Towards analysis on textual inference at ASSIN-2 dataset
Felipe O. da Silva, Giovana Meloni Craveiro, Vinícius F. da Silva, Vinícius João de Barros Vanzin · 2023
In this article, we conduct a preliminary analysis of different methods to address the Textual Entailment Recognition (RTE) task in Portuguese. We use the ASSIN-2 dataset as a benchmark to evaluate our models. Our work combines various textual representation approaches, including bag of words and word embeddings, with machine learning models. Additionally, we present a rule-based approach. Our highest performance was achieved by the BERTimbau-large model fine-tuned on ASSIN-2, which attained an F1 score of 0.89%, positioning it just 1% below the current state-of-the-art. Our ongoing experiment aims to combine our different approaches to leverage their full potential.