In today's digital landscape, users frequently turn to social media, email, and messaging platforms to access news, entertainment, and a wide range of content. This highly competitive attention economy has transformed how information is produced and consumed. In particular, click-based monetization has encouraged many content creators to adopt deceptive strategies, such as clickbait, which aims to attract attention through sensational headlines that often do not reflect the actual content. The IberLEF 2025 shared task TA1C was designed with two primary objectives: (1) the detection of clickbait and (2) the generation of informative content to counteract it, referred to as spoiling. This task reflects the growing need for tools that promote more transparent and usercentered communication. For Subtask 1, we implemented an ensemble of transformer-based models fine-tuned for clickbait detection, combined with a generative model utilizing a few-shot learning approach to enhance context understanding. For Subtask 2, we leveraged Claude's generative API, also in a few-shot setup, to generate spoiler content that accurately reflected the linked article. Our approach achieved competitive results: seventh place in Subtask 1, with an F1-score of 79.56%, and second place in Subtask 2, with a BLEU score of 42.81%. These results suggest that combining classification and generation strategies is a promising direction to address the challenges of online content consumption.
An important factor in communication is advertising, which was the primary means by which
magazines and newspapers sustained themselves. In the current economic context, distribution
methods have undergone significant changes. Now, the main way people receive information is
through online platforms—primarily Facebook, Twitter (now X), WhatsApp, email, and websites. These
media also need to adapt to survive. Their survival is now linked to visits to their platforms, which
is the leading way websites remain viable, while videos depend on the number of views [
The structure of this manuscript is organized as follows: in Section 2, we present a concise literature review on hate speech detection, beginning with a general overview and subsequently focusing on the specific context of the TA1C task. Section Section 3 details our approach, including the preprocessing steps, the models used, and the evaluation metrics. The results obtained are discussed in Section 4. Finally, Section 5 highlights the significance of our approach and outlines potential future research directions.
In this section, a brief literature review on automatic clickbait detection and clickbait spoiling is presented.
In early eforts to automate clickbait detection, Naeem et al. (2020) present a deep learning framework
that focuses on extracting clickbait-specific linguistic cues through part-of-speech (POS) analysis and
subsequent classification with an LSTM network [
A year later, in 2021, Al-Sarem et al. explore clickbait detection in the Arabic language context,
presenting the first corpus of around 55, 000 annotated headlines, of which 43.7% correspond to
clickbait and a hybrid approach combining lexical, user and metadata features with traditional machine
learning classifiers such as SVM [
In 2023, with the proliferation of transformative models, two comparable studies adopted
BERTbased [
More recently, in 2025, Muqadas et al. extended the research to Urdu language contexts, presenting
an annotated corpus of news headlines and evaluating a Bi-LSTM model that uses sentence embeddings.
Their system achieves approximately 88% accuracy, outperforming manual feature-based methods and
demonstrating that contextual embeddings can be successfully adapted to low-resource languages [
In 2022, the first formal papers addressing the generation of clickbait spoiling emerged. Hagen et al. define the clickbait spoiling task as:
The generation of a short text that responds to the information gap created by a sensational headline.
They introduce the Webis Clickbait Spoiling 2022 corpus (5, 000 Twitter, Reddit, and Facebook posts) and propose a two-phase scheme: first, they classify the type of spoiler required (short sentence, short passage or multipart) with around 80 percent accuracy; second, they generate the spoiling using a question answering system based on DeBERTa-large [16]. This QA model vastly outperforms simple retrieval strategies, establishing a new state of the art in 2022 [17]. In the same year, Johnson et al. collected data from the Facebook pages StopClickbait and the Reddit forum SavedYouAClick to train two parallel approaches: an extractive model (RoBERTa [18] fine-tuned for QA) and an abstractive model (T5 generative). Comparing the two strategies, they find that extractive RoBERTa obtains a slightly higher ROUGE, while generative T5 achieves a marginally better BERTScore, although both alternatives require domain-specific tuning. This work consolidates the initial foundations for automatic spoiling in English using QA and generative modeling [19].
