CS Net: A Coarse-to-Fine-Grained Summarization Network for Community-Based Question Answering Summarization

Community-based Question Answering (CQA) summarization refers to the task of automatically generating concise and informative summaries of question-answer interactions in community question answering platforms. It is essential to consider all the lengthy QA pairs without ignoring the valuable information in summarization. However, previous works directly summarized input QA pairs, leading to the omission of some effective information without the process of coarse-to-fine-grained information extraction in the summary. Meanwhile, the existing methods utilize a cross-attention mechanism to calculate the scores of all QA pairs, overlooking the noise information present in QA pairs. In this paper, a Coarse-to-fine-grained Summarization Network (CS Net) is proposed to make CQA summarization. The core modules in CS Net include an unsupervised similarity-based pair-level selector and a seed-QA constrained margin loss calculator where the former filters similar QA pairs from the coarse-grained perspective, and the latter makes training pay more attention to effective information for summarization from the fine-grained perspective. Experimental results show that CS Net achieves state-of-the-art performance in CQA summarization.