A corpus of drafts of NLP papers from non-native English speakers

We created an English parallel corpus of 3,005 sentence pairs, each containing a well-polished text from ACL Anthology Reference Corpus (ACL-ARC) [1] and corresponding restated drafts collected from 26 non-native writers. The purpose of this paper is to explore the writing features of the drafts from non-native English speakers, so as to benefit research in Academic Writing Aid Systems. We present a feature analysis of the corpus based on handcrafted features. To assess utility, we formulate a draft identification task to automatically recognize drafts from ground truth texts based on hybrid features. We show that the combination of deep semantic features with the optimal handcrafted features improves identification accuracy on the collected data, up to 84.57%.