Opinion-Aware Knowledge Embedding for Stance Detection

Zhenhui Xu, Qiang Li, Wei Chen, Yingbao Cui, Zhen Qiu, Tengjiao Wang

Abstract: As an emerging text classification task, stance detection is much helpful in reviewing subjective text and mining expressed attitudes of a person or organization towards an object. Due to the similarity with other text classification tasks, stance detection is always tackled by conventional classification methods. However, there is a big difference between stance detection and others, since stance detection depends much on human background knowledge while others do not. Therefore, to address such a unique problem, we propose a novel method, which leverages knowledge graph and incorporates text-mentioned knowledge with a deep classifier, by a key component named Opinion-aware Knowledge Embedding (OKE). The proposed OKE can integrate the objective knowledge facts and subjective text opinion well by a customized and effective attention mechanism. Our experiments also show that the proposed method comprehensively outperforms all the baselines on real datasets.