Paragraph-level Attention-Aware Inference for Multi-Document Neural Abstractive Summarization

Inspired by Google's Neural Machine Translation (NMT) that models the one-to-one alignment in translation tasks with an uniform attention distribution during the inference, this study proposes an attention-aware inference algorithm for Neural Abstractive Summarization (NAS) to regulate generated summaries to attend to source contents with the optimal coverage. Unlike NMT, NAS is not based on one-to-one transformation. Instead, its attention distribution for the input should be irregular and depend on the content layout of the source documents. To address this matter, we construct an attention-prediction model to learn the dependency between the optimal attention distribution and the source. By refining the vanilla beam search with the attention-aware mechanism, significant improvements on the quality of summaries could be observed. Last but not the least, the attention-aware inference has strong universality that can be easily adopted to different hierarchical summarization models to promote the models' performance.