Redundancy Detection in ESL Writings

Proceedings of the 14th Conference of the European Chapter of the Association for Computational Linguistics, pages 683–691, Gothenburg, Sweden, April 26-30 2014. c⃝2014 Association for Computational Linguistics Redundancy Detection in ESL Writings Huichao Xue and Rebecca Hwa Department of Computer Science, University of Pittsburgh, 210 S Bouquet St, Pittsburgh, PA 15260, USA {hux10,hwa}@cs.pitt.edu Abstract This paper investigates redundancy detec- tion in ESL writings. We propose a mea- sure that assigns high scores to words and phrases that are likely to be redundant within a given sentence. The measure is composed of two components: one cap- tures fluency with a language model; the other captures meaning preservation based on analyzing alignments between words and their translations. Experiments show that the proposed measure is five times more accurate than the random baseline. 1 Introduction Writing concisely is challenging. It is especially the case when writing in a foreign language that one is still learning. As a non-native speaker, it is more difficult to judge whether a word or a phrase is redundant. This study focuses on auto- matically detecting redundancies in English as a Second Language learners’ writings. Redundancies occur when the writer includes some extraneous word or phrase that do not add to the meaning of the sentence but possibly make the sentence more awkward to read. Upon re- moval of the unnecessary words or phrases, the sentence should improve in its fluency while main- taining the original meaning. In the NUCLE corpus (Dahlmeier and Ng, 2011), an annotated learner corpus comprised of essays written by pri- marily Singaporean students, 13.71% errors are tagged as “local redundancy errors”, making re- dundancy error the second most frequent prob- lem.1 Although redundancies occur frequently, it has not been studied as widely as other ESL errors. A 1The most frequent error type is Wrong collocation/idiom preposition, which comprises 15.69% of the total errors. major challenge is that, unlike mistakes that vio- late the grammaticality of a sentence, redundan- cies do not necessarily “break” the sentence. De- termining which word or phrase is redundant is more of a stylistic question; it is more subjective, and sometimes difficult even for a native speaker. To the best of our knowledge, this paper reports a first study on redundancy detection. In particu- lar, we focus on the task of defining a redundancy measure that estimates the likelihood that a given word or phrase within a sentence might be extrane- ous. We propose a measure that takes into account each word’s contribution to fluency and meaning. The fluency component computes the language model score of the sentence after the deletion of a word or a phrase. The meaning preservation com- ponent makes use of the sentence’s translation into another language as pivot, then it applies a statis- tical machine translation (SMT) alignment model to infer the contribution of each word/phrase to the meaning of the sentence. As a first experiment, we evaluate our measures on their abilities in picking the most redundant phrase of a given length. We show …

Content truncated to 500 words