Fact-Aware Abstractive Summarization Based on Prompt Information and Re-Ranking
Tianyu Cai, Yurui Yang, Yujie Wan, Xingjing Mao, Shenggen Ju · 2024
Abstractive text summarization helps people quickly obtain the key information of an article, and existing models generate fluent summaries but often suffer from factual consistency problems, a key issue that current research has not adequately addressed. In order to ensure the quality of summaries while improving their factualness, the article proposes a factaware summary generation method that combines reordering and prompting information to improve the quality and factual consistency of generated summaries. Keyword extraction and key phrases are first introduced and then fed into the generative abstract model along with the original text to obtain candidate abstracts that incorporate factual information from the original text. In order to combine abstract quality and factual consistency, the ROUGE metrics and FactCC metrics are combined, and a reordering rule is designed to guide the model in generating more realistic and content-rich summaries. In addition, the model is further incentivized to assign higher probability scores to more authentic summaries through the introduction of contrast loss. Experimental results on CNN/Daily Mail and XSum datasets show that the article's proposed method outperforms the strong baseline model in terms of quality and factual consistency, and ablation experiments validate the effectiveness of the proposed module.