Overview of ABSAPT at IberLEF 2024: Overview of the Task on Aspect-Based Sentiment Analysis in Portuguese

Alexandre Thurow Bender, Gabriel A. Gomes, Emerson P. Lopes, Ricardo M. Araujo, Larissa A. de Freitas, Ulisses B. Corrêa

Resumen


This paper reports the results of a competition on Aspect-Based Sentiment Analysis in Portuguese at ABSAPT 2024. The ABCD Team participated in the Aspect Extraction sub-task, using a BIO tagging scheme and Transformer-based models, with the best performance from the BERTimbau Large model. The teams ABCD and UTFPR participated in the Aspect Sentiment Classification sub-task. The UTFPR Team achieved the best results by fine-tuning a BERT model and using data augmented by ChatGPT to expand the training set. The ABCD Team experimented with both generation and classification approaches, with the classification method performing best. This competition showed that Transformer models and data augmentation are interesting techniques for improving Aspect-Based Sentiment Analysis in Portuguese.

Texto completo:

PDF