Jie Lin, Chao Wang, Qi Li
ABSTRACT Mining comparative opinions from user‐generated content (UGC) provides crucial preliminary insights into competitive relationships among products. However, in real‐world UGC scenarios, accurate identification remains challenging due to highly informal language, the low prevalence of comparative opinions in natural corpora leading to severe class imbalance, and the difficulty of fixed decision thresholds in accommodating fluctuating prediction confidences. In this study, we propose a BERT‐based unified framework enhanced with focal loss and a dynamic thresholding mechanism to effectively identify comparative opinions from UGC. The framework integrates three key components. (1) A context‐aware pre‐training strategy to better adapt to the linguistic characteristics of UGC. (2) A focal loss function to address severe class imbalance. (3) A dynamic threshold adjustment mechanism to refine classification decision boundaries. Experimental results on real‐world review datasets demonstrate that the proposed method outperforms baseline models in both accuracy and robustness, offering a practical solution for large‐scale comparative opinion detection in noisy online environments. In addition, it provides enterprises with a data‐driven tool for competitive analysis.