CompNA at SemEval-2021 Task 1: Prediction of lexical complexity analyzing heterogeneous features

Proceedings of the 15th International Workshop on Semantic Evaluation (SemEval-2021), pages 560–564 Bangkok, Thailand (online), August 5–6, 2021. ©2021 Association for Computational Linguistics 560 CompNA at SemEval-2021 Task 1: Prediction of lexical complexity analyzing heterogeneous features Giuseppe Vettigli Centrica plc giuseppe.vettigli@centrica.com Antonio Sorgente Institute of Applied Sciences and Intelligent Systems National Research Council antonio.sorgente@isasi.cnr.it Abstract This paper describes the CompNa model that has been submitted to the Lexical Complex- ity Prediction (LCP) shared task hosted at Se- mEval 2021 (Task 1). The solution is based on combining features of different nature through an ensembling method based on Decision Trees and trained using Gradient Boosting. We discuss the results of the model and highlight the features with more predictive capabilities. 1 Introduction Complex Word Identification (CWI) is a task fo- cused on the detection of complex (not easy to understand) words or expressions. One of the chal- lenges of natural language-based systems is to pro- vide informative content that is tailored to the needs of users in terms of content and level of understand- ing. For this aim, predicting Lexical Complexity plays a crucial role in simplifying the text so that it can be more easily understood by people with low literacy levels (for example children) or non na- tive speakers (Shardlow, 2013). The interest of the Computational Linguistics community on this topic has grown in recent years. Indeed, SemEval 2016 presented a challenge specifically on CWI (Paet- zold and Specia, 2016a; Zampieri et al., 2017). The task proposed to build models capable of predict- ing whether a word was easy or not for non-native English speakers. In this case, the proposed dataset was annotated with a binary label. In 2018 the CWI task was re-proposed at the workshop for Building Educational Applications (BEA) (Yimam et al., 2018). The data was again an- notated with binary labels but samples from more languages were considered, specifically English, German and Spanish. In addition to the word clas- sification task, a sub-task where the participants had to predict the probability of a word being com- plex was added. The best performing models for both tasks were based on ensembling techniques using features that were carefully selected (Paetzold and Specia, 2016b; Gooding and Kochmar, 2018). In this context, the Lexical Complexity Predic- tion (LCP) shared task hosted at SemEval 2021 proposes a challenge where the data is annotated according to the degree of complexity (Shardlow et al., 2021). This type of annotation allows re- gression models that predict a complexity index rather than just a binary label. This task is split into two sub-tasks, the first one is about the prediction of the complexity of single words and the second is about the prediction of the complexity of multi word expressions. In this work we present our submission to both the sub-tasks with a model that aggregates a large set of heterogeneous features that can capture a wide variety of linguistics aspects (morphological, semantic, distributional and lexicon-based) in a regression model based on Gradient Boosting. …

Content truncated to 500 words