A Multi-task Shared Cascade Learning for Aspect Sentiment Triplet Extraction Using BERT-MRC

The aspect sentiment triplet extraction (Triplet) aims at extracting aspect terms (AE), extracting aspect-oriented opinion terms (AOE), and discriminating aspect-level sentiment polarity (ASC) from the comments. To address the current study, the end-to-end framework-based approach suffers from the problem of contribution distribution among multiple components, while the pipeline framework-based approach is susceptible to error propagation. Moreover, the complexity of the model limits the detection of long-distance aspect terms and opinion terms. In this paper, we propose a framework based on multi-task shared cascade learning and machine reading comprehension (MRC), which is called Triple-MRC. The multi-task shared cascade learning can effectively avoid the problem of contribution distribution among components. The MRC approach leverages the prior knowledge from the question to reduce the error propagation between tasks and mitigate the limitation associated with model complexity. We conduct experiments on publicly available two benchmark datasets for the Triplet task. The experimental results demonstrate the superior performance of the Triple-MRC framework compared to the baseline model, which better achieves the Triplet task. Through the analysis of the comparison study, model training process, error analysis, ablation study, attention visualization, and case study, we have confirmed the effectiveness of introducing the multi-task shared cascade learning method and MRC method into the model.