katildakat at SemEval-2021 Task 1: Lexical Complexity Prediction of Single Words and Multi-Word Expressions in English

Proceedings of the 15th International Workshop on Semantic Evaluation (SemEval-2021), pages 700–705 Bangkok, Thailand (online), August 5–6, 2021. ©2021 Association for Computational Linguistics 700 katildakat at SemEval-2021 Task 1: Lexical Complexity Prediction of Single Words and Multi-Word Expressions in English Katja Voskoboinik Aalto University Helsinki, Finland ekaterina.voskoboinik@aalto.fi Abstract This paper describes systems submitted to Se- mEval 2021 Task 1: Lexical Complexity Pre- diction (LCP). We compare a linear and a non- linear regression models trained to work for both tracks of the task. We show that both systems are able to generalize better when supplied with information about complexities of single word and multi-word expression (MWE) targets simultaneously. This approach proved to be the most beneficial for multi- word expression targets. We also demonstrate that some hand-crafted features differ in their importance for the target types. 1 Introduction SemEval-2021 Task 1 is the task of Lexical Com- plexity Prediction (LCP) (Shardlow et al., 2021). The goal of the task is to assign a target in a con- text a continuous value ranging between 0 and 1, where 1 indicates complete unintelligibility and 0 signals perfect familiarity as perceived by a native speaker. The task has two tracks: predicting the complexity score of single words and predicting the complexity score of multi-word expressions (MWE). Such a task can be useful in applications like text simplification or automatic language profi- ciency evaluation. The CompLex dataset (Shardlow et al., 2020) used in this task is the first English dataset for the task of LCP. The dataset contains single words and MWEs annotated with their lexical complexity score in a specific context. The annotations were provided by native speakers of English. The tar- gets and their contexts were obtained from texts of different domains: the Bible, Europarl, and biomed- ical texts. The dataset opens several avenues for research. For instance, how does the perceived complexity of single words and MWEs differ? How does context affect the lexical difficulty of a target? How does text genre affect comprehensibility of words? We were interested if there is a difference in performance when using the same representation methods for single words and MWEs. It was de- cided to approach both tracks as the same problem. We did not distinguish between MWEs and single words and they both were treated as one lexical unit. The same array of features was extracted to represent the targets and linear and non-linear re- gressors were trained using both subcorpora. This strategy showed performance gains for both single targets and MWEs.1 In addition, we wanted to investigate how much the classic hand-crafted features like frequency and length together with subword information and con- textualized embeddings (not employed previously for LCP) contribute to complexity estimation of both single words and MWEs. We present the anal- ysis of feature imortance rankings in Section 6. 2 Related Work Complex Word Identification (CWI) is the task of determining how difficult a lexical unit is to a target audience (Shardlow, 2013). The knowledge about the lexical unit …

Content truncated to 500 words