Text-to-SQL systems allow users to query relational databases using natural language, but accuracy remains sensitive to the choice of language, model architecture, and prompting strategy. Although recent Large Language Models (LLMs) incorporate reasoning mechanisms that improve multi-step problem solving in other domains, their effects on multilingual Text-to-SQL are not yet well understood. This work evaluates a diverse set of LLMs on the BIRD benchmark and BIRD_PT, a Portuguese version produced by translating the questions and external knowledge while keeping the original English database schema and values unchanged. We compare four controlled scenarios that vary internal reasoning and guided reasoning for SQL generation. The results show a consistent decrease in accuracy when switching from English to Portuguese, with large variations in robustness across models. Reasoning alone does not reliably improve execution accuracy and can reduce performance in Portuguese, while combining reasoning with a guided plan provides the most stable improvements, although still weaker than in English. These findings highlight ongoing challenges in multilingual Text-to-SQL and emphasize the need to jointly consider language understanding, reasoning activation, and task-aligned planning when designing future systems.

This paper summarizes our work on the document classification subtask of Multilingual protest news detection of the CASE @ ACL-IJCNLP 2022 workshok. In this context, we investigate the performance of monolingual and multilingual transformer-based models in low data resources, taking Portuguese as an example and evaluating language models on document classification. Our approach became the winning solution in Portuguese document classification achieving 0.8007 F1 Score on Test set. The experimental results demonstrate that multilingual models achieve best results in scenarios with few dataset samples of specific language, because we can train models using datasets from other languages of the same task and domain.