Alignment-based Annotation of Proofreading Texts toward Professional Writing Assistance

International Joint Conference on Natural Language Processing, pages 753–759, Nagoya, Japan, 14-18 October 2013. Alignment-based Annotation of Proofreading Texts toward Professional Writing Assistance Ngan L.T. Nguyen University of Information Technology Hochiminh, Vietnam ngannlt@uit.edu.vn Yusuke Miyao National Institute of Informatics Tokyo, Japan yusuke@nii.ac.jp Abstract This work aims at constructing a corpus to satisfy such requirements to support re- search towards professional writing assis- tance. Our corpus is a collection of sci- entific work written by non-native speak- ers that has been proofread by native En- glish experts. A new annotation scheme, which is based on word-alignments, is then proposed that is used to capture all types of inarticulations and their correc- tions including both spelling/grammatical error corrections and paraphrases made by proofreaders. The resulting corpus con- tains 3,485 pairs of original and revised sentences, of which, 2,516 pairs contain at least one articulation. 1 Introduction Detection and correction of misspellings and grammatical errors have been recognized as key techniques for writing assistance, and have exten- sively been studied in natural language process- ing (NLP) (Whitelaw et al., 2009; Gamon, 2010; Tetreault et al., 2010; Park and Levy, 2011). How- ever, correcting misspellings and grammatical er- rors, which can be performed by normal English native speakers, does not satisfy all the require- ments of professional writing (Futagi, 2010). The core of the proofreading process, in reality, is para- phrasing inarticulations, which can only be done by expert proofreaders. Considering the two para- phrased sentences (1a) and (1b) below, we can see that sentence (1b) is likely to be considered better by most people (Williams and Colomb, 2010), al- though neither of them contains any misspellings or grammatical errors. (1a) The outsourcing of high-tech work to Asia by cor- porations means the loss of jobs for many middle-class American workers. SWA corpus Data Collection Preprocessing Corpus Annotation Annotation Scheme Annotation Scheme Design Scientific written work Proofreading Figure 1: Methodology for corpus annotation (1b) Many middle-class American workers are losing their jobs, because corporations are outsourcing their high-tech work to Asia. (Williams and Colomb, 2010) Most of the existing corpora are designed to cap- ture errors in spelling and grammar, but they have not paid enough attention to paraphrasing. We constructed a corpus that we called sci- entific writing assistance corpus (SWA), to sup- port research on assistance with scientific-writing that captures all types of inarticulations, including those in both mispellings/grammar and paraphras- ing. We have used the term inarticulation and inarticulation correction instead of error and er- ror correction in this paper, to include in our task the paraphrasing, which is actually not errors. Figure 1 overviews the methodology we pro- posed to construct the corpus. Scientific work written by non-native researchers or graduate stu- dents are collected (i.e., data collection, see Sec- tion 3), and this was then proofread by English native experts (i.e., proofreading). After that, we preprocessed the documents to convert them into a predefined format (i.e., preprocessing, see Section 3). Annotators with linguistic backgrounds were asked to strictly follow …

Content truncated to 500 words