Bidirectional generative prompt learning for aspect sentiment triplet extraction

Guangjin Wang, Wenya Wu, Fuyong Xu, Ru Wang, Zhenfang Zhu, Peiyu Liu, Jinlong Wang

Abstract: Generative methods based on prompt learning achieve competitive results on aspect sentiment triplet extraction (ASTE) task by injecting task-specific knowledge. However, existing methods ignore the decoupled learning of triplet representations, resulting in mutual interference between different forms of knowledge and thus affecting the generation of diverse results. To this end, we propose a bidirectional generative prompt learning (BGPL) framework that disentangles the learning of complementary representations to obtain experts with unique triplet knowledge to generate high-quality and diverse triplets. Specifically, we construct different representations of triplet as prompts to specialize aspect-to-opinion direction and opinion-to-aspect direction learning. With these two directions learning complementary triplet knowledge, high-quality and diverse results can be generated more comprehensively. To reduce the noise introduced by integrating the predictions, an ensemble strategy based on generation probability is utilized to select candidate triplets. Furthermore, we eliminate potentially conflicting triplets by balancing the generation probability of terms. Extensive experiments on four benchmark datasets with different data ratio settings demonstrate the effectiveness of the proposed BGPL framework.

External IDs:dblp:journals/mlc/WangWXWZLW25