|Título:||Detecting and Classifying Sexism by Ensembling Transformers Models|
This work presents the system with the highest results in terms of f1-score for tasks 1 and 2 of EXIST 2022. It is a challenge formed of two tasks, aimed at identifying and categorizing sexism in texts, respectively. First of all, a review of language models in Spanish and English is carried out, identifying the best performing models in each language. Also, a review of similar tasks and challenges (sexism detection in texts) is done. Then, models are trained in two phases. The first phase is for selecting the best hyperparameters for each model, while in the second phase these hyperparameters are used to learn with more training data. Finally, a simple ensembling strategy is used, which takes into account the performance of each model over a small validation set. This is compared against building a pure Transformers Ensemble, showing that the simple ensembling strategy obtains higher results. This leaves for future work the task of making such Ensembles work at least as good as the naive ensembling strategy.
Si te interesa esta publicación, puedes descargarla: