This paper presents the UKR team's participation in the SatiSpeech 2025 shared task, focused on the detection of satirical content in Spanish from both text-only and multimodal (text + audio) sources. We propose a supervised ifne-tuning approach using the Spanish monolingual BERT model (BETO) for Task 1, and extend it with MFCCbased acoustic features in Task 2 to capture prosodic information. For classification, we adapt the final layers of the transformer model to integrate textual and audio inputs in a unified architecture. Our system ranked 6th in Task 1 and 9th in Task 2, with validation macro F1 scores of 0.9648 and 0.9699, respectively. The results demonstrate that while textual information carries most of the discriminative power, audio features ofer complementary cues that slightly improve performance.
Satire is a nuanced and context-dependent form of expression that presents significant challenges for
traditional content classification systems. Unlike straightforward humor, it delivers critique indirectly
through rhetorical strategies such as irony, parody, and exaggeration. Understanding satire often
requires knowledge of cultural context and the speaker’s intent, which makes automated detection
more complex. The satirical meaning is influenced not only by the text itself but also by vocal features
like tone, pitch, and rhythm that shape how the message is communicated and received. As a result,
detecting satire in multimodal content requires models that can efectively combine both linguistic and
auditory information [
Interest in automatic satire detection has increased, driven by its potential to counter misinformation, support content moderation, and enrich media analysis. On digital platforms, satire often closely resembles genuine news, which heightens the risk of confusion and the dissemination of false narratives. Therefore, developing reliable automated systems for satire detection is essential to promote accurate interpretation of content, particularly in multilingual and culturally diverse contexts.
Traditionally, satire detection has concentrated on textual analysis. Transformer-based language
models trained on news and social media datasets have demonstrated strong performance in this area
[
Detecting satire in spoken or audiovisual content necessitates multimodal modeling due to the inherent limitations of relying on text alone. Satirical intent is often conveyed not just by the words themselves, but by how they are spoken. Prosodic features such as intonation, pitch variation, stress, and rhythm play a crucial role in signaling irony, sarcasm, or exaggeration. These vocal cues can fundamentally alter the meaning of an utterance. For instance, a deadpan or sarcastic tone can reverse the literal interpretation of the words, which would likely be misunderstood by models that analyze text in isolation. By incorporating audio, models can capture these non-verbal signals, leading to a more accurate understanding of speaker intent and more reliable satire detection in real-world, multimodal contexts.
Moreover, recent research has demonstrated that audio features can enhance performance in various
NLP tasks, such as emotion recognition, and across diferent domains [
To address the existing gap in multimodal satire detection, the SatiSpeech shared task [
We participated in both subtasks of the SatiSpeech challenge by adapting our prior approach from the
EmoSpeech task at IberLEF 2024. For Subtask 1 (text-only classification), we fine-tuned BETO, a
BERTbased model pre-trained on Spanish corpora [
For Subtask 2 (multimodal classification), we extended the architecture to integrate audio features. We extracted Mel-frequency cepstral coeficients (MFCCs) from each audio sample using librosa, resulting in 1D acoustic embeddings. These were concatenated with BETO’s contextual text embeddings to form joint multimodal vectors. A custom classification head was built to accommodate the expanded feature set, operating on top of a modified Bert model. Our implementation, structured as a subclass of BertForSequenceClassification class, combined pooled textual and acoustic representations before the final classification layer. The model was trained using binary cross-entropy loss for satire detection.
This paper is organized as follows: Section 3 provides an overview of the shared task and Section 2 a summary of related works; Section 4 describes our modeling approaches for both unimodal and multimodal configurations; Section 5 presents the experimental results and comparisons; and Section 6 concludes with insights and future directions.
The automatic detection of figurative language—such as irony, sarcasm, and satire—has garnered increasing attention in natural language processing due to its relevance in combating misinformation and supporting content moderation. Early research focused primarily on text-based features, using lexical, syntactic, and semantic cues to classify satirical or sarcastic content. Traditional machine learning approaches applied handcrafted features (e.g., n-grams, sentiment polarity) with classifiers such as SVMs or random forests. However, these methods often struggled with generalization across domains due to the context-dependent nature of satire.
Recent advancements in transformer-based language models, such as BERT and its monolingual
variants like BETO [
In the multimodal domain, several studies have explored the integration of speech and visual data
for figurative language detection. In [
The SatiSpeech 2025 shared task [
The SatiSpeech shared task, part of IberLEF 2025, focuses on the automatic detection of satirical content in Spanish using both text-only and multimodal (text + audio) inputs. This challenge stems from the subtle and context-dependent nature of satire and its growing relevance in areas such as media analysis, misinformation detection, and computational discourse understanding.
The task is organized into two subtasks: • Task 1: Text-Based Satire Detection. Participants must develop systems that classify whether a given transcript is satirical, relying solely on textual features. These may include word choice, syntactic patterns, and rhetorical devices like irony and exaggeration. • Task 2: Multimodal Satire Detection. This subtask extends the problem by incorporating audio.
