Summarizing Text on Any Aspects: A Knowledge-Informed Weakly-Supervised Approach

Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing, pages 6301–6309, November 16–20, 2020. c⃝2020 Association for Computational Linguistics 6301 Summarizing Text on Any Aspects: A Knowledge-Informed Weakly-Supervised Approach Bowen Tan1, Lianhui Qin2, Eric P. Xing1,3, Zhiting Hu1,4 1Carnegie Mellon University, 2University of Washington, 3Petuum Inc., 4UC San Diego {btan2,epxing}@andrew.cmu.edu, lianhuiq@cs.washington.edu, zhitinghu@gmail.com Abstract Given a document and a target aspect (e.g., a topic of interest), aspect-based abstractive summarization attempts to generate a sum- mary with respect to the aspect. Previous stud- ies usually assume a small pre-defined set of aspects and fall short of summarizing on other diverse topics. In this work, we study summa- rizing on arbitrary aspects relevant to the doc- ument, which significantly expands the appli- cation of the task in practice. Due to the lack of supervision data, we develop a new weak supervision construction method and an aspect modeling scheme, both of which integrate rich external knowledge sources such as Concept- Net and Wikipedia. Experiments show our ap- proach achieves performance boosts on sum- marizing both real and synthetic documents given pre-defined or arbitrary aspects.1 1 Introduction Remarkable progresses have been made in gener- ating generic summaries of documents (Nallapati et al., 2016; See et al., 2017; Narayan et al., 2018), partially due to the large amount of supervision data available. In practice, a document, such as a news article or a medical report, can span multiple topics or aspects. To meet more specific infor- mation need in applications such as personalized intelligent assistants, it is often useful to summa- rize a document with regard to a given aspect, i.e., aspect-based summarization. Recent research has explored the problem of aspect-based abstractive summarization (Krishna and Srinivasan, 2018; Frermann and Klementiev, 2019). A key challenge of the task is the lack of di- rect supervision data containing documents paired with multiple aspect-based summaries. Previous studies have created synthetic data from generic 1Code and data available at https://github.com/ tanyuqian/aspect-based-summarization news summarization corpora which have a small set of aspects (e.g., “sports”, “health” and other 4 aspects in (Frermann and Klementiev, 2019)). As a result, models trained on these data tend to be restricted to the pre-defined set and fall short of summarizing on other diverse aspects. This paper aims to go beyond pre-defined aspects and enable summarization on arbitrary aspects rel- evant to the document. The arbitrary aspect may not be explicitly mentioned but only implicitly re- lated to portions of the document, and it can be a new aspect not seen during training. To this end, we develop a new approach that integrates rich external knowledge in both aspect modeling and weak supervision construction. Specifically, we derive weak supervisions from a generic summa- rization corpus, where the ConceptNet knowledge graph (Speer et al., 2017) is used to substantially ex- pand the aspect scope and enrich the supervisions. To assist summarization model to better understand an aspect, especially a previously unseen one, we augment the model inputs with rich aspect-related information extracted from Wikipedia. Our …

Content truncated to 500 words