Enhancing Zero-shot and Few-shot Stance Detection with Commonsense Knowledge Graph

Abstract

In this paper, we consider a realistic scenario on stance detection with more application potential, i.e., zero-shot and few-shot stance detection, which identifies stances for a wide range of topics with no or very few training examples. Conventional data-driven approaches are not applicable to the above zero-shot and few-shot scenarios. For human beings, commonsense knowledge is a crucial element of understanding and reasoning. In the absence of annotated data and cryptic expression of users' stance, we believe that introducing commonsense relational knowledge as support for reasoning can further improve the generalization and reasoning ability of the model in the zero-shot and few-shot scenarios. Specifically, we introduce a commonsense knowledge enhanced model to exploit both the structurallevel and semantic-level information of the relational knowledge. Extensive experiments demonstrate that our model outperforms the state-of-the-art methods on zero-shot and fewshot stance detection task.

Zero-Shot Stance Detection via Contrastive Learning

“…[2] adapted a target-specific stance detection dataset [27] to ZSSD, and deployed adversarial learning to extract target-invariant transformation features in ZSSD. Further, to exploit both the structural-level and semantic-level information of the relational knowledge, [24] proposed a commonsense knowledge enhanced graph model based on BERT [9] to cope with ZSSD.…”

“…Table 4: Experimental results on three ZSSD datasets. The results with ♮ are retrieved from [1], with † are retrieved from [24], with ‡ are retrieved from [2], with ♯ are retrieved from [8], with ♭ are retrieved from [23]…”

“…and CKE-Net [24]. We also compare with a prompt-based method exploited in stance detection: Pattern-Exploiting Training (PET) [32].…”

“…To deal with ZSSD, some existing methods attempt to introduce attention mechanisms [1] or external knowledge to capture relationships between targets [24], in order to generalize to unseen targets. However, in practice, directly transferring stance features from seen targets to unseen ones may not lead to good results since there might be features specific to certain targets.…”

Few-Shot Stance Detection via Target-Aware Prompt Distillation

“…Zhang et al [45] proposed a semantic-emotion knowledge transferring model for cross-target stance detection, which used external knowledge as a bridge to enable knowledge transfer across different targets. Liu et al [19] introduced a commonsense knowledge enhanced model to exploit both the structural-level and semantic-level information of the relational knowledge. Besides, Zhang et al [46] leveraged multiple external knowledge bases as bridges to explicitly link potentially opinioned terms in texts to targets of interest.…”