This article details our participation in REST-MEX 2025, an evaluation task focused on sentiment analysis of tourism reviews about Pueblos Mágicos in Mexico. We addressed three subtasks: sentiment polarity classification, destination type identification, and Pueblo Mágico recognition. For the sentiment polarity and Pueblo Mágico recognition tasks, we employed fine-tuned BERT-based Transformer models. For the destination type identification, we explored an approach based on Word2Vec embeddings classified by a Multilayer Perceptron (MLP). We present a data analysis, a detailed methodology for each task-including implementation and fine-tuning/training procedures-the results obtained on the oficial dataset, and the conclusions drawn from our participation.
In this paper, we present our system developed for REST-MEX 2025. Our architecture leverages pre-trained Transformer models—specifically BERT—for the Polarity and Town classification tasks, and a custom-designed Artificial Neural Network for the Type classification task [ 12]. We describe the technical details of our approach, report on the experiments conducted, and analyze the performance achieved on the oficial competition dataset.
Sentiment analysis in Spanish has gained increasing attention in recent years, particularly within the tourism domain. Several corpora and shared tasks have been developed to support this line of research, including TASS [13] and REST-MEX [14, 15, 16]. Studies by Guerrero-Rodríguez et al. [17] and Álvarez-Carmona et al. [15] have shown that ensemble methods combining multiple classifiers enhance polarity-detection performance on user-generated content from platforms such as TripAdvisor.
The introduction of Transformer-based architectures, especially BERT, has significantly improved performance over classical models in opinion-mining tasks. Spanish-adapted variants such as BETO [5] have demonstrated robust results when fine-tuned for polarity and topic classification. In RESTMEX 2021, the top-ranking system employed BETO fine-tuned to predict star ratings from Mexican tourism reviews [14]. Subsequent editions in 2022 and 2023 confirmed this trend, with the majority of participating systems relying on BERT variants, thereby consolidating the dominance of Transformerbased models in this task domain [15, 16].
The REST-MEX shared tasks have evolved across editions. In 2021, tasks included site recommendation and sentiment-polarity classification; the leading system fine-tuned BETO to predict polarity on a ifve-point scale [ 14]. In 2022, a new task—classification of destination type (hotel, restaurant, attraction)—was introduced. The winning system integrated handcrafted linguistic features (UMUTextStats) with BERT, achieving a Macro-F1 score of 0.89 [18]. Another competitive approach involved preprocessing techniques such as translation and data cleaning prior to BERT training [15]. REST-MEX 2023 further expanded the task set by incorporating country identification; results indicated that predicting destination type and country (F1 ≈ 0.99 and ≈ 0.93, respectively) was less complex than predicting sentiment polarity [16].
Overall, the REST-MEX track at IberLEF has significantly contributed to advancing sentiment analysis in Spanish, particularly in tourism-related contexts. These shared tasks have reinforced the efectiveness and widespread adoption of Transformer-based architectures for multilingual and domain-specific NLP applications.
In this section we outline the theoretical and contextual foundations that support our system: (i) the Transformer architecture and its BERT adaptation; (ii) Spanish-trained BERT variants, with emphasis on BETO; (iii) the fundamentals of the Multilayer Perceptron (MLP); and (iv) the relevance of the Pueblos Mágicos program to Mexican tourism.
The Transformer introduced a self-attention mechanism that replaces traditional recurrence and enables the modeling of long-range dependencies with full training parallelism [19]. 1
BERT (Bidirectional Encoder Representations from Transformers) extends this architecture through self-supervised pre-training based on Masked Language Modeling (MLM) and Next Sentence Prediction (NSP), producing bidirectional contextual representations [12]. In numerous text-classification challenges—including TASS and REST-MEX—BERT systematically outperforms SVM, LSTM, or CNN 1The draft for this section was generated with assistance from the AI tool Gemini and subsequently reviewed, edited, and validated by the authors. approaches when a moderately sized annotated corpus is available and task-specific fine-tuning is applied [14, 16].
Models such as BETO [5], MarIA [6], and DistilBETO narrow the performance gap with English by training on large Spanish-language corpora. These variants retain the base BERT architecture (bert-base, ≈ 110 M parameters) but employ Spanish WordPiece vocabularies and weights adapted to the language. Comparative studies in sentiment polarity (TASS, MEX-A3T, REST-MEX) report absolute macro-F1 improvements of 3–8 points over BERT, as documented by [17, 15, 20], among others.
