This paper presents a methodology based on LSTM neural networks to perform sentiment analysis and classify tourist locations in written reviews about Mexico's Pueblos Mágicos. The natural language processing approach involves cleaning Spanish text, tokenizing it, and encoding both geographic and typological labels. The model uses a recurrent neural network architecture composed of embedding, LSTM, and dropout layers to predict three key variables from the reviews: sentiment polarity, the name of the Pueblo Mágico, and the category of the referenced tourist location. We trained and evaluated the model using a dataset provided by the Rest-Mex 2025 competition. The results demonstrate that LSTM networks have a strong potential to capture sequential dependencies in Spanish-language text.
Mexico’s Pueblos Mágicos represent a highly relevant tourist and cultural initiative by highlighting
towns with unique historical, architectural, or natural attributes [
This work proposes a methodology that combines Natural Language Processing (NLP) techniques and deep learning to analyze reviews written in Spanish and classify both the expressed sentiment and the location and type of tourist attraction mentioned. To achieve this, we use a neural network architecture based on Long Short-Term Memory (LSTM) [9, 10].
We applied the developed methodology to the dataset from the Rest-Mex 2025 competition [11, 12], which contains tourist reviews with varying tones and contents. Rest-Mex in a international shared task related to NLP over turist spanish texts [13, 14, 15]. After preprocessing, tokenization, and training, we built three independent models to predict opinion polarity, the type of the mentioned Pueblo Mágico, and the type of referenced location. The results confirm that deep neural networks can efectively extract valuable knowledge from unstructured data such as user comments.
Dataset
Description Training Set Size: 208,051 instances (70% of original dataset).
Columns: – Title: Title assigned by tourist (Text). – Review: Full review text (Text). – Polarity: Sentiment polarity (1 to 5). – Town: Town referred in review (Text). – Region: Mexican state of the town (Text). Not used for classification.
– Type: Place category (Hotel, Restaurant, Attractive).
Test Set
Size: 89,166 opinions (30% of original dataset).
Columns: – ID: Unique identifier (Integer). – Title: Title assigned by tourist (Text). – Review: Full review text (Text).
Contains data for polarity prediction.
The methodology used Natural Language Processing (NLP) and deep learning techniques to address three simultaneous tasks: opinion polarity classification, identification of the mentioned Pueblo Mágico, and classification of the type of tourist place. For this purpose, we used a dataset of Spanish tourist reviews provided by the Rest-Mex 2025 competition [11, 12]. Table 1 summarizes the main characteristics of the training and test datasets used in this work, including the number of instances and the available variables in each.
In the text preprocessing, we applied several cleaning and normalization steps to prepare the data, including: • Converting text to lowercase. • Removing URLs, special characters, and punctuation marks.
• Removing Spanish stopwords to reduce lexical noise.
We applied these transformations to both the training and test datasets. On the other hand, for tokenization and sequencing, we employed Keras’ Tokenizer class, configured to consider the 10,000 most frequent words. We assigned a special token for out-of-vocabulary words. We converted the cleaned texts into numerical sequences and normalized them using padding, ensuring all sequences had a fixed length of 100 tokens, which standardized input length for the model.
We then encoded the labels for the three tasks (polarity, town, and place type) into a numeric format using Scikit-learn’s LabelEncoder, ensuring compatibility with the categorical loss functions used during training. We formulated each task as a multiclass classification problem. We developed three independent LSTM-based models, one for each task, with the following structure: • Input layer receiving fixed-length sequences. • Embedding layer converting each token into a 128-dimensional vector. • LSTM layer with 64 units designed to capture temporal dependencies in the sequence. • Dropout layer with a rate of 0.5 to prevent overfitting.
• Dense layer with softmax activation to predict the corresponding class.
Each model was compiled using the categorical cross-entropy loss function (categorical_crossentropy) and optimized them with the Adam algorithm. Each model was trained for two epochs with a batch size of 32, reserving 20% of the training data for validation. Preliminary tests showed that the validation loss stabilized quickly, suggesting that most of the learning occurred in the early stages of training. Nevertheless, exploring longer training schedules remains a direction for future work. After training, each model generated predictions on the test set. We then post-processed the numerical codes back to their original labels (town names and place types). Finally, we exported the results to a text file in the format required by the competition, listing for each review the ID, polarity, Pueblo Mágico, and place type.
The task organizers evaluated the submitted predictions using standard metrics, including precision, recall, and F1. They evaluated each of the three tasks independently using macro-averaged scores. The following describes how they evaluated each task: • For polarity classification, which involves five sentiment classes ( = 1, 2, 3, 4, 5), the organizers computed the macro-averaged F1 across all classes, as shown in Equation 1, where () represents the F1 for class produced by system . • For type prediction, which includes three categories (Attractive, Hotel, and Restaurant), the organizers calculated the average of the individual F1 for each class, as expressed in Equation 2. • For Magical Town (MT) classification, the organizers used a predefined list of Magical Towns (MTL). They computed the macro F1 across all towns, following a similar approach to type prediction, as shown in Equation 3.