By 2023, several groups extend these approaches and overlap spoiler generation with multitask learning techniques and semi-supervised strategies. Maharani et al. address spoiling in a resourceconstrained Indonesian language context, using multilingual models such as XLM-RoBERTa [20], which are large in zero-shot mode. Although they do not report exact metrics in the paper, they show that these multilingual models produce competitive sentence- and passage-type spoilers, while mDeBERTa-base excels in generating multipart spoilers [21]. In the same SemEval-2023 call (Task 5), Sterz et al. (ML Mob team) propose a semi-supervised, multitask learning system that combines spoiler-type classification with the generation itself. Thanks to synthetic data distillation and joint training, they achieve an F1 score of 51.48%, one of the best oficial scores for the spoiler extraction subtask, and demonstrate that the joint task improves the coherence and completeness of the resulting text[22]. Also, this year, Pal et al. present a multitask model that simultaneously detects clickbait, categorizes the type of spoiler, and generates it. They incorporate a modified QA component and adjust the LongT5 model for long sequences, obtaining approximately 81.5% in BLEU-4 for multi-spoiler generation, which is equivalent to an increase of 60% over previous baselines in the Webis [23] corpus. This work evidences the transition from purely extractive methods to solutions that integrate classification and generation in a single learning framework.
In 2024, Panda et al. (2024) propose a hybrid framework for English with two distinct modules: a ifne-tuned GPT-2 (medium) for spoiler generation (achieving BLEU = 0.58) and a BERT + SVM for spoiler type classification (F1-score 0.80, Accuracy approximately 83%) [24]. Later that year, Woźny and Lango experiment with an ensemble of LLMs tuned to produce multipart spoilers, turning each clickbait into a question and generating multiple candidates that a final evaluator model selects. This approach consolidates the notion that combining several generative perspectives enhances the coverage of hidden content, outperforming automatic metrics such as BLEU, METEOR, and BERT-Score on baselines [25]. At the same time, Garcia-Ferrero and Altuna fill the gap in Spanish by introducing NoticIA, a dataset of 850 articles with clickbait headlines and “ultra-summaries” written by humans. In addition, they tune three versions of the ClickbaitFighter model (2B, 7B, and 10B parameters) on NoticIA, achieving near-human performance in summary quality and outperforming generic LLMs in zero-shot mode. This work lays the foundations for the Shared Task TA1C in Spanish, demonstrating the importance of resources and models adapted to the Spanish-speaking domain [26].
Finally, in 2025, Panda et al. resume their line of research with an extractive model based on local explainability using LIME. They identify the decoy words in the headline and extract the relevant sentences from the article body, obtaining BLEU = 0.61 and ROUGE = 0.72, which outperforms previous models. This approach highlights the contribution of XAI (Explainable Artificial Intelligence) techniques to guide content extraction and generate more relevant and interpretable spoilers [27].
This work tackles the TA1C 2025 shared task with two complementary Python pipelines: (i) a detection ensemble that decides whether a teaser is clickbait, and (ii) a spoiling module that generates the concise answer (“spoiler”) demanded by the teaser.
This section presents a comprehensive analysis of the TA1C clickbait detection dataset [
The TA1C dataset exhibits a moderate class imbalance, as illustrated in Figure 1. The dataset comprises 2,800 total instances, with non-clickbait content representing the majority class at 64.2% (1,798 instances), while clickbait content constitutes 35.8% (1,002 instances) of the dataset. This distribution reflects a realistic representation of online media content, where genuine news articles typically outnumber sensationalized clickbait headlines; however, the presence of clickbait remains substantial enough to warrant the implementation of detection mechanisms.
The observed class distribution is advantageous for machine learning applications, as it provides a suficient representation of both classes while maintaining the natural prevalence patterns found in real-world scenarios. The imbalance ratio of approximately 1.8:1 (non-clickbait to clickbait) falls within acceptable bounds for classification tasks and does not necessitate aggressive resampling techniques that could introduce artificial bias.
Analysis of textual features reveals some diferences in content length between clickbait and nonclickbait articles, as presented in Figure 2. The text length distribution analysis provides crucial insights into the linguistic patterns that characterize each content type.
The density plots reveal several key observations: 1. Distribution Shape: Non-clickbait content exhibits a broader, more right-skewed distribution with a longer tail extending toward higher character counts. 2. Central Tendency: Clickbait content demonstrates a more concentrated distribution with a pronounced peak at approximately 100–150 characters, indicating a preference for concise, attention-grabbing headlines optimized for social media sharing. 3. Range and Variability: Non-clickbait articles span a wider range of character lengths, extending beyond 300 characters in some instances, while clickbait content rarely exceeds 250 characters. This reflects the strategic brevity employed in clickbait headlines to maximize engagement within platform constraints. 4. Bimodal Characteristics: The non-clickbait distribution shows subtle bimodal tendencies, suggesting the presence of both brief news alerts and more detailed article teasers within the legitimate content category.
These distributional diferences provide empirical evidence for the hypothesis that clickbait content is systematically engineered for brevity and immediate impact, while legitimate news content prioritizes informational completeness over character economy. The distinct length patterns observed here serve as valuable features for automated clickbait detection algorithms.
