The HOMO-LAT25 shared task focuses on detecting sentiment polarity in online messages related to the LGBTQ+ community across Latin American Spanish dialects. The UC-UCO-CICESE_UT3-Plenitas team participated in two main subtasks: multi-dialect polarity detection-where the training and evaluation dialects match-and crossdialect polarity detection-where they difer. These subtasks challenge systems to recognize positive, negative, or neutral sentiment toward specific keywords across varying sociolinguistic contexts. The team employed transformer-based models from the Hugging Face ecosystem, particularly leveraging the Robertuito model from the pysentimiento library, which is pre-trained for Spanish sentiment classification. The overall methodology relied on deep learning techniques to manage unstructured social media text, emphasizing the adaptability of the transformers for multilingual sentiment analysis. Out of 31 teams and 67 total submissions, only 8 teams progressed to the final evaluation phase. The UC-UCO-CICESE_UT3-Plenitas team reached the finals and placed 7th in both subtasks. Their F1-scores were 0.2592 and 0.2588, respectively-demonstrating baseline performance but highlighting a clear gap from top-performing teams. Despite the modest scores, their progression to the final stage reflects the model's initial competence and provides a solid foundation for future enhancements. Key areas for improvement include better handling of dialectal variation, improved feature engineering, and optimization of classification strategies. Overall, the HOMO-LAT25 task represents a critical benchmark for developing inclusive, culturally aware NLP systems, supporting applications such as content moderation, public sentiment tracking, and digital policy eforts aimed at protecting LGBTQ+ communities in Latin America.
The HOMO-LAT (Human-Centric Polarity Detection in Online Messages Oriented Toward the Latin
American LGBTQ+ Population) initiative represents a collaborative efort among researchers across
various Latin American countries [
The HomoLat 2025 shared task centers on polarity detection for a curated set of generic terms
commonly used throughout Latin America . It is designed to engage a broad research community,
including both academic and industrial stakeholders, by ofering a challenging benchmark that combines
sentiment analysis with the cultural and linguistic variability inherent in Latin American Spanish
dialects [
Beyond academia, the task holds significant practical relevance. LGBTQ+ advocacy organizations can
employ polarity detection systems to identify and respond to harmful or exclusionary language in digital
discourse [
Deep learning and transformer-based architectures have revolutionized natural language processing
(NLP). Models such as BERT and its derivatives are capable of modeling linguistic nuance and
longdistance dependencies, making them ideal for multilingual tasks. Hugging Face, as an open-source
repository and community, facilitates access to these architectures via its Transformers and Datasets
libraries. Often referred to as the ´GitHub of Machine Learning´, it ofers an extensive repository of
pre-trained models, datasets, and interactive applications via its Model Hub. Deep learning is a subfield
of machine learning that employs artificial neural networks with multiple layers, hence the term ‘deep’,
to automatically learn complex patterns in large datasets. These models excel at processing unstructured
data such as text, images, and audio. Rather than relying on explicit rule-based programming, deep
learning systems learn to map input to output through iterative training, mimicking some aspects of
human cognitive learning [
This task ultimately invites the Latin American NLP and AI community to develop and refine
applications tailored to the sociolinguistic diversity of the region[
The oficial HOMO-LAT25 dataset was used. It included annotated social media messages with polarity labels (positive, neutral, negative), alongside dialectal information. The dataset was divided into two • Subtask 1: Multi-dialect (same dialects in train and test) • Subtask 2: Cross-dialect (dialects difer in train and test)
Subtask 1: Multi-Dialect Polarity Detection (Multi-Class Classification) The goal of this subtask
is to determine the polarity—positive, negative, or neutral—of a social media post with respect to a
given keyword. The distinguishing feature of this task is that the dialectal varieties of Spanish used in
the training set are the same as those appearing in the evaluation set, allowing models to learn and be
evaluated within consistent linguistic boundaries [
Subtask 2: Cross-Dialect Polarity Detection (Multi-Label Classification) This subtask also aims to classify the polarity of posts towards a specific keyword but introduces a greater level of dificulty: the Spanish dialects present in the evaluation set difer from those in the training set. This setup tests the cross-linguistic generalizability and robustness of the models, pushing them to adapt to previously unseen dialectal variations in the context of sentiment classification.
