Recovering Lexically and Semantically Reused Texts

Proceedings of the 10th Conference on Lexical and Computational Semantics, pages 52–66 August 5–6, 2021, Bangkok, Thailand (online) ©2021 Association for Computational Linguistics 52 Recovering Lexically and Semantically Reused Texts Ansel MacLaughlin∗, Shaobin Xu∗, David A. Smith Khoury College of Computer Science Northeastern University {ansel, shaobinx, dasmith}@ccs.neu.edu Abstract Writers often repurpose material from existing texts when composing new documents. Be- cause most documents have more than one source, we cannot trace these connections us- ing only models of document-level similarity. Instead, this paper considers methods for local text reuse detection (LTRD), detecting local- ized regions of lexically or semantically simi- lar text embedded in otherwise unrelated mate- rial. In extensive experiments, we study the relative performance of four classes of neu- ral and bag-of-words models on three LTRD tasks – detecting plagiarism, modeling jour- nalists’ use of press releases, and identifying scientists’ citation of earlier papers. We con- duct evaluations on three existing datasets and a new, publicly-available citation localization dataset. Our findings shed light on a num- ber of previously-unexplored questions in the study of LTRD, including the importance of in- corporating document-level context for predic- tions, the applicability of of-the-shelf neural models pretrained on “general” semantic tex- tual similarity tasks such as paraphrase detec- tion, and the trade-offs between more efficient bag-of-words and feature-based neural models and slower pairwise neural models. 1 Introduction When composing documents in many genres— from news reports, to scientific papers, to politi- cal speeches—authors obtain ideas and inspiration from source documents and present them in the form of direct copies, quotations, summaries, or paraphrases. In the simplest case, e.g. in con- gressional bills, writers include text from earlier versions of the same document along with new material (Wilkerson et al., 2015). In news me- dia, journalists often paraphrase or quote speeches, press releases, and interviews (Niculae et al., 2015; ∗Equal contribution. Tan et al., 2016). In academia, citations of papers usually appear along with summaries of their con- tributions (Qazvinian and Radev, 2010). These are instances of lexical and semantic local text reuse, where both source and target documents contain lexically or semantically similar passages, surrounded by text that is unrelated or dissimilar. Often, reused text is presented without explicit links or citations, making it hard to track infor- mation flow. While many state-of-the-art (SoTA) NLP archi- tectures have been trained on the closely-related tasks of document- and sentence-pair similarity de- tection (Reimers and Gurevych, 2019) and ad-hoc retrieval (Dai and Callan, 2019), prior methods for local text-reuse detection (LTRD) are mostly lim- ited to lexical matching (Lee, 2007; Clough et al., 2002; Leskovec et al., 2009; Wilkerson et al., 2015; Smith et al., 2014) with some dictionary expan- sion (Moritz et al., 2016). To our knowledge, only Zhou et al. (2020) has applied neural models to this problem, proposing hierarchical neural mod- els that use a cross-document attention mechanism to model local similarities between two candidate documents. In this paper, we conduct a large-scale evalu- ation of several lexical overlap and SoTA …

Content truncated to 500 words