Graph-enhanced multi-answer summarization under question-driven guidance

Bing Li, Peng Yang, Zhongjian Hu, Yuankang Sun, Meng Yi

Abstract: Multi-answer summarization for question-based queries in community Q&A requires comprehensive and in-depth analysis of lengthy and extensive information to generate concise and comprehensive answer summarization. Guided by the questions, capturing the relationships among candidate answers significantly benefits detecting salient information from multiple answers and generating an overall coherent summarization. In this paper, we propose a new Graph-enhanced Multi-answer Summarization under Question-driven Guidance model that enables explicit handling of the salience and redundancy of answer information. Specifically, the model first fully incorporates a pre-trained model to learn linguistic features through encoding and focuses on the role of questions in guiding answer generation during the encoding phase. The questions are utilized to explicitly constrain individual answers to ensure that the model more accurately identifies the information closely related to the questions in the answers and allocates more attention. Moreover, we utilize the question-driven answer graph information for encoding to capture the modeling relationships between answers and remove information redundancy. Finally, the graph-encoded information is exploited in the decoding stage to guide the generation of summaries to guarantee the informativeness, fluency, and conciseness of the summarization. Experimental results show that our proposed model brings substantial improvements compared to the state-of-the-art baseline, achieving the best outcomes on both of the community datasets ELI5 and MEDIQA, demonstrating the effectiveness of our model.