This paper presents our participation in the TA1C: Clickbait Detection and Spoiling in Spanish task at IberLEF 2025, which focuses on the automatic identification and mitigation of clickbait in Spanish-language online news shared on social media. The task comprises two subtasks: (1) a classification task, where the system must determine whether a given tweet constitutes clickbait, and (2) a generative task, in which the system-given a teaser (tweet and headline) along with the corresponding news article-must generate a concise spoiler that resolves the information gap and satisfies the curiosity induced by the teaser. The dataset includes Spanish tweets, headlines, and full news articles, annotated to support both classification and generation objectives. Our team participated in the first subtask, addressing the challenge of clickbait detection by developing several systems based on pre-trained transformer-based language models, applying fine-tuning strategies to improve prediction quality. The results confirm the efectiveness of our approach.
Clickbait is a prevalent phenomenon in online news media. It refers to the use of sensational, ambiguous, or deliberately incomplete headlines and teasers intended to provoke curiosity and drive user engagement, often at the expense of informativeness. Although its primary function is to increase trafic, clickbait often contributes to the spread of low-quality or misleading information, eroding trust in digital journalism.
The growing use of clickbait is not limited to soft news or dubious sources; increasingly, even reputable media outlets employ clickbait strategies for high-impact stories. The need for automatic tools to detect and mitigate the influence of clickbait content has become critical, not only from a technological standpoint but also from an ethical and communicative perspective.
Although early eforts in clickbait detection focused primarily on English, recent research has begun to address this issue in other languages, including Spanish. However, Spanish remains underexplored in terms of large-scale, annotated datasets, and shared evaluation frameworks.
This work presents our contribution to the shared task TA1C ("Te Ahorré Un Click") at IberLEF
2025 [
The remainder of the paper is structured as follows. In Sections 2 and 3, we introduce the task and describe the dataset and evaluation metrics. Section 4 presents our clickbait classification system along with our experimental results. Finally, Section 6 concludes the paper and discusses future work.
The TA1C shared task consists of two subtasks that address the identification and mitigation of clickbait in Spanish-language online news shared on social networks. The dataset includes tweets, news headlines and full news articles, annotated to support both classification and generation objectives. We have experimented with the first subtask, clickbait detection.
The task is a binary classification problem: Given a teaser consisting of a tweet and the corresponding
headline of the news article, the objective is to determine whether the content qualifies as clickbait. The
annotation relies on the definition proposed by Mordecki et al. [
The dataset for both tasks consists of Spanish-language tweets that link to online news articles. Data
were curated for the TA1C shared task [
The dataset for Subtask 1 includes 4200 tweets collected between October 2020 and October 2021 from 18 well-known media outlets in 12 Spanish-speaking countries. Each tweet is accompanied by the URL and clean HTML of the linked article, the headline, subheadline, article body (cleaned), images and captions, and embedded external links.
Each tweet was independently labeled by three human annotators to determine if it is clickbait. The 4200 samples of the dataset are split as shown in Table 1. We can appreciate an imbalance towards Non-clickbait labeled samples in the training and development set. Regarding the test set, we could not provide the distribution of Clickbait and Non-clickbait samples for that partition since the participants could not access the test labels. Examples of clickbait and non-clickbait samples are provided in Table 2.
We developed six systems for the Clickbait Detection task, each exploring diferent pre-training and
data augmentation strategies based on the Spanish-language RoBERTa model pre-trained with data
from the National Library of Spain (roberta-base-bne) [
All systems were trained with the same hyperparameters: 5 epochs, a learning rate of 3 × 10 −5 , batch size of 16, 50 warmup steps, and a weight decay of 0.01. The TA1C training split (2800 examples) was used as the training data, and the development set (700 examples) was used for validation, unless otherwise stated. The following configurations were explored.
This system serves as the reference point for evaluating the efectiveness of subsequent strategies. The roberta-base-bne model was fine-tuned directly on the TA1C training set. The development set was used to monitor validation performance and select the best checkpoint.
