Self-Attention Networks and Adaptive Support Vector Machine for aspect-level sentiment classification

Aspect-level sentiment classification aims to integrating the context predict polarity of aspect-specific in a text, which has been quite useful and popular, e.g., opinion survey products’ recommending e-commerce. Many recent studies exploit Long Short-Term Memory (LSTM) networks perform aspect-level classification, but limitation long-term dependencies is not solved well, so that semantic correlations between each two words text are ignored. In addition, traditional model adopts SoftMax function based on probability statistics as classifier, ignores words’ features space. Support Vector Machine (SVM) can fully use information characteristics, it appropriate make high-dimensional space, however, just considers maximum distance different classes similarities same classes. To address these defects, we propose two-stage novel architecture named Self Attention Networks Adaptive SVM (SAN-ASVM) for classification. first stage, order overcome dependencies, Multi-Heads (MHSA) mechanism applied extract relationships words; furthermore, 1-hop attention designed pay more some important related aspect-specific. second ASVM substitute effectively multi-classifications Extensive experiments SemEval2014, SemEval2016 Twitter datasets conducted, compared prove SAN-ASVM obtain better performance.