This study addresses balanced text classification in the Mexican tourism context using Natural Language Processing (NLP) and custom sampling strategies. Based on the Rest-Mex 2025 corpus, which contains over 200,000 annotated reviews, supervised models were trained to predict sentiment polarity, business type (hotel, restaurant, or attraction), and geographic location (state and municipality). A two-phase class balancing strategy was implemented: redundancy reduction in overrepresented classes using similarity metrics (Jaccard and Fuzzy), and synthetic data generation for underrepresented classes using controlled domain-specific vocabularies. The text preprocessing pipeline included normalization, lemmatization, and formatting for FastText compatibility. Trained models demonstrated strong performance for business type classification (F1=0.9687), moderate performance for geographic location prediction (F1=0.6397), and more challenging results for sentiment analysis (F1=0.4403). This approach enhances fairness and applicability in multi-label, multilingual AI models for tourism-related text analysis and supports downstream tasks such as recommendation systems and public policy insights.
In the field of artificial intelligence (AI), the performance of supervised learning models largely depends
on the balance and quality of the training data. One of the most persistent challenges is data imbalance,
where certain classes are significantly overrepresented or underrepresented compared to others. This
issue is particularly critical when developing models intended for classification tasks across multiple
categories and regions [
This work focuses on building a text classification model capable of analyzing tourist-related content
and assigning it a polarity score from 1 to 5. The model also predicts whether the text refers to a
restaurant, hotel, or tourist attraction, and determines its location among 40 Mexican states and their
municipalities [
To mitigate the imbalance, we employed a two-fold strategy. When overrepresented classes were identified, we applied a Jaccard distance metric to identify and randomly remove redundant samples until reaching the class mean. In contrast, for underrepresented classes, we leveraged preprocessed text data (already tokenized, lemmatized, and stripped of stop words, numbers, and punctuation) and expanded it by including contextually related words of similar category and length, supported by a custom vocabulary bank aimed at secure learning.
This problem is relevant in real-world AI applications where data collection does not always ensure
class uniformity. In our case, the tourism domain sufers from such imbalance, potentially biasing
models toward majority classes and limiting their usability in diverse environments[
Preliminary results show promising performance, including confusion matrices and predictions on a ifnal test set of 89,000 records. Training was conducted on a dataset of approximately 200,000 entries, using an 80/20 split for training and validation purposes.
This study contributes a practical methodology that combines advanced Spanish-language NLP preprocessing with FastText-based classification to build a multilingual, multi-label tourism sentiment model for Mexico.
Text classification is a fundamental task in natural language processing (NLP) that involves assigning predefined categories to textual data. In this project, the goal is to classify tourism-related user comments or descriptions according to three main aspects: sentiment polarity (from 1 to 5), category (restaurant, hotel, or attraction), and geographical location (states and municipalities of Mexico).
This multi-label classification problem requires a robust model capable of understanding contextual cues in the text and correctly mapping them to the corresponding labels. In tourism, such classification can support recommendation systems, improve customer experience, and assist government or private entities in analyzing public perception.
To ensure high-quality input for training our classification models, a comprehensive text preprocessing pipeline was implemented. The process involved several Natural Language Processing (NLP) libraries and custom strategies to clean and normalize Spanish-language tourist reviews.
Initially, each review was transformed to lowercase and stripped of trailing whitespace. We used regular expressions to replace date patterns and long numerical sequences with placeholder characters to prevent misleading token frequency patterns. Then, we applied two normalization strategies from the spanlp library: NumbersToVowelsInLowerCase and NumbersToConsonantsInLowerCase. These strategies replaced digits with semantically plausible letters, which helped maintain syntactic structure without numerical noise.
Accents and diacritics were removed using Unicode normalization (unicodedata), followed by ifltering all non-alphabetical characters. Tokenization was performed using the NLTK library, and Spanish stopwords were removed based on an extended list. After this step, we applied lemmatization using the spaCy model es_core_news_sm, which converts each token to its base form, improving the semantic consistency of inputs.
This preprocessing logic was encapsulated in a class named Preprocesador, which was used to transform each row in the dataset before being written to three FastText-compatible training files: one for sentiment polarity, one for business type (restaurant, hotel, or attraction), and one for geographic location (state and municipality). Only entries with valid and non-empty reviews and labels were included.
The resulting processed datasets were saved in plain text files following the FastText input format, with lines like: __label__positivo excelente comida y servicio atencion rapida __label__hotel lugar tranquilo y limpio en el centro __label__yucatan-merida paseo cultural interesante y economico
These datasets were then used to train three independent classification models using the FastText library, achieving eficient training with n-gram features and dimensional embeddings. All these steps were crucial in preparing data that is both clean and semantically rich for downstream classification tasks.
