CTRLsum: Towards Generic Controllable Text Summarization

Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 5879 - 5915 December 7-11, 2022 ©2022 Association for Computational Linguistics CTRLSUM: Towards Generic Controllable Text Summarization Junxian He∗†, Wojciech Kry´sci´nski‡, Bryan McCann‡, Nazneen Rajani‡, Caiming Xiong‡ †Carnegie Mellon University ‡Salesforce Research junxianh@cs.cmu.edu {kryscinski, bmccann, nazneen.rajani, cxiong}@salesforce.com Abstract Current summarization systems yield generic summaries that are disconnected from users’ preferences and expectations. To address this limitation, we present CTRLSUM, a generic framework to control generated summaries through a set of keywords. During training keywords are extracted automatically without requiring additional human annotations. At test time CTRLSUM features a control function to map control signal to keywords; through engineering the control function, the same trained model is able to be applied to control summaries on various dimensions, while nei- ther affecting the model training process nor the pretrained models. We additionally ex- plore the combination of keywords and text prompts for more control tasks. Experiments demonstrate the effectiveness of CTRLSUM on three domains of summarization datasets and five control tasks: (1) entity-centric and (2) length-controllable summarization, (3) contri- bution summarization on scientific papers, (4) invention purpose summarization on patent fil- ings, and (5) question-guided summarization on news articles. Moreover, when used in a standard, unconstrained summarization setting, CTRLSUM is comparable or better than strong pretrained systems. 1 1 Introduction Neural summarization systems aim to compress a document into a short paragraph or sentence while preserving key information. There are two com- mon categories of summarization systems: extrac- tive summarization where models find and copy important portions of a documents (Cheng and La- pata, 2016; Nallapati et al., 2017; Narayan et al., 2018), and abstractive summarization where mod- els freely generates novel sentences (Rush et al., ∗Work done during JH’s internship at Salesforce Research. 1Code, data, model checkpoints, and demos are available at https://github.com/salesforce/ctrl-sum. 2015; See et al., 2017; Paulus et al., 2018). The focus of this work is on abstractive summarization. In a common setting, summarization methods take a source document as input and yield a generic summary which covers content selected arbitrarily by the model. However, to be useful, automati- cally generated summaries should cover content considered important by the readers. For example, Figure 1 shows an NBA basketball news article, and the reference summary aggregates the results of several matches. Yet fans of certain players or teams might only be interested in the matches and statistics which include their entities of interest. Motivated by this observation, in this work we focus on controllable summarization which allows the users to control the generated summaries. We propose CTRLSUM, a framework to control sum- maries through a set of keywords. At training time, the model learns to predict summaries conditioned on both the source document and keywords, which are easily identified from training summaries. Dur- ing inference, a control function is designed de- pending on the specific control aspect to map user preference to keywords to control the summary. While simple, keywords provide a clean separa- tion of test-time …

Content truncated to 500 words