In today's digital landscape, social media information has become a valuable source for capturing readers' attention quickly through clickbait. However, this information sometimes leads users to low-quality or misleading information. For this reason, the TA1C shared task in IberLEF 2025 focuses on detecting clickbait from social media posts in Spanish. This paper presents an NLP framework developed by the UAEMex team that exploits lexical and semantic information from texts to detect this kind of information. This framework was also complemented with traditional classifiers (KNN, LR, NaiveBayes, and MLP) to create models that best fit the task. According to the obtained results, semantic representations provided by BETO (Spanish version of BERT) provide better representations with LR and MLP classifiers, obtaining 0.7538 in F1-score, which is a competitive performance compared to state-of-the-art approaches and baselines in the task.
According to Mordecki, clickbait is defined as “a technique for generating headlines and teasers
that deliberately omit part of the information with the goal of raising the readers’ curiosity,
capturing their attention and enticing them to click” [
In practice, clickbait has detrimental effects on the quality of journalism and the formation of
public opinion, as it displaces relevant information in favor of attractive but superficial or
misleading content. Various authors agree that clickbait undermines the quality of public discourse
and the healthy functioning of democratic societies. Palau-Sampio warns that leading newspapers
such as El País have adopted sensationalist writing strategies that blur the boundaries between
quality journalism and entertainment [
Potthast et. al., also criticizes media outlets that prioritize maximizing clicks for economic gain at
the expense of content quality [
To enable automatic clickbait detection using machine learning, some researchers have
compiled datasets to train models that have proven effective for natural language processing (NLP)
tasks. However, Mordecki and colleagues in [
To address these problems, Mordecki and collaborators recently introduced the TA1C (Te
Ahorré Un Click) dataset—the first manually annotated Spanish-language corpus for clickbait
detection, consisting of 3,500 tweets from 18 media outlets [
The dataset was publicly presented under the title Detecting and Spoiling Clickbait in
SpanishLanguage News [
This paper is organized as follows: Section 2 reviews related work on clickbait detection using various datasets and NLP methods. Section 3 describes the proposed system for clickbait classification. Section 4 discusses the results and hyperparameters for each method, and Section 5 outlines the main conclusions and feature experimentations.
Automatic clickbait detection has gained increasing attention in NLP community due to its
implications for digital misinformation, user trust, and the quality of online journalism [
To enable these models to operate effectively, a vectorization phase is required in which textual
data is transformed into numerical representations suitable for computational processing.
Vectorization techniques range from traditional approaches such as bag-of-words or TF-IDF [
Regarding machine learning algorithms, models commonly used for clickbait detection range
from traditional approaches such as Logistic Regression (LR), Support Vector Machines (SVM),
Random Forests (RF), and Multilayer Perceptron (MLP), to more recent and complex deep learning
architectures—including Convolutional Neural Networks (CNN), Recurrent Neural Networks
(RNN), and transformer-based models—designed to more effectively capture the contextual
information inherent in natural language [
One early approach is proposed in [
In [
Other researchers developed a dataset of headlines drawn from both reputable media outlets
and sources known for their clickbait [
In [
In a recent study, academics presented in [
This section provides a detailed overview of the proposed approach. First, we detail the data stratification process used for evaluation. Second, we provide a description of the preprocessing steps applied to the datasets. Next, we describe the text representation methods employed, and finally, it introduces the classification algorithms used, along with their corresponding hyperparameter settings.
Following the TA1C shared task (Te Ahorré Un Click – Clickbait Detection and Spoiling in Spanish), we used the official clickbait detection corpus made available for the competition.
The TA1C dataset consists of 3,500 news items collected from Twitter (currently known as X),
originating from 18 reputable media sources that are national or international (excluding local
outlets), with a generalist focus (not topic-specific) and high popularity. The corpus is divided into
three subsets: 2,100 instances for training, 700 for validation, and 700 for testing [
The TA1C shared task includes two subtasks. Subtask 1, called Clickbait Detection, aims to determine if the content of a tweet that links to a piece of news is clickbait. In contrast, Subtask 2, named Clickbait Spoiling, consists of generating or extracting from the article a short text – up to 280 characters – that, as concisely as possible, either satisfies the curiosity sparked by the headline or indicates that the article does not provide a clear answer. In this paper we focus on Subtask 1.
