Word Complexity is in the Eye of the Beholder

Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 4439–4449 June 6–11, 2021. ©2021 Association for Computational Linguistics 4439 Word Complexity is in the Eye of the Beholder Sian Gooding Dept of Computer Science and Technology University of Cambridge shg36@cam.ac.uk Ekaterina Kochmar Dept of Computer Science University of Bath ek762@bath.ac.uk Seid Muhie Yimam Language Technology group Universität Hamburg, Germany yimam@informatik.uni-hamburg.de Chris Biemann Language Technology group Universität Hamburg, Germany biemann@informatik.uni-hamburg.de Abstract Lexical complexity is a highly subjective no- tion, yet this factor is often neglected in lexical simplification and readability systems which use a “one-size-fits-all" approach. In this pa- per, we investigate which aspects contribute to the notion of lexical complexity in various groups of readers, focusing on native and non- native speakers of English, and how the notion of complexity changes depending on the profi- ciency level of a non-native reader. To facili- tate reproducibility of our approach and foster further research into these aspects, we release a dataset of complex words annotated by read- ers with different backgrounds. 1 Introduction Complex word identification (CWI) is the first step in a lexical simplification (LS) pipeline, concerned with identification of words in text that are in need of further simplification (Shardlow, 2013). For in- stance, in example (1) a CWI system might identify engulfed as a complex word, which would allow an LS system to replace it with a simpler alternative, e.g. flooded, in the next step (Paetzold and Specia, 2016a; Gooding and Kochmar, 2019b): (1) Water engulfed Beringia. ↓ Water flooded Beringia. It has been shown that accurate CWI can sig- nificantly reduce errors in simplification (Shard- low, 2014), thus improving the quality of an LS system output (Lee and Yeung, 2018). In addi- tion, CWI has been shown to be an important com- ponent in readability assessment systems (Mad- dela and Xu, 2018) and in vocabulary acquisition modules of educational applications (Zaidi et al., 2020). However, an important aspect of CWI and LS that is often neglected is that text complexity is not an objective notion homogeneous across vari- ous target populations: what is challenging for a reader with a particular background (for example, a non-native reader at a lower level of language proficiency) would not necessarily be challenging for readers with other backgrounds (for example, more proficient readers) (Bingel, 2018). A num- ber of factors may contribute to that, including the reader’s age and level of language proficiency, among others (Paetzold and Specia, 2016c). LS systems often aim to address the needs of specific reader populations, such as children, non-native speakers, or readers with particular cognitive im- pairments. Thus, personalization in LS typically results in specialized simplification tools aimed at certain groups of readers (Carroll et al., 1998; Rello et al., 2013; Evans et al., 2014), with only a few systems addressing adaptation to the readers’ needs in a more dynamic way (Bingel et al., 2018; Yimam and Biemann, 2018a,b; Scarton and Specia, 2018). Despite CWI being …

Content truncated to 500 words