LAST at SemEval-2021 Task 1: Improving Multi-Word Complexity Prediction Using Bigram Association Measures

Proceedings of the 15th International Workshop on Semantic Evaluation (SemEval-2021), pages 571–577 Bangkok, Thailand (online), August 5–6, 2021. ©2021 Association for Computational Linguistics 571 LAST at SemEval-2021 Task 1: Improving Multi-Word Complexity Prediction Using Bigram Association Measures Yves Bestgen Laboratoire d’analyse statistique des textes - LAST Institut de recherche en sciences psychologiques Universit´e catholique de Louvain Place Cardinal Mercier, 10 1348 Louvain-la-Neuve, Belgium yves.bestgen@uclouvain.be Abstract This paper describes the system developed by the Laboratoire d’analyse statistique des textes (LAST) for the Lexical Complexity Prediction shared task at SemEval-2021. The proposed system is made up of a LightGBM model fed with features obtained from many word fre- quency lists, published lexical norms and psy- chometric data. For tackling the specificity of the multi-word task, it uses bigram associ- ation measures. Despite that the only contex- tual feature used was sentence length, the sys- tem achieved an honorable performance in the multi-word task, but poorer in the single word task. The bigram association measures were found useful, but to a limited extent. 1 Introduction For more than half a century, many studies have been carried out to collect norms about formal and semantic properties of words, such as frequency of use, spelling regularity, familiarity, age of acquisi- tion, or emotional valence (Proctor and Vu, 1999). Some of these properties can be easily harvested through automatic counting procedures applied to corpora. Other properties, such as familiarity or emotional valence, are obtained by requiring par- ticipants, often more than ten, to rate the words on these dimensions. In psycholinguistics, these norms have been mainly used for selecting experi- mental materials (Wilson, 1988). In computational linguistics, they are used in opinion mining, in the evaluation of foreign language skills and in text simplification for instance (Pang and Lee, 2008; Kyle et al., 2018). Obtaining lexical norms that require human evaluations is extremely costly in time and resources, which greatly reduces their size. However, huge norms are essential in applications (Bestgen, 1994). This observation has led to the de- velopment of automatic techniques to extend such norms (Bestgen, 2002; Kamps et al., 2004; Esuli and Sebastiani, 2006; Bestgen and Vincze, 2012). The Lexical Complexity Prediction (LCP) shared task at SemEval-2021 requires exactly the development of such techniques (Shardlow et al., 2021a). It is indeed a question of estimating the lexical complexity, the degree of difficulty of the words in a text. This dimension is important in NLP applications for simplifying texts and assisting spe- cific populations such as people with reading dis- abilities or who are learning a foreign language. A specificity of the LCP task is that it relates not only to words but also to multi-word expressions which are very rarely taken into account in norms and in automatic extension techniques (Bestgen, 2014). Another important feature of the task is that the target tokens were presented to human judges in context and that a significant number of them were presented in several different contexts. Human an- notations are therefore likely to reflect the impact …

Content truncated to 500 words