Context-Guided BERT for Targeted Aspect-Based Sentiment Analysis

Aspect-based sentiment analysis (ABSA) and Targeted ASBA (TABSA) allow finer-grained inferences about to be drawn from the same text, depending on context. For example, a given text can have different targets (e.g., neighborhoods) aspects price or safety), with associated each target-aspect pair. In this paper, we investigate whether adding context self-attention models improves performance (T)ABSA. We propose two variants of Context-Guided BERT (CG-BERT) that learn distribute attention under contexts. first adapt context-aware Transformer produce CG-BERT uses context-guided softmax-attention. Next, an improved Quasi-Attention model learns compositional supports subtractive attention. train both pretrained (T)ABSA datasets: SentiHood SemEval-2014 (Task 4). Both achieve new state-of-the-art results our QACG-BERT having best performance. Furthermore, provide analyses impact in proposed models. Our work provides more evidence for utility context-dependencies self-attention-based language context-based natural tasks.