During the first phase, the organizers released two datasets: a training set and a development set. Only the training set included labels, allowing participants to build supervised classification models. At this stage, the development set labels were not disclosed; however, participants could evaluate their models using this data through the CodaLab platform [13]. Table 1 shows the class distribution of the training set.
In the second phase, the organizers re-released the training set along with the now-labeled development set and an unlabeled test set, which was used for the final evaluation. The class distribution of the development set is detailed in Table 2.
The class distribution imbalance, with 71.50% clickbait texts in the training set and 71.00% in the development set, indicates a high prevalence of clickbait content. This disproportion may bias models toward the majority of the class. Furthermore, detecting clickbait requires analyzing not only the superficial content of the text but also semantic and contextual aspects that enable understanding of the communicative intent and potential ambiguities present in the message.
Each dataset contains the followings columns: Tweet ID, Tweet Date, Media Name, Media Origin, Teaser Text and Tag Value. The data from Teaser Text column was processed and vectorized to generate input features for the classification models and the data from Tag Value column was used as the label to predict. Table 3 shows two representative examples of texts labeled as clickbait and two corresponding to the non-clickbait class.
The following examples illustrate certain linguistic and discursive features that differentiate content labeled as clickbait from non-clickbait. In general, texts classified as clickbait tend to employ questions, ambiguous phrasing, or curiosity-inducing expressions aimed at encouraging the reader to click in order to access the full content. In contrast, non-clickbait texts present information in a clear, direct, and specific manner, without omitting key elements of the text.
#ÚltimaHora México llegó a 70 mil 821 muertes por Covid-19, con 668 mil 381 casos confirmados de coronavirus
To ensure the quality and consistency of the input data before training the models, a preprocessing stage was applied to the TA1C dataset. This step is crucial in NLP tasks, as it reduces noise, standardizes textual input, and facilitates the extraction of relevant features. The processing aimed to transform raw tweets into structures and clean format for machine learning algorithms. Below we describe the preprocessing techniques considered in this study: • Tokenization: As an initial step, tokenization was applied to split the text into individual units (tokens), typically corresponding to characters or words to improve the association or “understanding” of words for each sentence. • Normalization: In this process we removed non-alphanumeric characters, and all texts were converted to lowercase. • Stopwords removal: In some cases, we removed common words that carry little semantic weight, such as “el”, “la”, “y”, or “de” in Spanish — were removed using a predefined list from the NLTK library. This step reduces noise and helps focus the model on more informative terms.
The model’s architecture is built upon a component-based framework called NLPClassKit. This toolkit consists of a set of software components designed to facilitate the development of classification models for NLP projects. It is currently available in a GitLab repository (https://gitlab.com/JohnRojas/NLPClassKit) and follows a process-oriented architecture as illustrated in Figure 1.
Once the text documents were extracted or preprocessed, they were fed into various vectorization methods. These techniques are crucial for representing and extracting relevant information from the text, whether at the lexical or semantic level. Some of these approaches are implemented in a software component called Text2Vec, developed in Python and hosted in a GitLab repository (https://gitlab.com/MLComponents1/Text2Vec). Technical aspects related to syntax; parameters and the execution process are described in the README file of the repository. This module encodes text using the following strategies:
One-hot-encoding (OHE): A straightforward technique that represents words as binary vectors. First, vocabulary is built from the text, assigning a unique index to each term. Then, each word is converted into a binary vector with a value of 1 in the position corresponding to its presence in the document and 0 elsewhere. Although OHE requires minimal computational resources, it produces high-dimensional vectors, especially with large vocabularies [14]. The length of the output vectors is variable and depends on the input vocabulary; for example, during our training phase, a vector of dimension 18,770 was produced using this method with the test set.
