Aspect Extraction from E-Commerce Reviews Based on Word-Word Relationship Classification

Aspect-level sentiment analysis of e-commerce reviews helps companies to understand customers' feedback and needs. However, the current unsupervised and sequence-annotation-based supervised aspect extraction methods do not fully utilize contextual information and cannot solve the problem of the multi-sense word ambiguity of aspect words. To address this issue, a new scheme is proposed to model aspect extraction as word-word relationship classification (WWRAE). WWRAE architecture fully exploits textual information by effectively modeling the next-word relationship and the last-first-word relationship between aspects. A neural network framework is designed based on the WWRAE scheme, where aspect words are modeled as a two-dimensional word-word lattice. Additionally, a multi-granularity two-dimensional convolution is proposed to enhance the lattice representation, and joint prediction is used to fully reason about the word-word relations. Experiments were conducted on the Restaurants and Laptop of SemEval-2014 and Restaurants of SemEval-2015 datasets, and F1 values reached 86.93, 82.99, and 75.51, respectively, outperforming the current baseline model.