To enrich the model with more diverse examples of clickbait-like language, we carried out an
intermediate pre-training phase using 20,000 English tweets that were machine-translated into Spanish. The
20,000 tweets were randomly sampled from the https://huggingface.co/datasets/christinacdl/clickbait_
detection_dataset dataset. Translation was performed using the Helsinki-NLP translation
EnglishSpanish model (https://huggingface.co/Helsinki-NLP/opus-mt-en-es) [
A data augmentation method was applied using multilingual backtranslation [
In this system, we applied a random masking strategy to simulate Masked Language Model (MLM)
noise. We generated five augmented versions of the TA1C training set by randomly masking 15% of the
tokens in each sentence. After concatenation and shufling, the final dataset contained approximately
14,000 augmented samples. This noisy training data was used to fine-tune the roberta-base-bne
model, again validated on the TA1C development set.
As an additional strategy to enhance the model’s generalization capabilities, Named Entity Recognition
(NER) masking was applied using spaCy [
This system trained on both the training and validation sets (3500 tweets in total) without a development set for evaluation. The roberta-base-bne model was fine-tuned directly on these 3500 tweets. No early stopping or checkpoint selection was applied; the final model corresponds to the last training epoch. All other training parameters were the same.
This section summarizes the performance of the systems on both the TA1C development set and the oficial blind test set provided for the shared task. All systems were evaluated using the F1-score as the primary metric.
Table 3 presents the F1 results. System T1-3, which incorporated backtranslation-based augmentation, achieved the highest F1 score on the test set (0.81481), ranking third place overall in the shared task leaderboard. Notably, although System T1-1 (baseline fine-tuning) already achieved a strong performance, augmenting the training data with synthetic samples from multilingual backtranslation in T1-3 ofered a meaningful improvement of more than two points in the test F1-score.
System T1-2 (two-phase pre-training) and T1-4 (random masking) showed moderate performance compared to the baseline in the development set, and did not surpass T1-3 on the test set. These findings suggest that while pre-training on loosely aligned translated data or introducing random masking noise may help generalization, carefully filtered backtranslated samples better preserve task-relevant semantics and are more efective for low-resource domain adaptation in Spanish clickbait detection.
System T1-5 (NER masking) showed solid generalization performance, achieving an F1-score of 0.8451 on the validation set and 0.7647 on the test set. Despite a drop from development to test, the results suggest that abstracting entities can still retain useful semantic patterns and may help reduce overfitting.
Finally, System T1-6, trained on all available data, achieved a strong test F1 (0.80342), showing the benefit of leveraging the full dataset at the expense of validation-based checkpoint selection.
Fig. 1 visually summarizes the results from Table 3, with systems ordered by Test performance. System T1-3 (Backtranslation Augmentation) achieved the highest F1 on the test set (0.81481), conifrming the efectiveness of multilingual backtranslation in generating useful training samples. System T1-6, which leveraged all available data (training + validation), also performed well on the test set (0.80342). The baseline system T1-1 already showed strong performance, suggesting that the pre-trained roberta-base-bne model is well-suited for the task even without additional augmentation. Systems T1-4 (Random Masking) and T1-2 (Two-phase Pre-training) yielded slightly lower test scores, indicating limited benefits from these strategies in this context. Finally, System T1-5 (NER Masking) had the lowest test F1-score (0.76471), but still demonstrated respectable generalization, highlighting its potential to reduce overfitting by abstracting entity-specific information. Overall, the figure underscores that not all augmentation techniques are equally efective, with carefully designed backtranslation yielding the most significant gains.
This study explored several approaches for clickbait detection in the TA1C competition in Spanish [
Among the techniques evaluated, backtranslation proved particularly efective, significantly enhancing system robustness and contributing to the highest test F1-score (0.81481) across all our models, achieving third place in the competition. While other strategies, such as random masking or NER-based masking, showed more modest gains, they still ofer potential for reducing overfitting and warrant further investigation.
Future work will focus on experimenting with larger language models and more sophisticated data augmentation strategies, aiming to further improve performance and generalization in clickbait detection tasks.
We have not used additional data to those provided by the competition. The pre-trained models used are obtained from HuggingFace models hub, under the Apache License 2.0.
This work is partially supported by MCIN/AEI/10.13039/501100011033 and "ERDF A way of making Europe" under grant PID2021-126061OB-C41. It is also partially supported by the Spanish Ministerio de Universidades under the grant FPU21/05288 for university teacher training.
During the preparation of this work, the author(s) used ChatGPT in order to: Grammar and spelling check. After using these tool(s)/service(s), the author(s) reviewed and edited the content as needed and take(s) full responsibility for the publication’s content.