The MLP is a feed-forward neural network composed of one or more dense layers with non-linear activation functions (e.g. ReLU). Its universal approximation power makes it a lightweight alternative—compared with complex Transformers—when the input is already a fixed vector representation (TF–IDF, averaged embeddings, etc.) and the output involves a large number of classes [21]. For the Type identification task (3 labels), an MLP provides: • Eficiency : fast training and inference even without a high-end GPU; • Simple regularization: techniques such as Dropout or L2 mitigate overfitting under class imbalance; • Flexibility: ability to combine lexical features with available metadata.
Sentiment analysis has become a strategic tool for understanding tourist perceptions and guiding decision-making in emerging destinations. In 2023, tourism accounted for 8.6 % of Mexico’s GDP, and the Pueblos Mágicos2 program has been instrumental in diversifying the ofer beyond beach resorts. As of January 2025, the oficial register lists 177 towns after the addition of 45 new destinations [9].
Spanish tourist reviews often exhibit colloquial language, regional dialect blends, and code-switching (Spanish–English), increasing preprocessing complexity. In REST-MEX 2025 these reviews are labelled with: Polarity (1–5 stars), Type (hotel, restaurant, attraction), and Town (40 Pueblos Mágicos in the corpus). Capturing the underlying semantics and emotional nuances motivates the use of contextual-attention models (BERT) for Polarity/Town, whereas the 3-class granularity of Type encourages a lighter, robust architecture (MLP) based on high-dimensional sparse vectors.
In summary, the combination of (i) deep BERT representations, (ii) Spanish-adapted variants such as BETO, and (iii) an eficient MLP for high-cardinality multi-class classification forms the backbone of our system for the three REST-MEX 2025 subtasks. The tourism context of the Pueblos Mágicos not only lends practical relevance to the research but also introduces linguistic and class-imbalance challenges that justify the proposed hybrid strategy.
This section provides a quantitative and qualitative overview of the dataset used in REST-MEX 2025, including its structure, class distributions, and textual characteristics, along with a summary of the preprocessing steps. 2Federal initiative launched in 2001 by the Secretaría de Turismo (SECTUR) to promote towns with high historical and cultural value.
69921 51410
Hotel Restaurant
At ractive
Type Distribution of the ’Town’ Variable (Top 20) 85993
The oficial REST-MEX 2025 dataset consists of a total of 297,218 Spanish-language tourism reviews,
divided into a training set of 208,052 examples and a test set of 89,166. Each instance includes a
short Title, a longer Review, and three classification labels: Polarity (
Distribution of the ’Type’ Variable
Distribution of the ’Polarity’ Variable 86720 136561 Sentiment Polarity. The class distribution is strongly skewed towards positive opinions. As shown in Figure 1, more than 65% of the reviews are labeled with 5 stars, while 1- and 2-star reviews collectively account for less than 10% of the total. The mean polarity is 4.45, with a median of 5. Destination Type. The Type label exhibits moderate imbalance: Restaurants dominate (86,720 reviews), followed by Attractions (69,921) and Hotels (51,410). This may reflect user interest bias or platform review frequency.
Town. A long-tailed distribution is observed for the Town variable. The most reviewed town is Tulum (45,345), followed by Isla Mujeres (29,826) and San Cristóbal de las Casas (13,060). On the opposite end, Tapalpa and Real de Catorce have fewer than 800 reviews each. This variability suggests challenges for classification models in minority towns. Sa0n_TCurliIusstmloab_aMl_udjeer_elsas_CasasValadolid BacalarPalenqueSaVyaulelit_ade_BraTveootihuacan LorTeotdoosSantosPatzcuaro TaTxlcaoquepaque TAejqijuicisquiapanMetepecTepoztlan Cholula Tequila
Town 23532 19439 Region. Although not used as a target label, the Region field is included as metadata. Reviews are unevenly distributed across states, with Quintana Roo (85,993) and Chiapas (23,532) representing the largest shares.
The reviews range widely in length. Based on tokenization statistics: • Mean length: 63.4 tokens • Median: 45 tokens • 95th percentile: ∼ 180 tokens
Most texts are short and highly subjective, often containing informal language, emojis, emphasis with repeated characters, and occasional code-switching to English. These traits introduce noise and motivate careful preprocessing and robust tokenization strategies.
