Personalizing Lexical Simplification

Proceedings of the 27th International Conference on Computational Linguistics, pages 224–232 Santa Fe, New Mexico, USA, August 20-26, 2018. 224 Personalizing Lexical Simplification John Lee, Chak Yan Yeung Department of Linguistics and Translation City University of Hong Kong jsylee@cityu.edu.hk, chak.yeung@my.cityu.edu.hk Abstract Given an input text from the user, a lexical simplification (LS) system makes the text easier to understand by substituting difficult words with simpler words. The best substitution may vary from one user to another, given individual differences in vocabulary proficiency level. Most current systems, however, do not consider these variations, and are instead trained to find one optimal substitution or list of substitutions for all users. This paper measures the benefits of using complex word identification (CWI) models to personalize an LS system. Experimental results show that even a simple CWI model, based on graded vocabulary lists, can help reduce the number of unnecessary simplifications and complex words in the output for learners of English at different proficiency levels. 1 Introduction Lexical simplification (LS) is the task of replacing difficult words with simple words in a text, while preserving its meaning and grammaticality. It aims to produce output text that is easier to understand for readers with special needs, such as language learners, children (Kajiwara et al., 2013), and those with language disabilities (Devlin and Tait, 1998; Carroll et al., 1999). Table 1 shows an example input sentence to an LS system, and the ranked list of possible substitutions for the target word, i.e., the word that should be simplified. Most LS systems first perform complex word identification (CWI) to detect target words (i.e., “avoid” in this case), and then find appropriate substitutions for them (i.e., “prevent”, “stop”, etc., in order of preference). CWI is thus an important first step in the LS pipeline. On the one hand, an overly conservative CWI model would fail to detect many complex words, leaving them unsimplified and limiting the utility of the LS system. On the other hand, an overly aggressive CWI model would be prone to misidentify simple words as complex, leading to unnecessary simplifications and increasing the risk of substitution errors. In an error analysis on LS systems, CWI-related error categories turned out to be among the most frequent (Shardlow, 2014). CWI has been receiving increasing attention in recent years, including a recent SemEval shared task (Paetzold and Specia, 2016b). Since the test set was annotated by a single learner, however, CWI performance on language learners at different levels of vocabulary proficiency continues to be under-explored. Indeed, most LS evaluations assume one best substitution or one fixed ranked list of substitutions (cf. Table 1), and do not take into account variations in vocabulary knowledge among users. This “one- size-fits-all” approach is suboptimal since word complexity is in the eye of the beholder: a word that is complex for a low-proficiency user may be perfectly familiar to a high-proficiency user, or even to a low-proficiency user whose native language has a cognate word. As a case in point, consider the …

Content truncated to 500 words