applied directly without need of manual boundary annotation.