TF-IDF (Term Frequency-Inverse Document Frequency): A frequency-based method that evaluates the importance of a word in a document relative to a corpus. It combines Term Frequency (TF), which counts how often a word appears in a document with Inverse Document Frequency (IDF) that measures its rarity across the corpus. This weighting reduces the impact of common terms and highlights those more relevant to the document’s content [15]. On the other hand, TF-IDF encoding outputs a sparse vector representation in which each dimension corresponds to a specific term from the corpus vocabulary. The value of each dimension reflects the weight of that term in the document based on its TF-IDF score. The length of the resulting vector depends on the size of the vocabulary, which is determined by steps such as tokenization, stop words removal, etc.
Doc2Vec: An extension of the Word2Vec model that generates vector representations for entire documents, paragraphs, or sentences rather than individual words. This approach captures the overall semantic context of the text and encodes it into a fixed-length vector, facilitating tasks like document comparison, classification, or clustering [16]. The length of the output vector is variable and can be controlled through its parameters. Other parameters such as context size, the use of iterations, or the use of bag-of-words or skip-grams, among others, can also be controlled by the model user.
BERT (Bidirectional Encoder Representations from Transformers): A pre-trained model based on attention mechanisms, capable of interpreting word context bidirectionally (analyzing both preceding and subsequent words). Unlike traditional methods, BERT uses a transformer architecture to capture complex relationships and deep contextual meanings. Additionally, it can be fine-tuned with domain-specific texts to enhance performance in specialized tasks [17]. BERT-base is a language model with 12 layers, 768 hidden units, 12 attention heads, and a vocabulary of 30,522 tokens. It can process sequences of up to 512 tokens and has approximately 110 million parameters. The extended version, BERT-large, reaches 340M parameters.
ASCII: An approach that analyzes the frequency of ASCII characters in a document to estimate the probability of each character’s occurrence. Though simpler than other techniques, it is useful for specific tasks like style or formatting classification [18]. This encoding always outputs a 256dimensional vector, in which the first 128 characters correspond to upper- and lower-case letters, numbers, punctuation marks and some control characters from the English language, and the remaining characters include symbols and accented letters from other languages.
The vector representations generated in the preceding stage serve as input for multiple supervised machine learning models, each specifically tailored to categorize textual data through pattern recognition. Below is a concise overview of the primary algorithms employed in this research:
Naive Bayes (NB): This probabilistic classification model operates on the foundation of Bayes' Theorem. It computes the likelihood of a text sample belonging to a specific category by analyzing its feature set. Although NB employs a simplified approach by assuming feature independence, it demonstrates notable efficacy in text classification scenarios, especially when this assumption approximately holds true.
K-Nearest Neighbors (KNN): As a non-parametric classification method, KNN determines class membership for new instances by identifying the predominant class among its K closest neighboring points within the vector space. The algorithm operates on a majority voting system and proves particularly valuable for complex, non-linear decision boundaries. The selection of neighbors depends crucially on distance computations, commonly employing measures such as Euclidean distance or cosine similarity.
Logistic Regression (LR): Logistic Regression is a linear classification algorithm used to predict binary outcomes. It models the relationship between a set of input features X and a binary target variable y ∈{0 , 1 }. The prediction is made using the sigmoid function:
P ( y = 1) = 1 +1e- z , z = ∑ wi xi , (1) where wi are the model parameters and xi are the input features. The output is a probability value between 0 and 1, indicating the likelihood that the input belongs to class 1.
Multilayer Perceptron (MLP): The MLP represents a fundamental deep learning architecture comprising three distinct fully connected layer types: (i) an input layer that receives feature representations, (ii) multiple hidden layers capable of learning hierarchical feature transformations, and (iii) an output layer that produces classification probabilities. This neural network architecture demonstrates effectiveness in modeling non-linear relationships within complex datasets, making it well-suited for text classification applications where it can learn intricate patterns in textual representations.
