This paper considers Aspect-based Opinion Summarization (AOS) of reviews onparticular products. To enable real applications, an AOS system needs toaddress two core subtasks, aspect extraction and sentiment classification. Mostexisting approaches to aspect extraction, which use linguistic analysis ortopic modeling, are general across different products but not precise enough orsuitable for particular products. Instead we take a less general but moreprecise scheme, directly mapping each review sentence into pre-defined aspects.To tackle aspect mapping and sentiment classification, we propose twoConvolutional Neural Network (CNN) based methods, cascaded CNN and multitaskCNN. Cascaded CNN contains two levels of convolutional networks. Multiple CNNsat level 1 deal with aspect mapping task, and a single CNN at level 2 dealswith sentiment classification. Multitask CNN also contains multiple aspect CNNsand a sentiment CNN, but different networks share the same word embeddings.Experimental results indicate that both cascaded and multitask CNNs outperformSVM-based methods by large margins. Multitask CNN generally performs betterthan cascaded CNN.
展开▼