Combining romance languages for semantic representation of portuguese texts

• Main Author: Silva, Hígor Uélinton da
• Publication Date: 2025
• Format: Master thesis
• Language: eng
• Source: Biblioteca Digital de Teses e Dissertações da PUC_RS
• Download full: https://tede2.pucrs.br/tede2/handle/tede/11797
• topic: Sentence Representation; Semantic Textual Similarity; Portuguese; Family Languages; Natural Language Processing; Representação de Sentenças; Similaridade Textual Semântica; Português; Família de Línguas; Processamento de Linguagem Natural; CIENCIA DA COMPUTACAO::TEORIA DA COMPUTACAO

With the advancement of technology, there is a growing need for machines to understand and process human language effectively. This requires capturing the semantic value of natural language, accurately identifying and interpreting diferente meanings. The lack of annotated data for Portuguese restricts model performance compared to high-resource languages. To address this, we leverage the similarities among Romance languages to augment existing data and develop a supervised model that better represents the semantics of Portuguese texts. In the STS task, which aims to determine how similar two texts are, we explore existing models and develop our own, fine-tuning and evaluating them using different data combinations. Our findings indicate that Spanish and Italian are the most suitable languages to pair with Portuguese, yielding results comparable to the state of the art. Our approach achieves results comparable to models such as BERTimbau and Jina, reaching 0.855 for Pearson and 0.317 for MSE.