Auto-regressive extractive summarization with replacement

Abstract Auto-regressive extractive summarization approaches determine sentence extraction probability conditioning on previous decisions by maintaining a partial summary representation. Despite its popularity, the framework has two main drawbacks: 1) representation is irresolutely denoted weighted summation of all processed sentences without any filtering, resulting in noisy and degrading effectiveness extracting subsequent sentences; 2) earlier are biased towards higher due to sequential nature sequence tagging. To address these problems, we propose Extractive Summarization with Replacement (AES-Rep), novel auto-regressive model. In particular, AES-Rep model consists modules: decision module that determines whether should be extracted, replacement locater enables extracted deficient replaced latter comparing their expressiveness respect idea document. These modules update explicit actions using elaborated multidimensional guidance. We conduct extensive experiments benchmark CNN DailyMail datasets. Experimental results show can achieve better performance compared various strong baselines terms multiple ROUGE metrics.