Hierarchical attention for aspect extraction using LSTM in fine-grained sentiment analysis and evaluation

In the current digital world, aspect-based sentiment analysis (ABSA) from a review sentence is a big challenge. In recent works, the aspect term detection from the reviews of the discourse is performed independently, even for contextually interdependent reviews. Further, the baseline techniques could not correlate valid multi-word aspect extraction from different review sentences, which further degrades the aspect extraction performance. Therefore, this paper aims to improve the performance of the aspect term detection subtask using a supervised hierarchical attention-based model. A sentence alignment step is added as preprocessing before ABSA on multisentence reviews. This step removes deficient knowledge, captures aspect alignment from different review sentences, and properly relates the aspect to the context domain. Next, the context information is used to extract contextual aspects using word-vector representation. The input sentences and labeled data using sequence tagging are supplied to the proposed supervised hierarchical attention network. The proposed model can capture the semantics of words for different domains, enhancing the extraction of domain-specific relevant aspect terms. The experimental results shown in the SemEval-16 reviews present the effectiveness of the proposed attentionbased approach for aspect term detection. The F-score is improved by approximately 3 % as compared to the recent hybrid unsupervised approach for both the laptop and restaurant domains when coreference resolution is not performed. On the other side, F-score is improved by approximately 4% and 3 % as compared to the recent hybrid unsupervised approach for the laptop and restaurant domains, respectively, when coreference resolution is not performed. When the results of the proposed approach are compared with those of recent supervised approaches, the F-score of the proposed model is improved by 5 % and 9 % for laptops and restaurants, respectively, when coreference resolution is not performed. Whereas the F-score of the proposed model is improved by 9% and 12 % for laptops and restaurants, respectively, when coreference resolution is performed before the aspect extraction subtask.