【摘要】 方面级抽象式文本摘要（Aspect-based Abstract Summarization,ABAS）是一项旨在为特定用户定制关注特定方面摘要的具有挑战性的新任务。该文提出了互补注意力记忆（Complementary Attentional Memory,CoAM）方法，通过记忆机制增强ABAS任务中的方面-上下文交互建模。该文将CoAM与摘要模型BART集成，实现特定方面与上下文特征更好的聚合，生成更高质量的摘要。在多个现有数据集上的实验结果表明，CoAM模型优于现有的包括大模型在内的基线模型，并具有跨领域的鲁棒泛化能力。为了检验CoAM模型在不同语言环境下的效果，该文构建了中文方面级抽象式文本摘要数据集CABAS，并在该数据集上进行了人工标注和模型评估，以推动中文精细化方面级文本摘要的发展。更多还原

【Abstract】 Aspect-based Abstractive Summarization(ABAS)is a challenging task aiming to summarize specific aspects for specific users.This paper proposes Complementary Attentional Memory(CoAM),a new memory mechanism to enhance aspect-context interactive modeling.It combines the state-of-the-art summarization model,BART,with CoAM so as to better aggregate aspect-specific features into the context and generate higher-quality summaries.The experimental results on various datasets,including a self-constructed Chinese Aspect-based Abstract Summarization dataset(CABAS),show that the proposed model outperform existing baselines including large language models and yield robust generalization ability across different domains.更多还原