This FastText-compatible structure allowed eficient supervised training using n-gram word representations and low-dimensional embeddings, while remaining computationally lightweight and suitable for large-scale experimentation.
One of the main challenges in designing supervised classification models is class imbalance, a condition where some labels have significantly more samples than others. In this project, we observed this phenomenon especially in classes representing small municipalities or less common types of businesses within the tourism sector, such as "attraction."
Imbalance can generate biases in machine learning, favoring dominant classes during training and reducing the model’s ability to generalize to minority classes. To mitigate this problem, we developed a two-phase balancing strategy: redundancy reduction in overrepresented classes and data augmentation for underrepresented classes.
First, for classes with samples above the global average (calculated as the mean of the number of instances per label), we removed redundant examples using a hybrid text similarity approach. Both Jaccard similarity, based on sets of tokens, and fuzzy similarity using Python’s SequenceMatcher algorithm were used. These metrics allowed us to detect highly similar or near-duplicate phrases, which were eliminated to retain only unique and representative examples.
Subsequently, for underrepresented minority classes, new synthetic examples were generated. Using a JSON dictionary with representative vocabulary per category, random phrases were synthesized through controlled word sampling, ensuring semantic consistency. These phrases were tokenized and formatted in the style required by FastText, preserving the structure of the original corpus.
Finally, all classes were truncated or augmented until they equaled the average number of instances per class, thus generating a balanced dataset. This procedure helped reduce class bias during training, improving prediction fairness for lower-frequency categories.
This approach not only improves overall model performance but also increases the interpretability and reliability of predictions, especially in sensitive applications such as tourism and localized services.
In total, X redundant examples were removed and Y synthetic examples were generated, resulting in Z uniformly distributed examples per class.
This study is based on the Rest-Mex 2025 corpus, a large-scale dataset of tourist reviews focused on the most iconic and visited towns across Mexico. The dataset consists of 208,051 annotated entries, each representing a tourist’s opinion and metadata collected from multiple sources. It was released as part of the “Sentiment Analysis Magical Mexican Towns” research initiative, and is intended exclusively for academic and research purposes.
Each review includes a title, a textual review, and three key labels: • Polarity: A sentiment score from 1 (very dissatisfied) to 5 (very satisfied). • Type: The category of place described, labeled as Hotel, Restaurant, or Attractive. • Geographic Location: The name of the town and its corresponding region (state) in Mexico.
The corpus spans opinions from 40 carefully selected Mexican towns, such as Tulum, Isla Mujeres, San Cristóbal de las Casas, and Valladolid — places known for their cultural, historical, or ecological significance. These locations are distributed across 40 diferent states of Mexico, reflecting a wide geographic and touristic diversity. Tulum alone accounts for over 45,000 reviews, while towns like Tapalpa and Real de Catorce have under 1,000 reviews each, illustrating a natural class imbalance in the distribution of the data.
From a classification perspective, this dataset presents a multi-label challenge that involves: 1. Assigning a sentiment polarity (ordinal classification). 2. Identifying the type of business or site (nominal classification). 3. Predicting the corresponding municipality and state (geographic classification).
Such a setting closely simulates real-world tourism dynamics where both subjective experiences (e.g., satisfaction) and structured information (e.g., location and service type) coexist. The dataset not only enables experiments in multilingual NLP and sentiment analysis but also provides a practical foundation for regional tourism recommendation systems or public policy insights.
All preprocessing and model training in this research strictly adhere to the terms of academic use specified by the Rest-Mex 2025 initiative.
Our methodology consists of three main phases: text preprocessing, class balancing, and model training. Each phase is designed to prepare the dataset and mitigate issues such as noise, imbalance, and redundancy in tourist-related textual reviews.
To ensure semantic clarity and consistency across the dataset, we implemented a custom preprocessing class named Preprocesador, designed specifically for Spanish-language tourist reviews. The main steps of the normalization pipeline are illustrated in Algorithm 1.
Algorithm 1 Normalization Pipeline 1: Remove dates and long numerical patterns using regex 2: Replace digits using NumbersToVowelsInLowerCase 3: Normalize Unicode accents (NFKD form) 4: Tokenize and lemmatize using spaCy (es_core_news_sm) 5: Filter out Spanish stopwords using an extended list
This pipeline ensures that the resulting tokens retain only meaningful, context-rich elements suitable for classification.