Finally, the outcome of the classification process is presented in the final stage, where each algorithm described in the previous section produces the following output: • • •
Confusion Matrix (CM): Displays the number of texts correctly and incorrectly classified. Predictions: Lists the predicted labels generated during the training phase of each algorithm. Model Performance: Provides a detailed evaluation of each classification model based on metrics such as Recall, Precision, and F1-score.
During the experimentation and testing phase, the team members conducted over 1,500 experiments to explore various parameter configurations in both vectorization methods (VecM) and classifiers (Clf). We shared the best results among ourselves, aiming to explore as many options as possible. In the final phase, predictions were made on the test set of 700 unlabeled tweets, which were used to generate the official scores shown in Table 4. The performance of each model is presented in the Precision (P), Recall (R), and F1-score (F1) metrics. Moreover, we show the dimensions of vector representations (Dvec) and the number of parameters of BERT-based models ( N p) for each experiment.
Table 4 presents a comparative analysis of different features and classifiers tested in the NLPClassKit system. The best overall performance on the test set was achieved by UAEMex-1 and UAEMex-2, both using the pre-trained model albert-base-spanish combined with a LR classifier. These models obtained the highest F1-score of 0.7538, indicating a strong balance between precision and recall. UAEMex team corresponded to the set of classifiers with simple majority voting.
Interestingly, UAEMex-3, which used the same text vectorization but replaced LR with a MLP, showed a slightly lower precision (0.7273) but improved recall (0.7665) resulting in a competitive F1-score of 0.7463. In contrast, when we used ASCII vectorization combined with bert-basemultilingual-cased (UAEMex-4-2 and UAEMex-4-3), we achieved the lowest performance with F of 0.6610 and 0.6369, suggesting that this combination is less effective for the clickbait detection task. Meanwhile, configurations that combined bert-base-spanish-wwm-cased with albert-base-spanish (UAEMex-4-4 and UAEMex-4-5) outperformed the multiligual-based system, reaching an F1-score of 0.7122 and 0.7352. However, these results were slightly below those of the best models, where only alBERT features were used. These results demonstrate that not only the classifier architecture but also the quality and linguistic adequacy of the vector representations are key factors in optimizing performance in Spanish text classification tasks.
Finally, we combined the embedding generated by bert-base-spanish-wwm-cased and albertbase-spanish. Although this may initially seem redundant, given that both models are pre-trained on Spanish; However, we decided to carry out this experiment because they have different architectures, which means they learn and represent distinct aspects of the language. On one hand, BERT has a greater capacity to capture complex semantic relationships; while albert, being a lighter and more efficient model, tends to generalize better in simpler or noisier contexts. ABS: albert-base-spanish, ASC: ASCII, ABBM: ASCII + bert-base-multilingual-cased, and BSWC: bert-base-spanish-wwm-cased + albert-base-spanish.
In general, we expected that combination of both features would allow us to leverage their complementary strengths and thus enhance the classification system is performance, but in this case, the results showed a slight decrease in the F-measure compared to using albert-base-spanish. This drop in performance could be attributed to redundancies or conflicts in the embeddings generated by the two models, which may introduce noise rather than providing additional value.
Regarding vectorization methods, the best performance was obtained with albert, a linguistic model based on BERT, trained specifically for the Spanish language by the Computer Science Department of the University of Chile and available in the Hugging Face repository [13]. It should be noted that other Spanish-oriented BERT models were also evaluated, such as bert-base-spanishwwm-uncased and bert-base-multilingual-cased, which showed lower performance than albert when used with the same classifiers. Table 5 presents the hyperparameter settings corresponding to the machine learning models that obtained the best results.
After the evaluation phase was completed, the results were published on the CodaLab platform,
an open-source system designed to host and manage scientific challenges [
According to the F1 scores, our team ranked 10th, 11th, 12th, and 13th. Compared to the topperforming team (gsdeyson), the difference was 0.047 points. It is worth noting that all the metrics presented in the table were calculated using a combined training set composed of the 2,800 labeled tweets provided during the development phase and an additional 700 labeled tweets from the evaluation phase.
