Bilingual Zero-Shot Stance Detection

Zero-shot stance detection (ZSSD) aims to determine whether the author of a text is in support, against, or neutral toward a target that is unseen during training. In this paper, we investigate ZSSD within a bilingual framework and compare it with cross-lingual and monolingual scenarios, in settings that have not previously been explored. Our study focuses on both noun-phrase and claim targets within indomain and out-of-domain bilingual ZSSD scenarios. To support this research, we assemble Bi-STANCE, a comprehensive bilingual ZSSD dataset consisting of over 100,000 annotated text-target pairs in both Chinese and English, sourced from existing datasets. Additionally, we examine a more challenging aspect of bilingual ZSSD by focusing on claim targets with a low occurrence of shared words with their corresponding texts. As part of Bi-STANCE, we created an extended dataset that emphasizes this challenging scenario. To the best of our knowledge, we are the first to explore this difficult ZSSD setting. We investigate these tasks using state-of-the-art pre-trained language models (PLMs) and large language models (LLMs). We release our dataset and code at https://github.com/chenyez/BiSTANCE.