Differentiable N-gram objective on abstractive summarization

Authors：Zhu, Yunqi; Yang, Xuebing^*; Wu, Yuanyuan; Zhu, Mingjin; Zhang, Wensheng^*

Source：EXPERT SYSTEMS WITH APPLICATIONS, 2023, 215: 119367.

DOI：10.1016/j.eswa.2022.119367

• Abstractive summarization
• Differentiable N-gram objective
• Neural network language model

ROUGE is a standard automatic evaluation metric based on N-gram for sequence-to-sequence tasks like abstractive summarization, while cross-entropy loss is an essential objective that optimizes at unigram level for neural network language models. In this paper we present differentiable N-gram objectives, attempting to alleviate the discrepancy between training and evaluating criteria. The novelty of our work is the objective does not ceil the number of matched sub-sequences by the ground truth count of N-gram in reference sequence and weights the matched sub-sequences equally. Therefore, our proposed objective can maximize the probabilistic weight of matched sub-sequences. We jointly optimize cross-entropy loss and the objective, providing decent evaluation scores enhancement including ROUGE over abstractive summarization datasets CNN/DM and XSum, outperforming competitive N-gram objectives.

• Institution
  Chinese Academy of Science; Hainan University