Key components of transformers include multi-head self-attention, positional encoding to preserve
word order, and feed-forward networks for feature transformation. Layer normalization and residual
connections facilitate stable and eficient training. Owing to their scalability and performance,
transformers form the foundation of many state-of-the-art models such as BERT, GPT, and T5, and have
become the dominant architecture not only in NLP but increasingly in computer vision and speech
recognition domains as well.
2.1.1. Robertuito: A Transformer-Based Model for Spanish Sentiment Analysis
Robertuito is a pre-trained natural language processing (NLP) model specifically fine-tuned for
sentiment analysis in Spanish. Distributed as part of the pysentimiento library, Robertuito is designed to
classify textual data into sentiment categories such as Positive, Negative, Neutral, and in some cases,
Mixed or Other. The model is built upon transformer-based architectures, primarily variants of BERT
(Bidirectional Encoder Representations from Transformers), and has been adapted to perform efectively
on Spanish-language tasks [
Messages were lowercased and cleaned of non-textual noise (e.g., emojis, links). No stemming or lemmatization was applied to preserve dialectal features.
To assess the performance of sentiment analysis models submitted by competition participants, the
F1-score is employed as the primary evaluation metric [
The evaluation used the macro-F1 score for Subtask 1 and the sample-based F1 for Subtask 2. This evaluation framework ensures a fair and comprehensive comparison of models, particularly in settings with class imbalance or multiple sentiment labels.
The lowest F1-score was observed for the Positive class, particularly in the cross-dialect scenario. Manual inspection revealed false positives caused by irony or dialect-specific expressions, see table 1.
Subtask Multi-dialect Cross-dialect
Class Positive Neutral Negative Positive Neutral Negative
Precision
F1-score
Competition Results and Performance Analysis: The HOMO-LAT 2025 shared task attracted participation from 31 teams, generating a total of 67 submissions over the course of the competition. From these, only 8 teams advanced to the final evaluation stage, highlighting the selective and competitive nature of the task.
The UC-UCO-CICESE_UT3-Plenitas team (ymlopez) was among the finalists, ultimately ranking 7th out of the finalist teams. The team submitted entries for two subtasks (tracks), achieving the following results: • Track 1 – Multi-Dialect Polarity Detection (Multi-class) F1 Score: 0.2592, Final Rank: 7th • Track 2 – Cross-Dialect Polarity Detection (Multi-label) F1 Score: 0.2588, Final Rank: 7th
Although the UC-UCO-CICESE_UT3-Plenitas team successfully advanced to the final phase of a highly competitive evaluation involving 31 participating teams, its final placement—7th out of finalist teams—highlights the need for further methodological enhancements. Future research should prioritize the refinement of classification strategies, the development of more robust and representative data features, and the improvement of model generalization. These eforts aim to elevate performance metrics, particularly the F1 score, towards the levels achieved by mid-ranking teams, which consistently reached the 0.40–0.45 range. Such improvements are essential for achieving competitive parity in complex multilingual and sentiment analysis tasks. While the performance of our system ranked among the finalists, our results indicate clear areas for improvement. These include enhancing dialectal generalization, leveraging domain adaptation techniques, and incorporating class-weighted loss functions to handle imbalanced data. Future work will focus on integrating sociolinguistic features and leveraging multilingual embeddings to capture deeper cultural context in sentiment classification.
During the preparation of this work, the authors used GPT-4 and Grammarly in order to: Grammar and spelling check. After using these services, the authors reviewed and edited the content as needed and takes full responsibility for the publication’s content.
The results are available via • HOMOLAT25 results, • HOMOLAT Codabench.