Identifying which words present in a text may be difficult to understand by common readers is a well-known subtask in text complexity analysis. The advent of deep language models has also established the new state-of-the-art in this task by means of end-to-end semi-supervised (pre-trained) and downstream training of, mainly, transformer-based neural networks. Nevertheless, the usefulness of traditional linguistic features in combination with neural encodings is worth exploring, as the computational cost needed for training and running such networks is becoming more and more relevant with energy-saving constraints. This study explores lexical complexity prediction (LCP) by combining pre-trained and adjusted transformer networks with different types of traditional linguistic features. We apply these features over classical machine learning classifiers. Our best results are obtained by applying Support Vector Machines on an English corpus in an LCP task solved as a regression problem. The results show that linguistic features can be useful in LCP tasks and may improve the performance of deep learning systems.

In this work, we present a novel approach to lexical complexity prediction (LCP) that combines diverse linguistic features with encodings from deep neural networks. We explore the integration of 23 handcrafted linguistic features with embeddings from two well-known language models: BERT and XLM-RoBERTa. Our method concatenates these features before inputting them into various machine learning algorithms, including SVM, Random Forest, and fine-tuned transformer models. We evaluate our approach using two datasets: CompLex for English (a high-resource language) and CLexIS2 for Spanish (a relatively low-resource language in ), allowing us to study performance issues from a cross-lingual perspective. Our experiments involve different combinations of linguistic features with encodings from pretrained deep learning models, testing both token-based and sequence-related encodings. The results demonstrate the effectiveness of our hybrid approach. For the English CompLex corpus, our best model achieved a mean absolute error (MAE) of 0.0683, representing a 29.2% improvement over using linguistic features alone (MAE 0.0965). On the Spanish CLexIS2 corpus, we achieved an MAE of 0.1323, a 19.4. These findings show that handcrafted linguistic features play a fundamental role in achieving higher performance, particularly when combined with deep learning approaches. Our work suggests that hybrid approaches should be considered over full end-to-end solutions for LCP tasks, especially in multilingual contexts.

Evaluar la complejidad léxica en documentos utilizando los modelos Generative Pre-trained Transformer (GPT-4, GPT-4o y GPT-4 Turbo) permitió analizar su impacto en la comprensión del lenguaje específicamente en textos estatales ecuatorianos en español. Se aplicó la técnica de few-shot learning en todos los modelos, tomando el conjunto de datos GovAIEc. El enfoque aplicado de la investigación es cuantitativo y descriptivo, con un diseño cuasi experimental. Los modelos se evalúan mediante métricas del error común: MAE, MSE, RMSE. El objetivo de esta investigación es evaluar el rendimiento del modelo GPT-4 y sus modelos sucesores en la predicción de la complejidad léxica en textos públicos ecuatorianos mediante Few-Shot Learning. Los resultados indican que el modelo GPT-4 obtiene un mayor rendimiento frente a sus sucesores con un MAE = 0.2464, mientras que GPT-4o presenta un MAE = 0.3889, y finalmente los resultados aplicando el modelo GPT-4 Turbo dieron un MAE = 0.2540.