This section presents a comprehensive analysis of the TA1C clickbait spoiling dataset [
The fundamental challenge in clickbait spoiling lies in generating responses that provide suficient information to satisfy user curiosity while maintaining conciseness and clarity. Figure 3 illustrates the distributional characteristics of clickbait teasers compared to their corresponding spoiling responses across the 300 instances in our dataset.
The analysis reveals several critical observations regarding the teaser-response dynamic: 1. Distribution Overlap and Divergence: The teaser text distribution exhibits a concentrated peak of around 150–200 characters, consistent with social media optimization constraints and clickbait brevity principles. In contrast, spoiling responses demonstrate a broader, more right-skewed distribution with a pronounced tail extending toward higher character counts, indicating greater variability in response generation strategies. 2. Response Expansion Patterns: The majority of spoiling responses exceed their corresponding teaser lengths, with the response distribution showing a secondary peak of around 250–300 characters. This pattern suggests that efective spoiling requires elaboration beyond the original teaser length to provide meaningful content resolution. 3. Density Concentration Diferences: While teaser texts show high-density concentration within a narrow range (indicating standardized clickbait formulation), spoiling responses exhibit lower peak density but broader coverage, reflecting the diverse approaches humans employ when crafting explanatory content. 4. Tail Behavior: The extended tail in the response distribution, reaching beyond 500 characters in some instances, indicates that certain clickbait topics require substantial elaboration for efective spoiling, particularly for complex or nuanced subject.
These distributional diferences provide empirical evidence for the hypothesis that efective clickbait spoiling requires strategic content expansion while maintaining informativeness. The observed patterns suggest that spoiling responses typically require 1.5-2 times the character count of the original teaser to achieve satisfactory content resolution.
Understanding how spoiling responses relate to original article content is crucial for developing automated spoiling systems. Figure 4 presents a scatter plot analysis examining the relationship between full article lengths and their corresponding spoiling response lengths, revealing patterns in content compression and information distillation.
The compression analysis yields significant insights into the spoiling process: 1. Compression Ratio Consistency: The linear trend line demonstrates a consistent compression ratio across diferent article lengths, with spoiling responses typically representing 0.8-1.2% of the original article length. This consistency suggests that human annotators employ systematic approaches to content distillation, regardless of the volume of source material. 2. Scale Independence: The relationship maintains linearity across the full range of article lengths (from 2,000 to over 25,000 characters), indicating that the compression strategy scales efectively with content complexity and volume. This finding has significant implications for automated spoiling systems, suggesting that response length can be predicted based on the characteristics of the source article. 3. Variance Patterns: The scatter around the trend line reveals controlled variance in response generation, with most points clustering within a predictable band around the regression line. This controlled variance suggests that while individual spoiling strategies may vary, the overall approach remains within bounded parameters. 4. Outlier Analysis: Notable outliers above the trend line represent instances where longer responses were deemed necessary, potentially indicating articles with complex narratives, multiple key points, or nuanced conclusions requiring additional elaboration for efective spoiling. 5. Lower Bound Constraints: The absence of data points significantly below the trend line suggests a minimum information threshold for efective spoiling, below which responses would be insuficient to address the clickbait curiosity gap.
To solve task 1, we propose the following ensemble as depicted in Fig. 5
Pre–processing and Feature Design. Tweets are first normalized through a Spanish-aware cleaning routine that removes diacritics, user mentions, hashtags, and URLs and that canonicalizes interrogative and exclamation punctuation. A bespoke lexical feature extractor then encodes seven binary or bounded real attributes that capture surface cues of clickbait (question marks, interjections, interrogative pronouns, hyperbole, urgency, promises, and numeral phrases). These handcrafted signals are later concatenated to the contextual embeddings produced by a transformer backbone, and prove particularly valuable in short tweets that lack syntactic depth.
classification. Five stratified folds drive a nested ensemble, each trained for at most five epochs with early stopping. To counter class imbalance, every odd mini-batch minimizes weighted cross-entropy, whereas every even mini-batch uses Focal Loss (=2 , =0.25 ), thus dynamically emphasizing hard positive instances. A cosine learning–rate schedule with warm-up and per-parameter weight decay further stabilizes optimization. Gradient clipping at ℓ2=1 prevents exploding updates. Data Augmentation. Positive (clickbait) instances undergo two stochastic augmentations: synonyms are swapped through a curated Spanish lexicon, and one random token is dropped in long teasers. The policy roughly doubles the minority class without altering semantics, and introduces linguistic variety to mitigate overfitting.
