Lexical Complexity Prediction using Deep Learning with Sentence and Word Features

Lexical complexity predictions are one of the critical Natural Language Understanding (NLU) tasks to be solved. Currently, the best approach to this task is to use contextual pre-trained embeddings and deep learning. In order to get better performance, this research focuses on adding word and sentence features to the contextual pre-trained embedding-based model and ensemble the model using a stacking method. Based on the experiments, using the word and sentence features is proven to improve the model’s performance and the stacking results. The results of the best stacking model managed to rank first in subtask 1 with a Pearson value of 0.7887. In subtask 2, we managed to rank second with a Pearson score of 0.8590. Based on further analysis, the characteristics of the built model tend to predict the complexity of words or phrases that are rarely used higher than words or phrases that are often used.