Subjectivity-Aware Aspect-Based Sentiment Analysis using BERT: Improving Performance through Objective Sentence Filtering

This study investigates the impact of incorporating subjectivity detection as a preprocessing step in aspect-based sentiment analysis (ABSA) using the BERT model.The proposed approach, named Sub-BERT, filters out objective sentences from the dataset before performing ABSA, allowing the model to focus on classifying the sentiment polarity of subjective sentences.Experiments are conducted on two datasets, SemEval14 and MEMD, covering various domains such as laptops, restaurants, books, clothing, and hotels.The results demonstrate that Sub-BERT outperforms the baseline BERT model, achieving improvements of up to 22% in accuracy and 28% in F1 score for the SemEval14 dataset, and up to 5% in accuracy and 33% in F1 score for the MEMD dataset.The findings highlight the importance of subjectivity detection in enhancing the performance of ABSA, particularly in scenarios with limited training data.However, the study's limitations, such as the reliance on manual subjectivity filtering and the impact of dataset imbalance, should be considered when interpreting the results.Future research directions include exploring techniques for addressing dataset imbalance and developing a unified framework for end-to-end ABSA with subjectivity detection.