Inference and Threshold Calibration. Each fold exports its best checkpoint (the one with the highest validation 1 score). At prediction time, probabilities are averaged with weights proportional to those 1 scores. A second optimization scans thresholds ∈ [0.1, 0.9] on the validation split to maximize macro-1, yielding a per-model ⋆ stored alongside performance metadata.
To solve task 2, we propose the following pipeline as depicted in Fig. 6 1Early experiments with microsoft/deberta-v3-small were replaced after pilot runs showed better robustness on Spanish data.
LLM Prompt Engineering. The spoiling task leverages claude-sonnet-4-20250514 through the Anthropic API. The script first ingests all training spoilers and auto-categories them into five semantically homogeneous buckets (questions, WH-questions, price mentions, factual statements, no-answer) using light keyword heuristics. Given a new teaser, it retrieves up to =15 exemplars prioritizing bucket relevance (e.g., a price teaser preferentially draws from the price bucket). The final prompt packs: (i) a role header with five mandatory rules (Respond exactly what the tweet suggests, DO NOT add context or explanations, Be EXTREMELY concise and to the point, If there is no response in the article: “No response”, Maximum 280 characters); (ii) numbered reference pairs {tweet, article snippet, answer}; and (iii) the target instance. Article JSON is parsed on the fly to strip boilerplate and truncate to 2,000 chars for cost control.
Generation Loop and Robustness. Temperature is fixed to 0.1 for determinism and BLEU maximization. If a transient API overload (HTTP 529) occurs, the script retries with exponential back-of. Outputs exceeding 280 characters are forcibly clipped; empty or single-character generations default to “No responde”. To respect the evaluation protocol, each prediction pauses for 0.4 s, staying well inside the 100-QPM rate limit.
The detector combines transformer representations with explicit linguistic priors, achieving a balanced precision-recall trade-of through hybrid losses and adaptive thresholds. Meanwhile, the spoiler generator demonstrates that many-shot prompting—when driven by linguistically informed retrieval—can compress the essence of a full article into a tweet-sized answer without resorting to fine-tuning. Together, these modules constitute a language-specific baseline for TA1C that is easily extendable to neighboring tasks such as stance detection or rumor verification.
In this section, we show the results obtained from our methodology for solving tasks 1 and 2.
Table 1 reports the central evaluation figures for the binary classification subtask. The ensemble yields an accuracy of 0.7385, correctly labeling almost three-quarters of the test tweets. Class imbalance is handled efectively, as evidenced by a recall of 0.8623, which shows the model retrieves the vast majority of true clickbait instances. Precision and recall trade of to a macro– 1 score of 0.7956, underscoring the benefit of the hybrid loss schedule and handcrafted lexical features described in Section 3.
Metric Score
Accuracy 0.7385
Recall 0.8623
For the sentence-level spoiler generation subtask, the system achieves a BLEU score of 0.4281 (Table 2). While automatic metrics only partially capture human-perceived answer quality, this value indicates that the many-shot prompting strategy (Section 3) successfully produces concise spoilers that overlap with the reference answers in more than 42 % of the -gram space.
Metric Score
BLEU 0.4281
These results validate the design choices of the dual-pass–pipeline architecture: explicit linguistic priors paired with transformer representations for detection and linguistically informed example retrieval for generation. Together, they provide a solid foundation for future research on multilingual clickbait detection and mitigation.
The experimental findings corroborate the design hypotheses that guided the two-stage architecture introduced in Section 3. For Task 1, the detector’s hybrid representation acts as high-precision anchors for tweets whose semantic content is terse or figurative, while the transformer backbone absorbs subtler pragmatic hints. This synergy explains the markedly high recall (0.8623): the model rarely overlooks an actual clickbait instance because at least one feature subspace fires even under noisy wording. The unavoidable trade-of is a pool of borderline false positives that caps overall accuracy at 0.7385; nevertheless, the macro–1 of 0.7956 attests to a balanced precision-recall compromise and vindicates the alternating cross-entropy+Focal Loss training regime that explicitly rewards correct minority-class predictions while preventing over-confidence on easy negatives. Threshold calibration a posteriori further tightened the precision gap, contributing roughly two percentage points to the final 1.
For Task 2, the spoiler generator reached a BLEU of 0.4281, a competitive score given that no gradient updates were performed on the language model itself. The many-shot prompting strategy is chiefly responsible for this performance. By presenting the LLM with tightly aligned precedents, the prompt narrows the solution manifold and discourages stylistic drift, thereby increasing -gram overlap with the gold answers.
The authors gratefully acknowledge the Instituto Politécnico Nacional (Secretaría Académica, COFAA, SIP under Grant 20230140 and 20251352, Centro de Investigación en Computación) and the Secretaría de Ciencia, Humanidades, Tecnología e Innovación (SECIHTI) for their economic support to develop this work.
