Towards Summary Candidates Fusion

Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 8488 - 8504 December 7-11, 2022 ©2022 Association for Computational Linguistics Towards Summary Candidates Fusion Mathieu Ravaut♣♢, Shafiq Joty♣♠, Nancy F. Chen♢ ♣ Nanyang Technological University, Singapore ♢ Institute of Infocomm Research (I2R), Singapore ♠ Salesforce Research {mathieuj001@e.ntu, srjoty@ntu}.edu.sg nfychen@i2r.a-star.edu.sg Abstract Sequence-to-sequence deep neural models fine-tuned for abstractive summarization can achieve great performance on datasets with enough human annotations. Yet, it has been shown that they have not reached their full potential, with a wide gap between the top beam search output and the oracle beam. Re- cently, re-ranking methods have been proposed, to learn to select a better summary candi- date. However, such methods are limited by the summary quality aspects captured by the first-stage candidates. To bypass this limita- tion, we propose a new paradigm in second- stage abstractive summarization called Sum- maFusion that fuses several summary candi- ates to produce a novel abstractive second- stage summary. Our method works well on several summarization datasets, improv- ing both the ROUGE scores and qualitative properties of fused summaries. It is espe- cially good when the candidates to fuse are worse, such as in the few-shot setup where we set a new state-of-the-art. We will make our code and checkpoints available at https: //github.com/ntunlp/SummaFusion/. 1 Introduction Leading abstractive summarization methods typ- ically rely on transfer learning and the pre-train- then-finetune paradigm. In this approach, a deep sequence-to-sequence neural model is first pre- trained on a very large text corpus, either with a general-purpose text generation objective like masked span generation as in T5 (Raffel et al., 2019), BART (Lewis et al., 2020) and ProphetNet (Qi et al., 2021), or with a pre-training objective specific to summarization as in PEGASUS (Zhang et al., 2020) and TED (Yang et al., 2020). Then, the model is fine-tuned on the downstream sum- marization dataset(s) of interest, which can have a wildly varying amount of human labels, from a few thousands to hundreds of thousands. Source document: this happen yesterday afternoon. i been trying to dual boot mac os and windows on my wife’s macbook pro. it’s a late 2011 model so support from apple is almost nonexistent which is great when they wanted to charge me chat with them. i convinced them to a free chat and learned that apparently my hardware is to out dated to have boot camp make a bootable usb. ...boot camp assistant on this macbook only does cd iso img. 3: delete the wrong hard dive part and corrupt the hard drive and have to re format the whole computer and lose every file that was saved on her computer since 2011. i got yelled at for a good hour. i knew it was my fault but at the same time... how in the world have you not backed your things in 4 years! Summary candidates (PEGASUS with diverse beam search): 1: got yelled at for 4 years. 2: i tried to back up my wife’s …

Content truncated to 500 words