Integration of new generation embeddings and current linguistic resources to identify complex words in Spanish with machine learning

Abstract

The complexity of words can pose a limitation to the accessibility of information, which could affect millions of Spanish-speaking people. The objective of this study is to develop a machine learning model for the binary task of identifying complex words in Spanish, using next-generation embeddings, current linguistic resources, and lexical properties. To this end, the Spanish dataset from the CWI Shared Task 2018 was used, obtaining embeddings generated by the text-embedding-3-large model and word frequencies extracted from resources such as the Corpus del Español del Siglo XXI, the Corpus de Referencia del Español Actual, the Spanish Billion Word Corpus and Embeddings, and Wordfreq. To select features and find their best combination, a 5-fold cross-validation using XGBClassifier was employed. After comparing several machine learning algorithms, the final model, based on LGBMClassifier, achieved a macro F1 score of 0.7993, surpassing the best team from that competition, more recent studies that used neural networks, and some large language models. This demonstrates the potential of these resources that are constantly being updated and that can contribute to improving the accuracy of this task.