Generating personalized article edits on collaborative editing platforms

Anonymous

Abstract: NLP methods to generate edits on collaborative editing platforms can help users to edit more efficiently and suggest locations within an article for editing. Existing methods have largely ignored the personalized aspect of editing--the diverse styles, interests, and editing intentions that affect user edits. In this paper, we analyze two personalization methods: augmenting models with user behavior clusters and user tags. We demonstrate that these methods, when combined with a new architecture, generate edits that are closer to ground-truth Wikipedia edits when compared to an existing strong baseline. Our experiments test edits for both edit type (insertion or deletion) and word choice, and include a user study collecting feedback from human evaluators. Finally, we introduce a new dataset of Wikipedia edits to facilitate future innovation.

Paper Type: short