Multi-target opinion words extraction: Multi-target opinion words extraction

Target-oriented Opinion Words Extraction (TOWE) emerges as a novel subtask of Aspect-Based Sentiment Analysis (ABSA), exerting profound impacts on numerous downstream tasks and real-world applications. Current research on target-oriented opinion words extraction focuses solely on extracting opinions for a single target. However, when handling multiple “target-opinion” pairs within a single data instance, inefficiency and time-consuming issues arise. Therefore, we propose an alternative solution to target-oriented opinion words extraction, termed Multi-Target Opinion Words Extraction (MTOWE). To achieve multi-target opinion words extraction, we conduct a statistical analysis and reconstruction of popular datasets(14lap, 14res, 15res, and 16res) in the target-oriented opinion words extraction task, denoted as M-14lap, M-14res, M-15res, and M-16res, respectively. We propose a Graph-based Collaborative Understanding Network for MTOWE. Experimental results demonstrate that the graph-based collaborative understanding network significantly outperforms the baseline models on the four datasets of the multi-target opinion words extraction task, achieving F1 scores of 74.04, 82.30, 74.92, and 83.49, respectively. More importantly, when processing the same batch of information, the inference time of MTOWE is significantly reduced. On the four datasets, the average inference time of MTOWE is only 70% of that of TOWE, while the F1 scores of MTOWE are only slightly lower than those of TOWE.