Two-stage framework for identifying and extracting Vietnamese comparative opinion quintuple extraction

Comparative opinion mining is an important subtask of opinion mining. It aims to identify comparative reviews and extract the comparative elements in quintuples. This task, called Comparative Opinion Quintuple Extraction (COQE), has two main sub-tasks: Comparative Sentence Identification (CSI) and Comparative Element Extraction (CEE). In this paper, we introduce an effective two-stage framework for the COQE task specifically designed for the Vietnamese language. The first stage leverages the power of fine-tuning different BERT-based language models to identify the comparative sentences. We then formulate the comparison extraction task as a conditional text generation problem and apply a multi-task instruction prompting architecture based on generative language models. Furthermore, we also employ a data augmentation technique to increase the training data samples. Our experimental results on the VCOM dataset [15] demonstrate that our framework outperforms existing methods and achieves state-of-the-art performance on the test set. We also conduct a detailed analysis to provide insights for future research on this topic.