Adversarial dual decision-based model for event-related opinion sentence recognition

Event-related opinion sentences recognition aims to mine the valuable comments discussing about the specific event from the mass of microblogs comments. It’s difficult to label sufficient comments for a new microblogs event, and domain adaptation methods are suitable for this task, which can transfer the knowledge from the source domain(labeled event) to the target domain (unlabeled event). However, existing domain adaptation models do not make full use of the domain invariance to realize categories alignments of the two domains. Therefore, we propose a adversarial dual decision model for opinion sentences recognition, which takes advantage of the domain invariance that the opinion sentences are correlated with the event description. The proposed model has two generators to obtain different event-correlated features, and two classifiers to make dual decisions based on the two features. Through the adversarial process of generators and classifiers, the model completes the category feature alignment and realizes the opinion sentences recognition for the new event. The comparative experiment results show that our model outperforms the state-of-the-art models on six opinion sentences recognition tasks. In addition, we verify the effectiveness of each component of our model by ablation experiments.