Stance detection aims to capture users’ attitudes towards specific topics on social media at a fine-grained level. Recently, several two-stage topic-stance extraction methods have been proposed to address the challenge of stance detection without prior topic knowledge. However, existing methods still have two significant limitations. The first is the neighboring dependency, where nearby words receive higher attention. The second is the error propagation between the two stages. To address these issues, we propose a syntax-aware generative framework for topic-stance extraction. Specifically, at the topic extraction stage, we develop a dual-channel topic generation model, including a semantic and syntactic graph encoding. This model can “pull” words that are semantically related but located far away from topic words using syntactic information, thus effectively alleviating the problem of neighboring dependency. Moreover, to mitigate the error propagation, we propose a multi-task joint unlikelihood training strategy at the second stage, which includes stance detection and topic modification. The erroneously generated topics from the first stage are treated as negative samples. By the unlikelihood learning strategy, we minimize the probability of these negative samples to reduce topic error propagation. Empirical results on three popular datasets demonstrate that our method significantly outperforms state-of-the-art methods.