Multitask Learning and BERT Embedding: A Comprehensive Approach to Subjectivity Detection and Aspect-Based Sentiment Analysis

The proliferation of online platforms has intensified the need for fine-grained sentiment analysis of customer reviews. While traditional Aspect-based Sentiment Analysis (ABSA) focuses on sentiment classification, it often overlooks the crucial distinction between subjective opinions and objective statements. This paper introduces the Subjectivity Multitask Aspect-based Sentiment Analysis framework, which integrates Subjectivity Detection as an auxiliary task with Aspect Sentiment Classification through a unified multitask learning architecture. The model leverages BERT embeddings, a Bidirectional LSTM, a self-attention mechanism, and a Neural Tensor Network to effectively model the interplay between subjectivity and sentiment. Crucially, our experimental results on the comprehensive MEMD multi-domain dataset demonstrate the framework’s effectiveness, achieving a state-of-the-art F1-score of 87.26%. This represents a significant 26.89 percentage point improvement over the previous baseline, showcasing superior performance in both precision-recall balance and cross-domain stability. Key contributions include: (1) manual annotation of aspect-level subjectivity labels for a large-scale multi-domain dataset, (2) development of an integrated multitask framework that addresses precision-recall imbalance in ABSA, and (3) empirical validation showing consistent improvements across diverse domains. Results confirm that joint optimization of related tasks enhances sentiment classification accuracy while maintaining computational efficiency.