Additionally, the confusion matrices corresponding to the best results are displayed in Figure 2. It can be observed that the models exhibit a slight bias toward predicting that a given text does not contain clickbait. This tendency is primarily attributed to the class imbalance in the training data, with 1,001 clickbait examples versus 2,499 non-clickbait examples (see Tables 1 and 2). On the other hand, we observe that the best-performing model (UAEMex-2) does not show a better prediction of clickbait in the training stage compared to UAEMex-3 and UAEMex-4-5 models. We assume that this may be an indicator of generalization in the proposed model, enabling a balance between non-clickbait (0) and clickbait (1).
In this study, the team collaborated to address the clickbait detection task proposed by Mordecki and colleagues as part of the IberLEF 2025 forum, with the goal of contributing to the development of algorithms capable of eliminating unethical practices exhibited by some digital media creators and promoting high-quality journalism.
During the competition, various text vectorization techniques were explored, allowing for a comparative analysis between several custom BERT-based models for Spanish and traditional vectorization approaches. This experimentation enabled the identification of the most suitable method for this specific natural language processing task. Among the tested models, Albert proved to be the most effective, consistently outperforming other Spanish-language BERT models during the evaluation phase.
Additionally, four classic supervised machine learning algorithms were tested, yielding several noteworthy findings. For instance, while the Multilayer Perceptron achieved strong results during training, Logistic Regression demonstrated superior generalization performance during the competition, which is the primary goal of a classification model. Similarly, the ensemble classifier based on majority voting achieved comparable performance to logistic regression, likely due to its internal inclusion of similar models.
For future work, we recommend further exploration of BERT-based Spanish word embedding models and experimentation with alternative classifiers beyond those evaluated in this study. Furthermore, no dimensionality reduction techniques were applied in this project; consequently, classification efficiency could potentially be enhanced through the implementation of methods such as Principal Component Analysis (PCA), Genetic Algorithms (GAs), or Information Gain, among others. Additionally, the implementation of a probability-based ensemble classifier is suggested as a promising line of research.
During the preparation of this work, the authors used Grammarly tool in order to: Grammar and spelling check. After using this tool, the authors reviewed and edited the content as needed and take full responsibility for the publication’s content. advances in social networks analysis and mining (ASONAM). pp. 9-16. https://doi.org/10.1109/ASONAM.2016.7752207 [13] Pavao, A., Guyon, I., Letournel, A.-C., Tran, D.-T., Baro, X., Escalante, H. J., Escalera, S., Thomas, T., & Xu, Z. (2023). CodaLab Competitions: An Open Source Platform to Organize Scientific Challenges. Journal of Machine Learning Research, 24(198), pp. 1-6. [14] Rojas-Simon, J., Ledeneva, Y., & Garcia-Hernandez, R. A. (2022). Evaluation of Text Summaries Based on Linear Optimization of Content Metrics (Vol. 1048). Springer International Publishing. https://doi.org/10.1007/978-3-031-07214-7. [15] Le, Q., & Mikolov, T. (2014). Distributed representations of sentences and documents. In
International conference on machine learning. 32(2), pp. 1188-1196. [16] Abubakar, H. D., Umar, M., & Bakale, M. A. (2022). Sentiment classification: Review of text vectorization methods: Bag of words, Tf-Idf, Word2vec and Doc2vec. SLU Journal of Science and Technology, 4(1), pp. 27-33. https://doi.org/10.56471/slujst.v4i.266 [17] Devlin J., Chang M.-W., Lee K., & Toutanova, K. (2019). BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding. pp. 4171–4186, Accessed: May 20, 2025. [Online]. Available: https://github.com/tensorflow/tensor2tensor. [18] Departamento de Ciencias de la Computación de la Universidad de Chile (DCC UChile). (2023).
ALBERT Base Spanish BERT: dccuchile/albert-base-spanish. Available: https://huggingface.co/dccuchile/albert-base-spanish.