GSum: A General Framework for Guided Neural Abstractive Summarization

Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 4830–4842 June 6–11, 2021. ©2021 Association for Computational Linguistics 4830 GSum: A General Framework for Guided Neural Abstractive Summarization Zi-Yi Dou, Pengfei Liu, Hiroaki Hayashi, Zhengbao Jiang, Graham Neubig Language Technologies Institute, Carnegie Mellon University {zdou, pliu3, hiroakih, zhengbaj, gneubig}@cs.cmu.edu Abstract Neural abstractive summarization models are flexible and can produce coherent summaries, but they are sometimes unfaithful and can be difficult to control. While previous studies at- tempt to provide different types of guidance to control the output and increase faithfulness, it is not clear how these strategies compare and contrast to each other. In this paper, we propose a general and extensible guided summarization framework (GSum) that can effectively take different kinds of external guidance as input, and we perform experi- ments across several different varieties. Ex- periments demonstrate that this model is ef- fective, achieving state-of-the-art performance according to ROUGE on 4 popular summariza- tion datasets when using highlighted sentences as guidance. In addition, we show that our guided model can generate more faithful sum- maries and demonstrate how different types of guidance generate qualitatively different sum- maries, lending a degree of controllability to the learned models.1 1 Introduction Modern techniques for text summarization gener- ally can be categorized as either extractive meth- ods (Nallapati et al., 2017; Narayan et al., 2018b; Zhou et al., 2018), which identify the most suit- able words or sentences from the input document and concatenate them to form a summary, or ab- stractive methods (Rush et al., 2015; Chopra et al., 2016; Nallapati et al., 2016; Paulus et al., 2018), which generate summaries freely and are able to produce novel words and sentences. Compared with extractive algorithms, abstractive algorithms are more flexible, making them more likely to pro- duce fluent and coherent summaries. However, the unconstrained nature of abstractive summarization can also result in problems. First, it can result 1Code is available at https://github.com/ neulab/guided_summarization. Source Document Guidance Signal Encoder Oracle Decoder Source Document Guidance Signal Encoder Auto. Extracted/ User Specifed Decoder Training Testing Figure 1: Our framework generates summaries using both the source document and separate guidance sig- nals. We use an oracle to select guidance during train- ing and use automatically extracted or user-specified guidance at test time. in unfaithful summaries (Kry´sci´nski et al., 2019), containing factual errors as well as hallucinated content. Second, it can be difficult to control the content of summaries; it is hard to pick in advance which aspects of the original content an abstrac- tive system may touch upon. To address the issues, we propose methods for guided neural abstractive summarization: methods that provide various types of guidance signals that 1) constrain the summary so that the output content will deviate less from the source document; 2) allow for controllability through provision of user-specified inputs. There have been some previous methods for guiding neural abstractive summarization models. For example, Kikuchi et al. (2016) specify the …

Content truncated to 500 words