Transforming Aspect-based Sentiment Analysis to Domain-Specific Mining by Integrating Machine Learning Ensemble (MLE) Model

Aspect-based Sentiment Analysis (ABSA) extract sentiment with respect to a particular aspect of the given text. An ensemble is a combination of multiple models to build a composite model that is more reliable and accurate. We propose a novel technique that integrates Machine Learning Ensemble (MLE) approaches to allow ABSA to be adapted for domain-specific mining. We collect and annotate domain-specific data on Indian Budget, with an aspect label and sentiment polarity. The presented model utilizes numerous machine learning models as base learners and concatenated the predictions using MLE approaches in combination with Long Short-Term Memory (LSTM) and Bidirectional Encoder Representations from Transformers (BERT) as a meta-learner to improve the overall performance. We cross-validate and hyper-parameter tune the entire ensemble. Our MLE model’s accuracy, precision, recall and f1 score are recorded as $\mathbf{9 2 \%}, \mathbf{9 4 \%}, \mathbf{8 7 \%}$, and 0.93 respectively, which is undoubtedly higher than other baseline models. The result we received, indicates an improvement in terms of accuracy and applicability. This research has proven to be a promising approach for understanding domain-specific sentiments.