부산대학교 도서관

The aspect sentiment quadruple extraction (ASQE) task aims to extract all the opinion quadruples (aspect term, aspect category, opinion term, sentiment polarity) from the review text. However, there are often implicit expressions in the review text, and implicit opinion quadruples are difficult to be represented and extracted. To this end, we propose a novel end-to-end approach for the explicit and implicit ASQE tasks. Specifically, we use a multi-scale convolutional neural network (MS-CNN) and bidirectional long short-term memory neural network (BiLSTM) to capture the local features and contextual features. Then we design a novel sentence-guided grid tagging scheme to extract explicit and implicit opinion quadruples contained in the reviews, in which the grid of the representation of the sentence's overall meaning is used to mark the implicit expression. Extensive experimental results indicate that our model outperforms strong baselines significantly and achieves state-of-the-art performance. [ABSTRACT FROM AUTHOR]