A Meta-Learning Approach for Few-Shot (Dis)Agreement Identification in Online Discussions

Anonymous

Abstract: Online discussions are abundant with different opinions for a common topic, and identifying agreement and disagreement between online posts enables many opinion mining applications. Realizing the increasing needs to analyze opinions for emergent new topics (e.g., from "mask mandate" to "COVID vaccination") that however tend to lack annotations, we present the first meta-learning approach for few-shot (dis)agreement identification on a new topic with few labeled instances. We further design a lexicon based regularization loss and propose domain-aware task augmentation for meta-training to enable the meta-learner to learn both domain-invariant cues and domain-specific expressions for (dis)agreement identification. Extensive experiments on two benchmark datasets and evaluation on three topic domains demonstrate the effectiveness of the meta-learning approach that consistently and noticeably outperforms the conventional transfer learning approach based on fine-tuning.