During the preparation of this work, the author(s) used Claude (Anthropic) in order to: analyze computational linguistics tasks, interpret the spoiling content, assist in the generation of a proper spoil, and explore methodological approaches for the TA1C 2025 shared tasks to evaluate the scope of generative AI models in research assistance. The author(s) also used ChatGPT and Grammarly in order to: grammar and spelling check, paraphrase and reword. After using these tools/services, the author(s) reviewed and edited the content as needed and take(s) full responsibility for the publication’s content. [15] B. Gamage, A. Labib, A. Joomun, C. H. Lim, K. Wong, Baitradar: A multi-model clickbait detection algorithm using deep learning, 2025. URL: https://arxiv.org/abs/2505.17448. arXiv:2505.17448. [16] P. He, X. Liu, J. Gao, W. Chen, Deberta: Decoding-enhanced bert with disentangled attention, 2021.
URL: https://arxiv.org/abs/2006.03654. arXiv:2006.03654. [17] M. Hagen, M. Fröbe, A. Jurk, M. Potthast, Clickbait spoiling via question answering and passage retrieval, 2022. URL: https://arxiv.org/abs/2203.10282. arXiv:2203.10282. [18] Y. Liu, M. Ott, N. Goyal, J. Du, M. Joshi, D. Chen, O. Levy, M. Lewis, L. Zettlemoyer, V. Stoyanov, Roberta: A robustly optimized bert pretraining approach, 2019. URL: https://arxiv.org/abs/1907. 11692. arXiv:1907.11692. [19] O. Johnson, B. Lou, J. Zhong, A. Kurenkov, Saved you a click: Automatically answering clickbait titles, 2022. URL: https://arxiv.org/abs/2212.08196. arXiv:2212.08196. [20] A. Conneau, K. Khandelwal, N. Goyal, V. Chaudhary, G. Wenzek, F. Guzmán, E. Grave, M. Ott, L. Zettlemoyer, V. Stoyanov, Unsupervised cross-lingual representation learning at scale, 2020.
URL: https://arxiv.org/abs/1911.02116. arXiv:1911.02116. [21] N. P. I. Maharani, A. Purwarianti, A. F. Aji, Low-resource clickbait spoiling for indonesian via question answering, 2023. URL: https://arxiv.org/abs/2310.08085. arXiv:2310.08085. [22] H. Sterz, L. Bongard, T. Werner, C. Poth, M. Hentschel, ML mob at SemEval-2023 task 5: “breaking news: Our semi-supervised and multi-task learning approach spoils clickbait”, in: A. K. Ojha, A. S. Doğruöz, G. Da San Martino, H. Tayyar Madabushi, R. Kumar, E. Sartori (Eds.), Proceedings of the 17th International Workshop on Semantic Evaluation (SemEval-2023), Association for Computational Linguistics, Toronto, Canada, 2023, pp. 1818–1823. URL: https://aclanthology.org/ 2023.semeval-1.251/. doi:10.18653/v1/2023.semeval-1.251. [23] S. Pal, S. Das, R. K. Srihari, Mitigating clickbait: An approach to spoiler generation using multitask learning, in: J. D. Pawar, S. Lalitha Devi (Eds.), Proceedings of the 20th International Conference on Natural Language Processing (ICON), NLP Association of India (NLPAI), Goa University, Goa, India, 2023, pp. 486–490. URL: https://aclanthology.org/2023.icon-1.43/. [24] I. Panda, J. P. Singh, G. Pradhan, K. Kumari, A deep learning framework for clickbait spoiler generation and type identification, Journal of Computational Social Science 7 (2024) 671–693. URL: https://doi.org/10.1007/s42001-024-00252-z. doi:10.1007/s42001-024-00252-z. [25] M. Woźny, M. Lango, Generating clickbait spoilers with an ensemble of large language models, in: Proceedings of the 16th International Natural Language Generation Conference, Association for Computational Linguistics, 2023, p. 431–436. URL: http://dx.doi.org/10.18653/v1/2023.inlg-main.32. doi:10.18653/v1/2023.inlg-main.32. [26] I. García-Ferrero, B. n. Altuna, NoticIA: A clickbait article summarization dataset in spanish,
Procesamiento del Lenguaje Natural (2024). [27] I. Panda, J. P. Singh, G. Pradhan, Local explainability-based model for clickbait spoiler generation, Journal of Computational Social Science 8 (2024) 4. URL: https://doi.org/10.1007/ s42001-024-00329-9. doi:10.1007/s42001-024-00329-9.