Let be the number of samples in class , and let ¯ denote the global average number of samples across all classes. The balancing procedure is defined by:
Action() = ⎧Redundancy Reduction, if > ¯ ⎪ ⎨ ⎪⎩Data Augmentation, if < ¯ Redundancy was addressed by comparing intra-class instances using a hybrid similarity metric:
Sim(, ) = · Jaccard(, ) + (1 − ) · FuzzyMatch(, ) Algorithm 2 describes the pruning process applied to high-similarity entries. (1) (2) Algorithm 2 Redundancy Pruning 1: for each pair (, ) in do 2: Compute hybrid similarity Sim(, ) 3: if Sim > 0.8 then 4: Remove instance 5: end if 6: end for
For underrepresented classes, new synthetic samples were created using a domain-specific vocabulary stored in structured JSON files. Random sequences of words were generated while preserving semantic consistency, as illustrated in Figure 1.
All classes were finally normalized to a balanced state of approximately ¯ samples.
This section formalizes the key mathematical and algorithmic components of the proposed methodology, from text preprocessing to data balancing and model training using FastText.
Tourist reviews were preprocessed using a custom class Preprocesador, which implemented the following steps: 1. Conversion to lowercase and trimming of whitespace. 2. Replacement of dates and long numerical sequences using regular expressions. 3. Normalization via Unicode NFD to remove accents and non-ASCII characters. 4. Digit substitution using two transformation strategies: NumbersToVowelsInLowerCase and
NumbersToConsonantsInLowerCase. 5. Tokenization using nltk.tokenize.word_tokenize. 6. Stopword filtering using an extended list of Spanish stopwords.
7. Lemmatization using SpaCy’s es_core_news_sm.
Let = {1, 2, . . . , } be the set of reviews in class . For each class , we define its cardinality = || and compute the global mean ¯= 1 ∑︀=1 where is the total number of classes.
Each class was processed as follows:
Action() = {︃PruneRedundancy(), if > ¯
AugmentSynthetic(), if < ¯ Redundancy Detection: A hybrid similarity score was computed between pairs of sentences using:
Sim(, ) = · Jaccard(, ) + (1 − ) · Fuzzy(, ) where:
Jaccard(, ) = | ∩ | , | ∪ |
Fuzzy(1, 2) = 2 · match(1, 2) |1| + |2| (3) Synthetic Augmentation: For underrepresented classes, phrases were generated from a domainspecific vocabulary associated with category , where: frase = random_sample(, ), 5 ≤ ≤ 10
This ensured semantic relevance while compensating for data scarcity.
We trained three separate supervised classifiers using FastText: • Sentiment Polarity Model on a 5-point ordinal scale. • Business Type Model (Hotel, Restaurant, Attraction).
• Geographic Model (State and Municipality).
FastText’s softmax formulation estimates the probability of class given input vector as: (,) (|) = ∑︀′ (,′) where: (, ) = ∑︁ · + =1 with representing the weight of feature for class , and the bias.
Once trained, the models were used to predict each of the three labels per review using: 1. Model Loading: FastText models were restored using fasttext.load_model(). 2. Prediction: Each preprocessed comment was passed through the three classifiers, outputting both labels and confidence probabilities. 3. Integration: Final results were saved as an Excel file including predicted labels and their associated probabilities for analysis and visualization.
This mathematically grounded and computationally optimized pipeline enabled accurate and scalable analysis of tourism-related sentiment across multiple dimensions.
The models were evaluated on a held-out test set (20% of 208,051 samples) with the following class distributions: • Business Type: 3 classes (Hotels, Restaurants, Attractions) • Polarity: 5 classes (Rating scales 1-5) • Location: 40 classes (Mexican tourist destinations)
• Uniform performance across all classes (F1 > 0.96) • Minimal confusion between categories (Fig. 2) • Balanced precision-recall tradeof (Fig. 3)
Notable patterns: • Strong performance bias toward Class 5 (Fig. 5) • Frequent 3↔4 misclassifications (Fig. 4) • Macro F1 depressed by minority classes (2-4)
Figure 6: Sparse confusion matrix for 40 location classes Figure 7: F1-scores for top 15 performing locations
Critical findings:
The analysis reveals: • High variance in class performance (Fig. 7) • Phonetic confusion visible in of-diagonals (Fig. 6) • Top performers: Chiapas-chiapa_de_corzo (F1=0.7577), QuintanaRoo-tulum (F1=0.7413) • Class imbalance significantly impacts minority classes • Ordinal bias afects sentiment analysis • Geographic complexity requires specialized handling for location names
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. The sources for the ceur-art style are available via • GitHub, • Overleaf template.