Participants receive aligned audio-transcript pairs and must combine linguistic and prosodic cues—such as rhythm, intonation, and stress—for binary satire classification.
We participated in the task using the SatirA dataset, a curated collection of Spanish-language audio segments sourced from YouTube. The dataset features a diverse range of Spanish dialects and regional varieties to promote linguistic diversity and reduce potential regional bias. Audio segments were generated using automatic speaker diarization, filtering out clips longer than 25 seconds. Transcriptions were produced using Whisper. The annotation process followed a semi-supervised approach: initial automatic labels were refined by a team of three expert annotators to ensure high-quality ground truth.
The final dataset contains approximately 25 hours of labeled content and was split 80/20 into training and validation sets. All submissions and evaluations were carried out through the Codalab platform.
Table 1 summarizes the class distribution in both splits. There is a mild class imbalance, with nonsatirical samples being slightly more frequent and typically longer on average. Satirical transcripts tend to show more variation in length. While these patterns may reflect real-world stylistic diferences, models should avoid exploiting them directly during classification.
Set Train Validation
Metric
Samples Avg. Length
Std. Dev.
Samples Avg. Length
Std. Dev.
Non-satirical Satirical 3168 60.83 12.01 633 61.08 12.23 2832 55.95 17.60 567 55.83 17.15
This section describes the approach we used for both subtasks of the SatiSpeech Shared Task: text-based satire detection (Task 1) and multimodal satire detection (Task 2). Instead of a traditional SVM-based pipeline, we fine-tuned neural architectures built on top of the Spanish BERT model (BETO), using custom classification heads tailored to each modality.
Our system was developed in PyTorch using the HuggingFace Transformers library. For Task 1, we relied exclusively on textual inputs. For Task 2, we extended this model by incorporating MFCCbased acoustic features extracted from the corresponding audio segment. The full architecture for the multimodal model is illustrated in Figure 1.
For Task 1 (text-only satire detection), we fine-tuned the dccuchile/bert-base-spanish-wwm-uncased model, commonly known as BETO. The transcriptions were preprocessed using standard tokenization, with padding and truncation applied to ensure a consistent input length of 512 tokens. We extracted the 768-dimensional contextual embedding associated with the [CLS] token from the final hidden layer of BETO to serve as a fixed-size representation of each utterance, which was subsequently passed through a lightweight classification head for binary satire prediction.
This embedding was passed to a custom classification head consisting of a dropout layer followed by a linear projection to the number of output labels. The entire model, including BETO and the classification head, was fine-tuned using cross-entropy loss.
Training was conducted over 10 epochs using a batch size of 16 and a learning rate of 2 × 10− 5. We selected the best model based on weighted F1-score on a validation set (10% of the training data).
For Task 2 (multimodal satire detection), we extended the textual model by integrating prosodic information. We extracted Mel-Frequency Cepstral Coeficients (MFCCs) from each audio segment using the librosa library. These features were averaged across the temporal axis to yield a fixedlength acoustic vector of 40 dimensions. This audio representation was then concatenated with the 768-dimensional BETO text embedding.
This MFCC vector was concatenated with the 768-dimensional BETO embedding to form a 808dimensional multimodal input. We then adapted the classification head to project from this larger combined feature space.
The decision to combine BETO with MFCCs rather than using a full acoustic language model (e.g., Wav2Vec 2.0) was driven by both computational constraints and interpretability. MFCCs provide a compact and well-understood representation of prosodic and timbral characteristics of speech, which are highly relevant for capturing nuances of satirical delivery such as tone, rhythm, and emphasis.
The multimodal classification head was custom-designed to handle the concatenated embeddings eficiently. By treating the audio and text branches independently until the fusion point, the architecture maintains modularity, making it adaptable for other modalities or downstream tasks.
All training was conducted using PyTorch and Hugging Face Transformers, with evaluation based on weighted and macro F1-scores to account for mild class imbalance. The entire pipeline—from preprocessing to final prediction—is reproducible and compatible with GPU acceleration, ensuring scalability for larger datasets or multilingual adaptations.
The rest of the architecture remained the same. The model was trained for 20 epochs using a learning rate of 1 × 10− 5 and the same batch size. As in Task 1, we used cross-entropy loss and selected the best model according to weighted F1-score on the validation split.
For Task 1, the fine-tuned BETO model on Spanish satirical texts achieved a macro F1 score of 0.9648 on the validation split. This demonstrates that BETO, as a monolingual transformer pretrained on Spanish, is highly efective at capturing linguistic features relevant to satire, such as irony, hyperbole, and rhetorical structure.
For Task 2, the addition of prosodic features through MFCC vectors led to a modest improvement, with a macro F1 score of 0.9699. The gain, while small, confirms that acoustic signals—particularly related to tone, rhythm, and emphasis—can reinforce textual cues in satire detection. However, the improvement is incremental rather than transformative, suggesting that simple early fusion of features (concatenation) may not fully exploit the expressive richness of the audio modality.
