Incorporating Question Answering-Based Signals into Abstractive Summarization via Salient Span Selection

Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics, pages 575–588 May 2-6, 2023 ©2023 Association for Computational Linguistics Incorporating Question Answering-Based Signals into Abstractive Summarization via Salient Span Selection Daniel Deutsch† Google dandeutsch@google.com Dan Roth University of Pennsylvania danroth@seas.upenn.edu Abstract In this work, we propose a method for in- corporating question-answering (QA) signals into a summarization model. Our method identifies salient noun phrases (NPs) in the input document by automatically generating wh-questions that are answered by the NPs and automatically determining whether those questions are answered in the gold summaries. This QA-based signal is incorporated into a two-stage summarization model which first marks salient NPs in the input document us- ing a classification model, then condition- ally generates a summary. Our experiments demonstrate that the models trained using QA-based supervision generate higher-quality summaries than baseline methods of identify- ing salient spans on benchmark summarization datasets. Further, we show that the content of the generated summaries can be controlled based on which NPs are marked in the input document. Finally, we propose a method of augmenting the training data so the gold sum- maries are more consistent with the marked in- put spans used during training and show how this results in models which learn to better ex- clude unmarked document content.1 1 Introduction Abstractive sequence-to-sequence summarization models have become very effective methods of easily generating summaries of input documents (Rush et al., 2015; Nallapati et al., 2016; Lewis et al., 2020). Previous work has demonstrated that condition- ing the summary generation on salient document sentences results in higher-quality summaries and more controllable summarization models (Chen and Bansal, 2018; Dou et al., 2021). Salient sen- tences are typically identified during training by 1http://cogcomp.org/page/publication_view/997 †Work done while at the University of Pennsylvania Incumbent Goodluck Jonathan phoned former military leader Muhammadu Buhari on Tuesday to concede defeat in Nigeria’s presidential elections, Buhari’s party says. Jonathan acknowledged the phone call and his defeat in a written statement to his countrymen. “I thank all Nigerians once again for the great opportunity… I promised the country free and fair elections. I have kept my word…” Buhari ruled Nigeria from late 1983 until August 1985… The 72-year-old retired major general’s experience… Input Document Incumbent President Goodluck Jonathan acknowledges defeat, says he delivered on promise of fair elections. Muhammadu Buhari ’s party says Jonathan called to concede even before final results are announced. Buhari is a 72-year-old retired major general who ruled in Nigeria in the 1980s. Gold Summary Figure 1: Salient spans identified by QA-based signals (shown in color) more precisely identify salient docu- ment content than those that identify salient sentences based on lexical overlap (shown in bold). Our method classifies the salient spans, marks them in the input doc- ument, and then generates a summary. lexical overlap with the gold summaries (Nallapati et al., 2017) and predicted during inference. Although marking different sentences as salient allows for some controllability over the content of the summary, desired summary content …

Content truncated to 500 words