CEFR-based Lexical Simplification Dataset

CEFR-based Lexical Simplification Dataset Satoru Uchida†, Shohei Takada‡, Yuki Arase‡ † Faculty of Languages and Cultures, Kyushu University ‡ Graduate School of Information Science and Technology, Osaka University † uchida@flc.kyushu-u.ac.jp, ‡{takada.syouhei, arase}@ist.osaka-u.ac.jp Abstract This study creates a language dataset for lexical simplification based on Common European Framework of References for Languages (CEFR) levels (CEFR-LS). Lexical simplification has continued to be one of the important tasks for language learning and education. There are several language resources for lexical simplification that are available for generating rules and creating simplifiers using machine learning. However, these resources are not tailored to language education with word levels and lists of candidates tending to be subjective. Different from these, the present study constructs a CEFR-LS whose target and candidate words are assigned CEFR levels using CEFR-J wordlists and English Vocabulary Profile, and candidates are selected using an online thesaurus. Since CEFR is widely used around the world, using CEFR levels makes it possible to apply a simplification method based on our dataset to language education directly. CEFR-LS currently includes 406 targets and 4912 candidates. To evaluate the validity of CEFR-LS for machine learning, two basic models are employed for selecting candidates and the results are presented as a reference for future users of the dataset. Keywords: Lexical simplification, ESL learners, paraphrasing, CEFR 1. Introduction There is no doubt that vocabulary is the key to success- ful international communication. Laufer (1989) points out that learners of a foreign language need to know 95% of the words in the input text to be able to successfully un- derstand the text message, which implies the importance of vocabulary in language learning. However, authentic En- glish passages may contain difficult words that may hinder the readers’ or listeners’ comprehension. One solution is to simplify words in the input text using statistical and computer-based methods (Horn et al., 2014; Biran et al., 2011; Glavas and Stajner, 2015). This task is called lexical simplification whose aim is to replace a dif- ficult word (referred to as the target word in this paper) with a simpler word selected from candidates (referred to as candidate words in this paper), which has been featured at SemEval20121. Although previous studies have shown that data-driven approaches to find a simpler word are use- ful to some extent (Horn et al., 2014; Glavas and Stajner, 2015), they do not place the focus on educational aspects and hence tend to discuss only technical issues. Also, the criteria for the word difficulty tend to be vague and sub- jective, which would imply that previous research methods may not be useful for suggesting simplifications for differ- ent proficiency levels of learners. The present study attempts to construct a dataset that uses Common European Framework of Reference for Languages (CEFR) as a criterion for word levels using introductory parts of university textbooks whose contents are academi- cally reliable, educationally important, and covering a vari- ety of topics. In this sense, our dataset is education-oriented based on a solid and widely-used framework for …

Content truncated to 500 words