The original Title and Review fields were merged into a single column ( read_text). We then applied a standard cleaning routine: 1. Mojibake correction using a custom decoding function. 2. Lower-casing and Unicode NFD normalization to remove diacritics. 3. Noise filtering: removal of non-alphabetic characters using regex. 4. Stop-word removal using the NLTK Spanish stop-word list. 5. Resulting tokens were re-joined and stored in a new column clean_text.
The cleaned datasets were saved as train_clean.csv and test_clean.csv, and used for all downstream experiments.
The strong polarity imbalance led us to explore weighted loss functions. For the Type task, the longtailed label distribution and varying text lengths informed the choice of a lightweight architecture with regularization mechanisms to avoid overfitting on rare classes.
This section outlines the methodological approaches implemented to address the three subtasks of the REST-MEX 2025 challenge: sentiment polarity classification ( Polarity), prediction of the type of destination mentioned (Type), and identification of the specific Pueblo Mágico (Town).
For the Polarity and Town tasks, we adopted advanced Transformer-based models, specifically BETO—a Spanish-pretrained variant of BERT—capable of capturing the linguistic intricacies of the domain more efectively.
In contrast, the Type classification task posed lower complexity and involved a balanced label set. Therefore, we opted for a classical machine learning approach using a Multilayer Perceptron (MLP) as the primary model.
The objective of this task is to assign a sentiment label ranging from 1 (very negative) to 5 (very positive) to each tourist review. We implemented a pre-trained Transformer model and fine-tuned it on the task-specific training data provided in the competition. 5.1.1. BERT The core model employed for this task was BERT, specifically the ‘dccuchile/bert-base-spanish-wwmcased‘ variant. This model is pre-trained on a large Spanish-language corpus using Whole Word Masking as its training strategy. Its selection was motivated by the model’s established performance across a range of Spanish-language NLP tasks.
The architecture retains the standard BERT-base configuration: 12 encoder layers, 768 hidden units per layer, 12 attention heads per self-attention mechanism, and approximately 110 million pre-trained parameters.
Leveraging BERT’s self-attention layers allows the model to capture complex contextual relationships between words, which is essential for accurately identifying sentiment in tourism reviews. A linear classification layer was added on top of the [CLS] token representation from the final hidden layer to perform sequence-level classification. This layer was trained to map the contextualized representation to one of five polarity classes.
The oficial REST-MEX 2025 training set served as the starting point. Initial preprocessing addressed encoding issues in the ‘Title’ and ‘Review’ columns, which were concatenated into a unified input field named ‘Texto_Leido’. No aggressive preprocessing such as stopword or accent removal was applied, in order to preserve as much contextual information as possible.
The polarity labels, originally ranging from 1 to 5, were shifted to a 0–4 scale to meet the requirements of PyTorch’s CrossEntropyLoss. The dataset was partitioned into training and validation sets in an 80/20 split.
Tokenization was handled using the AutoTokenizer utility, with sequences truncated to a maximum length of 128 tokens. Shorter sequences were padded to ensure input uniformity.
Model training was conducted using Hugging Face’s Transformers library and its Trainer class. The main training parameters were configured as shown in Table 1:
The goal of this task was to identify the specific Pueblo Mágico referenced in each tourist review, selecting from the set of localities included in the REST-MEX 2025 database. Given the nature of this multi-class classification task with a high number of categories, a pre-trained Transformer model was again selected. 5.2.1. BERT The selected model for the Pueblo Mágico identification task was, once again, BETO. We used the same variant as in the Polarity task: ‘dccuchile/bert-base-spanish-wwm-cased‘. The underlying architecture remained unchanged from the previously described configuration.
We hypothesized that BETO’s capacity to model context and semantics would be advantageous in this setting, particularly in cases where location names are implied rather than explicitly mentioned. The model’s ability to extract relevant contextual cues from descriptions and characteristics was considered essential for this classification task.
For the final classification layer, a linear head was added on top of the [CLS] token representation from the final hidden layer. This layer was trained to map each review’s contextual embedding to one of the defined Town labels.
The dataset provided by REST-MEX 2025 was used as the input source. Data preprocessing followed
the same steps as in the Polarity task, with one key diference: encoding the target variable Town.
A dictionary was created to map each unique Pueblo Mágico name to a numerical identifier (
Tokenization was carried out using the AutoTokenizer from the ‘dccuchile/bert-base-spanishwwm-cased‘ model, with a maximum sequence length of 128 tokens. Sequences exceeding this limit were truncated, and shorter sequences were padded as needed.
