REST-MEX 2026 introduces a fully generative evaluation paradigm for tourism-related natural language processing in Spanish. Unlike previous editions, this task does not evaluate systems by directly classifying real tourist reviews. Instead, the challenge focuses on the generation of synthetic tourist opinions and on measuring their effectiveness for downstream sentiment analysis.

The problem is defined as follows:

"Given a limited collection of real tourist reviews written in Spanish, each participating team must develop a generative model capable of producing a synthetic corpus of 3,000 original tourist opinions related to Mexican Pueblos Mágicos."

Each generated review must:

• Be entirely original (no repetition or paraphrasing of the provided real reviews),

• Reflect realistic tourist narrative styles, descriptive richness, and communicative intentions,

• Include an explicit sentiment polarity label ranging from 1 (very negative) to 5 (very positive), assigned by the generative system itself.

Indirect Evaluation Setting

The core scientific challenge of REST-MEX 2026 lies in its indirect evaluation strategy.

Rather than assessing the surface quality of the generated texts, the organizers will:

1. Train a sentiment classification model exclusively on the synthetic reviews submitted by each team.

2. Evaluate the resulting classifier on a hidden set of real tourist opinions from Mexican Pueblos Mágicos.

3. Measure performance using a weighted macro F1-based metric, designed to emphasize minority sentiment classes.
   
   

Under this framework, a synthetic dataset is considered effective only if a classifier trained solely on artificial data generalizes well to real-world reviews.

Key Challenges

This task poses several significant research challenges:

• Distributional alignment between synthetic and real tourist reviews, particularly under strong sentiment imbalance.

• Narrative diversity, avoiding generic, repetitive, or stylistically uniform generations.

• Faithful sentiment control, ensuring that generated texts genuinely reflect the assigned polarity labels.

• Data efficiency, as only synthetic data may be used for training the classifier.

• Generalization, testing whether synthetic corpora can meaningfully support real-world sentiment analysis in Spanish.
  
  

REST-MEX 2026 serves as an experimental testbed to explore how generative artificial intelligence can replace or complement real annotated data, advancing research on synthetic data generation, sentiment modeling, and tourism analytics in Spanish

Example

Synthetic Review (generated by a system):

“Aunque el lugar tiene potencial, la atención fue deficiente y varias áreas se encontraban descuidadas. La experiencia no cumplió con nuestras expectativas y no regresaríamos.”

Assigned Polarity:

• 2 (Negative)