Identifying Semantic Edit Intentions from Revisions in Wikipedia

Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pages 2000–2010 Copenhagen, Denmark, September 7–11, 2017. c⃝2017 Association for Computational Linguistics Identifying Semantic Edit Intentions from Revisions in Wikipedia Diyi Yang♣, Aaron Halfaker♦, Robert Kraut♠, Eduard Hovy♣ ♣ Language Technologies Institute, Carnegie Mellon University {diyi,hovy}@cmu.edu ♦ Wikimedia Foundation ahalfaker@wikimedia.org ♠ Human-Computer Interaction Institute, Carnegie Mellon University robert.kraut@cmu.edu Abstract Most studies on human editing focus merely on syntactic revision operations, failing to capture the intentions behind revision changes, which are essential for facilitating the single and collaborative writing process. In this work, we develop in collaboration with Wikipedia editors a 13-category taxonomy of the semantic in- tention behind edits in Wikipedia articles. Using labeled article edits, we build a computational classifier of intentions that achieved a micro-averaged F1 score of 0.621. We use this model to investigate edit intention effectiveness: how differ- ent types of edits predict the retention of newcomers and changes in the quality of articles, two key concerns for Wikipedia today. Our analysis shows that the types of edits that users make in their first ses- sion predict their subsequent survival as Wikipedia editors, and articles in different stages need different types of edits. 1 Introduction Many online text production communities, in- cluding Wikipedia, maintain a history of revi- sions made by millions of participants. As Wikipedia statistics as of January 2017 show, English Wikipedia has 5.3 million articles with an average of 162.89 revisions per article, with revisions growing at a rate of about 2 revisions per second. This provides an amazing corpus for studying the types and effectiveness of revisions. Specifically, differences between revisions contain valuable information for modeling document qual- ity or extracting users’ expertise, and can addition- ally support various natural language processing (NLP) tasks such as sentence compression (Ya- mangil and Nelken, 2008), lexical simplification (Yatskar et al., 2010), information retrieval (Aji et al., 2010), textual entailment recognition (Zan- zotto and Pennacchiotti, 2010), language bias de- tection (Recasens et al., 2013), spelling errors and paraphrases (Zesch, 2012; Max and Wisniewski, 2010). To avoid building different approaches to ex- tract the information needed by different NLP tasks (Ferschke et al., 2013), a unified framework to recognize edits from revisions is needed. Prior research on revision editing primarily develop syntactic edit action categories, from which they try to understand the effects of edits on meaning (Faigley and Witte, 1981; Yang et al., 2016). For instance, Daxenberger and Gurevych (2012) categorized edits based on whether edits affect the text meaning, resulting in syntactic edit categories such as file deletion, reference modification, etc. However, simply understanding the syntactic re- vision operation types does not provide the in- formation we seek: why do editors do what they do? how effective are their actions? For exam- ple, syntactic edit type taxonomies cannot tell the difference between simplifying a paragraph and maliciously damaging that paragraph, since both involve deleting a sentence. In this work, we focus explicitly on revision intention. We introduce a fine-grained taxonomy of the …

Content truncated to 500 words