Training was performed using the Hugging Face Transformers library and the Trainer class. To address class imbalance resulting from uneven mention frequencies among towns, a class weighting strategy was applied. We computed class-specific weights using compute_class_weight with the ‘balanced‘ option from scikit-learn. These weights were integrated into the loss function (CrossEntropyLoss) via a custom WeightedTrainer wrapper.
The main training parameters are summarized in Table 2.
For the Type task, which aims to identify whether a review refers to a “Hotel”, “Restaurant”, or “Attraction”, we explored an alternative approach to Transformer-based models. Specifically, we implemented a combination of Word2Vec vector representations and a Multilayer Perceptron (MLP) classifier.
The input representation for the MLP consisted of vectors generated via Word2Vec embeddings with a dimensionality of 100. The neural network was structured with three hidden layers comprising 256, 128, and 64 neurons respectively (‘hidden_layer_sizes=(256, 128, 64)‘). The ReLU (Rectified Linear Unit) activation function was used for all hidden layers due to its computational eficiency and its capacity to mitigate the vanishing gradient problem, enabling the model to learn complex, non-linear representations of the input data.
The MLP’s output layer was designed for multi-class classification, predicting one of the three destination types. A Softmax activation was applied at the output to produce class probabilities, with the model trained using a log-loss (cross-entropy) function in combination with the Adam optimizer.
This architecture was selected as an eficient and efective solution for this specific task, especially when paired with dense input representations generated via Word2Vec. The MLP was trained to learn the underlying patterns in the document vectors and to accurately classify the type of destination.
The implementation and training process consisted of two main stages: generating vector representations using Word2Vec and training the MLP classifier.
The MLP was trained using the scikit-learn library, while Word2Vec embeddings were generated with Gensim. Data manipulation was handled using pandas and NumPy.
As noted earlier, the Adam optimizer was employed. The MLPClassifier in scikit-learn was used, which implicitly applies log-loss (cross-entropy) as the default loss function. The batch size was automatically set as min(200, n_samples). The number of iterations (epochs) was capped at 300. L2 regularization was also applied with an alpha value of 1 × 10− 4 to reduce overfitting.
The trained MLP was subsequently used to classify the test set review vectors, generating the predictions for the Type classification task.
Word2Vec Representation: The Word2Vec model was trained using the parameters listed in Table 3.
This section presents the experimental results obtained. We report performance across the three subtasks: sentiment polarity classification, destination type identification, and recognition of specific Pueblos Mágicos. For each task, we provide detailed evaluation metrics—including precision, recall, and F1-score—on both the training and oficial test sets whenever applicable. The goal is to assess the generalization capabilities of our models under real-world conditions and to identify strengths and limitations in each approach.
On the test set, we observed a moderate drop in overall performance. The macro F1-score decreased to 0.6057 and accuracy fell to 73.75%. Nonetheless, the classifier preserved consistent patterns, maintaining high precision and recall for class 5 (F1 = 0.85), and competitive results for class 1. However, performance on classes 2 and 4 remained weak, with class 2 notably achieving an F1-score of only 0.39, confirming the dificulty of distinguishing mid-range sentiment levels in real-world reviews.
These results highlight the efectiveness of the model in capturing polar sentiment extremes, while suggesting the need for specialized handling of ambiguous or neutral cases. Future improvements could explore ordinal classification loss functions or contrastive learning approaches tailored to sentiment gradation.
Tables 6 and 7 present the classification performance for the Type prediction task, using only the training set and the complete dataset (train + test), respectively. As described in Section Methodology, this task was tackled with a Multilayer Perceptron (MLP) classifier fed with Word2Vec vector representations.
On the training set, the MLP achieved an overall accuracy of 96% and a macro F1-score of 0.95. The performance across all three destination types—Attractive, Hotel, and Restaurant—was consistently strong, with F1-scores ranging from 0.94 to 0.97. This indicates that the model efectively captured the semantic patterns associated with each category from the vectorized representations.
When evaluated on the complete dataset, which includes the test set, the model maintained high generalization performance. It achieved a macro F1-score of 0.9437 and an accuracy of 94.6%. Notably, the classes Attractive and Restaurant obtained the highest F1-scores (0.9565 and 0.9500, respectively), confirming the robustness of the model even on unseen data.
These results demonstrate the efectiveness of the Word2Vec + MLP pipeline in distinguishing between types of tourism destinations. The high precision, recall, and F1-scores in both training and test scenarios highlight the suitability of dense word embeddings combined with deep learning architectures for fine-grained text classification tasks in the tourism domain.
