Transformers and Attention: Decoding and Understanding of Aspect-based Opinions in User-Generated Contents

Aspect-based opinion mining has become a significant information extraction technique based on natural language processing, driven by the growing volume of online user-generated content. This approach aims to determine the opinion polarity of specific aspects within a given context. The existing models primarily target explicit aspects, often neglecting the identification of implicitly mentioned aspect-based opinion polarity. Consequently, these existing models result in low classification accuracy and struggle to identify multiple aspects within a specific context. This paper proposes an aspect-based attention model (AAM) to address these limitations. We integrate a pre-trained BERT model with an attention mechanism to perform aspect detection. The AAM model is trained and evaluated on the benchmark SemEval-2014 Task 4 dataset. Experimental results demonstrate that the proposed AAM model outdoes other existing methods. Additionally, the robustness and generalizability of the proposed model are calculated using raw textual datasets.