Overview of Rest-Mex at IberLEF 2026: Research on Synthetic Text Modeling for Mexican Magical Towns

Resumen

This paper presents an overview of the REST-MEX 2026 shared task, organized as part of IberLEF 2026. Unlike previous editions, which focused on sentiment classification, REST-MEX 2026 introduced a new evaluation framework centered on synthetic tourism review generation. Participants generated synthetic reviews in Mexican Spanish to improve a domain-adapted RoBERTa sentiment classifier through fine-tuning. All submitted corpora were evaluated on the same hidden test set of authentic reviews using a weighted Macro-F1 metric that emphasizes minority sentiment classes. The shared task attracted 28 participating systems exploring diverse approaches, including prompt engineering, supervised fine-tuning, prototype retrieval, multi-agent architectures, diffusion language models, lexical augmentation, bio-inspired methods, and automatic filtering strategies. The results show that well-designed synthetic corpora can substantially improve downstream sentiment classification, with several systems outperforming the baseline model trained without synthetic data.