The results were submitted under the team name BRIAN_1, which presented the performance of the proposed methodology only once. The team achieved 51st place in the overall results table with a global P() = ∑︀∈ ()
|| T() = () + () + ()
3 MT() = 1 | |
∑︁ MTL () | | =1
Finally, the organizers combined the scores of the three subtasks using diferent weights to compute each system’s global score. Since they considered polarity and Magical Town classifications more critical, they assigned these tasks two and three times the weight of the type classification, respectively. Equation 4 shows how they computed the final score: () = 2 · P() + 3 · MT() + T() 6
For greater clarity, Figure 1 presents the pseudocode summarizing the overall system flow, from data loading to the generation of the output file. (1) (2) (3) (4) score of 0.3396 on the Track Score metric, as shown in Table 2. The evaluators based their assessment on multi-class classification metrics, specifically Macro F 1, precision, recall, and accuracy, analyzing each task separately:
In the polarity prediction task (ranging from 1 to 5), the model achieved modest performance (see Table 3). It recognized middle polarities better but struggled with the extreme classes (positive and negative). Class imbalance and the limited number of training epochs likely caused these dificulties. Table 4 details the F1 for each class.
In the classification of place type (attraction, restaurant, or hotel), the model achieved robust results, as shown in Table 5. The model efectively distinguished among the categories to determine the type of place mentioned in the reviews, likely due to more evident linguistic patterns and better representation of these classes in the data. The F1 per category are displayed in Table 6.
The most challenging task was predicting the specific Magical Town mentioned in the text due to the large number of classes (over 50). Nevertheless, the model achieved acceptable results (see Table 7). A subset of towns achieved higher F1, as detailed in Table 8. Other towns, especially those with fewer data representations, had F1 scores close to zero.
The results obtained in the Rest-Mex 2025 competition demonstrate that the proposed LSTM model performs competitively in general classification tasks, particularly in detecting the type of tourist place mentioned in Spanish reviews. The precision and F1 achieved in this task show that the implemented architecture efectively captures relevant natural language patterns to identify categories such as attraction, hotel, or restaurant.
However, the model demonstrated limited performance in more complex tasks, such as sentiment polarity prediction and identifying Magical Town. The metrics reflect dificulties in distinguishing between extreme polarity classes (very positive or very negative) and in accurately recognizing a large variety of diferent locations.
One issue that negatively impacted the results was an error in preprocessing the polarity labels: the code transformed the original range from 1 to 5 into a range from 0 to 4. Although this modification suits some models, it did not align with the oficial evaluation format, which contributed to the lower performance in this task. Since this was the first and only submission, the results can serve as an initial baseline for building significant improvements in future system versions.
Based on these findings, this paper proposed the following directions for future work: • Correct the handling of polarity labels by preserving the original range (1 to 5) as required by the evaluation. • Increase the number of training epochs to allow the model to learn more complex patterns and avoid underfitting or overfitting. • Apply class balancing techniques, such as oversampling minority classes or weighted loss functions, to improve the representation of underrepresented classes. • Explore multitask learning architectures that enable the model to learn related tasks jointly, such as polarity and place type, which could enhance generalization. • Use pre-trained Spanish language models like BERT or BETO, which have demonstrated superior performance on NLP tasks compared to models trained from scratch. • Expand preprocessing and enrich the corpus by incorporating lemmatization, named entity recognition (NER), and contextual analysis to improve semantic understanding of the text. • These improvements could significantly enhance the robustness of the proposed methodology and enhance its applicability in real-world automated tourism analysis scenarios.
We declare that the present manuscript has been written entirely by the authors and that no generative artificial intelligence tools were used in its preparation, drafting, or editing. [9] S. Hochreiter, J. Schmidhuber, Long short-term memory, Neural Computation 9 (1997) 1735–1780.
doi:10.1162/neco.1997.9.8.1735. [10] Y. Zhang, B. C. Wallace, A sensitivity analysis of (and practitioners’ guide to) convolutional neural networks for sentence classification, CoRR abs/1510.03820 (2015). URL: http://arxiv.org/abs/1510. 03820. arXiv:1510.03820. [11] M. Á. Álvarez-Carmona, Á. Díaz-Pacheco, R. Aranda, A. Y. Rodríguez-González, L. Bustio-Martínez, V. Herrera-Semenets, Overview of rest-mex at iberlef 2025: Researching sentiment evaluation in text for mexican magical towns, volume 75, 2025. [12] J. Á. González-Barba, L. Chiruzzo, S. M. Jiménez-Zafra, Overview of IberLEF 2025: Natural Language Processing Challenges for Spanish and other Iberian Languages, in: Proceedings of the Iberian Languages Evaluation Forum (IberLEF 2025), co-located with the 41st Conference of the Spanish Society for Natural Language Processing (SEPLN 2025), CEUR-WS. org, 2025. [13] M. Á. Álvarez-Carmona, R. Aranda, S. Arce-Cárdenas, D. Fajardo-Delgado, R. Guerrero-Rodríguez, A. P. López-Monroy, J. Martínez-Miranda, H. Pérez-Espinosa, A. Rodríguez-González, Overview of rest-mex at iberlef 2021: Recommendation system for text mexican tourism, Procesamiento del Lenguaje Natural 67 (2021). doi:https://doi.org/10.26342/2021-67-14. [14] M. Á. Álvarez-Carmona, Á. Díaz-Pacheco, R. Aranda, A. Y. Rodríguez-González, D. Fajardo-Delgado, R. Guerrero-Rodríguez, L. Bustio-Martínez, Overview of rest-mex at iberlef 2022: Recommendation system, sentiment analysis and covid semaphore prediction for mexican tourist texts, Procesamiento del Lenguaje Natural 69 (2022). [15] M. Á. Álvarez-Carmona, Á. Díaz-Pacheco, R. Aranda, A. Y. Rodríguez-González, L. Bustio-Martínez, V. Muñis-Sánchez, A. P. Pastor-López, F. Sánchez-Vega, Overview of rest-mex at iberlef 2023: Research on sentiment analysis task for mexican tourist texts, Procesamiento del Lenguaje Natural 71 (2023).