Tables 8 and 9 present the classification results for the Town prediction task on the training set and the oficial train/test split, respectively. This task posed a substantial challenge due to the high number of classes (up to 40 diferent Pueblos Mágicos) and severe class imbalance.
In the training set, the model achieved a macro F1-score of 0.70 and a weighted F1 of 0.74, with notable performance on well-represented towns such as Tulum (F1 = 0.84), Teotihuacan (F1 = 0.84), and San Cristóbal de las Casas (F1 = 0.72). Some towns with fewer training examples (e.g., Tapalpa, Dolores Hidalgo) showed moderate to low F1-scores, indicating data scarcity as a limiting factor.
When evaluated on the test set (Table 9), the model obtained a macro F1-score of 0.5958. Although performance dropped relative to the training set, the classifier still performed competitively for several towns. For instance, Tulum, Teotihuacan, and Chiapa de Corzo maintained strong F1-scores above 0.75, showcasing the model’s ability to generalize well in high-frequency cases. However, low recall values in underrepresented classes highlight the dificulty in managing class imbalance in real-world deployment.
Overall, these results suggest that while the model generalizes reasonably well for dominant towns, future work should consider strategies such as data augmentation or cost-sensitive learning to improve performance in minority classes.
This paper has presented the approach of the CIMAT CC team to the REST-MEX 2025 shared task, where we addressed three distinct NLP challenges within the domain of Mexican tourism: sentiment polarity classification, destination type identification, and Pueblo Mágico recognition. Our participation demonstrated the value of combining state-of-the-art Transformer architectures with well-established classical models, selecting the appropriate methodology according to the complexity and structure of each subtask.
For sentiment polarity classification, the fine-tuned BERT model achieved a macro F1-score of 0.6057 and an accuracy of 73.75% on the test set. These results confirmed the model’s strength in capturing semantic subtleties, particularly at the extremes of the sentiment scale. However, performance remained lower for mid-scale values, where expressions of sentiment tend to be more ambiguous. This observation suggests the need for models that explicitly account for ordinal relationships between classes.
In contrast, the destination type classification task was tackled using a Multilayer Perceptron (MLP) fed with Word2Vec embeddings. This classical approach yielded a macro F1-score of 0.9437 and an accuracy of 94.6%, indicating that dense vector representations coupled with a well-configured neural classifier remain highly efective in tasks with balanced data and clearly separable class boundaries.
For the Pueblo Mágico identification task, a BERT-based model was again employed. Despite the challenge posed by a high number of classes and strong class imbalance, the model attained a macro F1-score of 0.5958. Performance was highest for towns with greater representation in the dataset (e.g., “Tulum”, “Teotihuacan”), but declined for those with few training instances. This underscores the dificulty of high-cardinality classification under data scarcity.
Overall, our results suggest that while BERT models ofer robust solutions for semantically rich tasks, classical models such as MLPs can outperform on more structured, low-ambiguity tasks. The hybrid strategy employed here provides a flexible and efective framework for addressing heterogeneous NLP problems in domain-specific contexts.
Future work will focus on addressing class imbalance in the Town classification task, potentially through advanced data augmentation or loss re-weighting techniques such as focal loss. In the polarity classification task, incorporating ordinal regression or contrastive learning may help improve accuracy on intermediate sentiment levels. Finally, multi-task learning approaches that simultaneously optimize for all three tasks may ofer benefits through shared representation learning, enhancing generalization across related subtasks in the tourism review domain.
We thank the organizers of REST-MEX 2025 and IberLEF 2025 for coordinating the task and providing the dataset. We also highlight that undertaking this challenge aforded us a pleasant experience and valuable learning.
