Natural Language Inference for Portuguese Using BERT and Multilingual Information

Recognizing Textual Entailment, also known as inference recognition, aims to identify when the meaning of a piece of text contains the meaning of another fragment of text. In this work, we investigate multiples approaches for recognizing inference in the ASSIN dataset, an entailment recognition corpus for Portuguese. We also investigate the consequences of adding external data to improve training in two different forms: multilingual data and automatically translated corpus. Our results outperform, using the multilingual pre-trained BERT model, the current state-of-the-art for the ASSIN corpus. Finally, we show that using external data did not improve the performance of the model or the improvements are not significant.