Adversarial Training for a Hybrid Approach to Aspect-Based Sentiment Analysis

The increasing popularity of the Web has subsequently increased abundance reviews on products and services. Mining these for expressed sentiment is beneficial both companies consumers, as quality can be improved based this information. In paper, we consider state-of-the-art HAABSA++ algorithm aspect-based analysis tasked with identifying towards a given aspect in review sentences. Specifically, train neural network part using an adversarial network, novel machine learning training method where generator tries to fool classifier by generating highly realistic new samples, such robustness. This method, yet never its classical form applied analysis, found able considerably improve out-of-sample accuracy HAABSA++: SemEval 2015 dataset, was from 81.7% 82.5%, 2016 task, 84.4% 87.3%.