During the preparation of this work, we made limited use of generative AI tools—namely ChatGPT (OpenAI) and Gemini (Google)—for general text drafting, language refinement, and code assistance. All AI-generated material was subsequently reviewed, edited, and validated by us, and we take full responsibility for the final content. [2] B. Liu, Sentiment Analysis and Opinion Mining, volume 5 of Synthesis Lectures on Human Language
Technologies, Morgan & Claypool, 2012. [3] W. Medhat, A. Hassan, H. Korashy, Sentiment analysis algorithms and applications: A survey,
Ain Shams Engineering Journal 5 (2014) 1093–1113. doi:10.1016/j.asej.2014.04.011. [4] M. A. Álvarez-Carmona, R. Aranda, A. Y. Rodríguez-Gonzalez, D. Fajardo-Delgado, M. G. Sánchez, H. Pérez-Espinosa, J. Martínez-Miranda, R. Guerrero-Rodríguez, L. Bustio-Martínez, Ángel DíazPacheco, Natural language processing applied to tourism research: A systematic review and future research directions, Journal of King Saud University - Computer and Information Sciences 34 (2022) 10125–10144. URL: https://www.sciencedirect.com/science/article/pii/S1319157822003615. doi:https://doi.org/10.1016/j.jksuci.2022.10.010. [5] J. Cañete, G. Chaperón, R. Fuentes, J. Pérez, C. Bizarreta, Spanish pre-trained bert model and evaluation data, in: Proceedings of the PML4DC Workshop at ICLR 2020, 2020. [6] A. Gutiérrez-Fandiño, M. G. Pino, R. Pérez, et al., Maria: Spanish language models, corpora and benchmark, Procesamiento del Lenguaje Natural 68 (2022) 39–60. [7] A. Diaz-Pacheco, M. A. Álvarez-Carmona, R. Guerrero-Rodríguez, L. A. C. Chávez, A. Y. RodríguezGonzález, J. P. Ramírez-Silva, R. Aranda, Artificial intelligence methods to support the research of destination image in tourism. a systematic review, Journal of Experimental & Theoretical Artificial Intelligence 0 (2022) 1–31. doi:10.1080/0952813X.2022.2153276. [8] S. de Turismo de México, Informe de resultados del sector turismo 2023, https://www.gob.mx/ sectur/documentos/informe-turismo-2023, 2024. Accessed: 3 Jun 2025. [9] S. de Turismo de México, Programa Pueblos Mágicos: Listado oficial de localidades 2025, https: //www.gob.mx/sectur/acciones-y-programas/pueblos-magicos, 2025. Accedido: 3 jun 2025. [10] 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. [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. Devlin, M.-W. Chang, K. Lee, K. Toutanova, Bert: Pre-training of deep bidirectional transformers for language understanding, in: Proceedings of NAACL-HLT 2019, 2019, pp. 4171–4186. [13] E. Martínez-Cámara, J. Villena-Román, F. García-Sánchez, et al., Overview of tass 2020: Sentiment analysis and emotion detection in spanish, in: Proceedings of the SEPLN 2020 Workshop on TASS, 2020, pp. 13–27. [14] M.-A. Álvarez Carmona, et al., Rest-mex 2021: Resources and evaluations for tourism sentiment analysis in spanish, in: Proceedings of IberLEF 2021, 2021, pp. 638–651. [15] M.-A. Álvarez Carmona, V. Guzmán-Flores, L. González-Gurrola, G. Bel-Enguix, Overview of the rest-mex 2022 shared task on sentiment and tourism text classification, in: Proceedings of IberLEF 2022, 2022, pp. 684–703. [16] V. Guzmán-Flores, M.-A. Álvarez Carmona, et al., Rest-mex 2023: Adding country identification to tourism sentiment tasks, in: Proceedings of IberLEF 2023, 2023, pp. 712–728. [17] G. Guerrero-Rodríguez, K. Hernández-Figueroa, A. Sandoval-Sánchez, Combining ensembles for ifne-grained sentiment analysis of tripadvisor reviews, in: Proceedings of the TASS Workshop at SEPLN 2021, 2021, pp. 123–132. [18] L. García-Villalba, E. Martínez-Cámara, Umutextstats at rest-mex 2022: Combining linguistic statistics with bert for tourism sentiment tasks, in: Proceedings of IberLEF 2022, 2022, pp. 704–714. [19] A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. Kaiser, I. Polosukhin, Attention is all you need, in: Advances in Neural Information Processing Systems 30 (NeurIPS 2017), 2017, pp. 5998–6008. [20] Á. Díaz-Pacheco, R. Guerrero-Rodríguez, M. Á. Álvarez-Carmona, A. Y. Rodríguez-GonzÁlez, R. Aranda, A comprehensive deep learning approach for topic discovering and sentiment analysis of textual information in tourism, Journal of King Saud University - Computer and Information Sciences 35 (2023) 101746. URL: http://dx.doi.org/10.1016/j.jksuci.2023.101746. doi:10.1016/j. jksuci.2023.101746. [21] I. Goodfellow, Y. Bengio, A. Courville, Deep Learning, MIT Press, 2016. https://www. deeplearningbook.org.