Topic-Selective Graph Network for Topic-Focused Summarization

Due to the success of pre-trained language model (PLM), existing PLM-based summarization models show their powerful generative capability. However, these are trained on general-purpose datasets, leading generated summaries failing satisfy needs different readers. To generate with topics, many efforts have been made topic-focused summarization. works a summary only guided by prompt comprising topic words. Despite success, methods still ignore disturbance sentences non-relevant topics and conduct cross-interaction between tokens attention module. address this issue, we propose topic-arc recognition objective topic-selective graph network. First, is used training, which endows capability discriminate for model. Moreover, network can topic-guided based results recognition. In experiments, extensive evaluations NEWTS COVIDET datasets. Results that our achieve state-of-the-art performance.