As the gold labels for the test set were not publicly released, a 20% stratified split from the training data was used as a held-out validation set to better understand model performance. Table 5 provides the macro-averaged metrics for both tasks.
A class-wise breakdown revealed that the models maintain extremely high performance across both satire and no-satire classes. In particular: • Task 1 showed excellent balance, with precision and recall values near 0.99 for both classes, indicating that BETO is capable of generalizing from lexical and syntactic satire cues alone. • Task 2 demonstrated that audio features marginally improved recall for satirical utterances.
This suggests the model learned prosodic patterns such as exaggerated intonation or rhythmic anomalies common in satirical speech. • However, the improvement was not drastic, which likely reflects the limitation of using MFCC features and early fusion. Richer acoustic representations or attention-based fusion may be needed to fully leverage the audio modality.
In sum, both models achieved results well above the challenge baseline, confirming the validity of our pipeline. The performance gap between our internal validation and leaderboard scores could be explained by slight domain shift, noise in test set audio, or diferences in source domains.
To better understand the limitations of our models, we conducted a qualitative analysis of several misclassified examples from the validation set. We selected representative cases from both Task 1 (text-only) and Task 2 (multimodal) that illustrate common failure patterns.
In Task 1, many errors occurred when the satirical content employed subtle rhetorical cues such as irony, sarcasm, or absurd exaggeration that were not easily distinguishable from factual reporting. These examples often relied on cultural or contextual background knowledge that the model was unlikely to capture from text alone.
In Task 2, despite the inclusion of prosodic features through MFCCs, some predictions still failed to recognize satirical tone. This may be due to the limitations of using static acoustic embeddings and early fusion strategies, which do not fully exploit temporal speech dynamics such as intonation or speech rate.
Table 6 presents a selection of misclassified samples with their full transcriptions and classification results.
Task Task 1 Task 1 Task 1 Task 2 Task 2
Prediction no-satire
Gold Label satire no-satire no-satire no-satire satire satire satire satire no-satire
Transcription Este parque es súper importante para el movimiento del barrio, que además ahora se usa para los rodajes, que es otra fuente de ingresos. Pero claro, como lo usa la tele, ya nos lo van a cerrar.
El 5G no deja de sorprender. Ayer se pudo ver a una vecina del barrio que, tras vacunarse, sintonizaba los canales rusos con sólo tocarse la frente.
No nos responden. Luego tuve ocasión de, con los servicios secretos franceses, hacerme con una grabación de Macron escuchando la COPE mientras se duchaba.
Bueno, pues me temo que vamos a tener que dejar las fronteras abiertas otra vez. Parece que el virus ha pedido vacaciones.
Por su altura de 1 metro y 52 centímetros, nunca pensaron que sería capaz de trepar esa valla de seguridad. Pero lo logró. Y con elegancia.
This qualitative review reveals that improving satire detection may require enhanced modeling of pragmatic and cultural cues, and more sophisticated multimodal fusion strategies that account for the temporal structure of speech. Incorporating prosody-aware embeddings, pitch contours, or pre-trained acoustic transformers such as Wav2Vec2 could help address these limitations in future work.
In this paper, we presented the UKR team’s system for the SatiSpeech 2025 shared task, addressing both textual and multimodal satire detection in Spanish. Our approach relied on fine-tuning the BETO language model, with and without the integration of acoustic features derived from MFCCs. Despite using a simple early fusion strategy for the multimodal task, our system achieved strong results, placing 6th in Task 1 and 9th in Task 2 on the oficial leaderboard.
The validation results showed that BETO alone is highly capable of capturing the linguistic and rhetorical markers of satire, with a macro F1 score of 0.9648. The addition of MFCC features led to a modest gain (macro F1 of 0.9699), suggesting that prosodic cues—although helpful—were not fully leveraged with the current feature representation. These findings underscore the need for more efective fusion strategies to harness the full potential of multimodal features.
Our system outperformed the baseline by a large margin, and remained competitive among teams that adopted more complex multimodal architectures. As future work, we plan to explore richer audio embeddings (e.g., Wav2Vec 2.0 fine-tuning), attention-based fusion layers, and the use of prosodicspecific features like pitch contours and speech rate, which may reveal subtler satire signals.
Overall, our results confirm the efectiveness of Transformer-based models in satire detection and highlight the potential of integrating speech features to enhance detection in multimodal communication settings. Additionally, we aim to experiment with LLMs, given their demonstrated efectiveness in various classification tasks within domains such as hate speech and satire detection [ 10, 11]. Their capacity for contextual understanding and generalization could further enhance performance in nuanced tasks like SatiSpeech.
During the preparation of this work, the author(s) used DeepL in order to Grammar and spelling check. [9] V. Tiwari, Mfcc and its applications in speaker recognition, International journal on emerging technologies 1 (2010) 19–22. [10] A. Salmerón-Ríos, J. A. García-Díaz, R. Pan, R. Valencia-García, Fine grain emotion analysis in spanish using linguistic features and transformers, PeerJ Computer Science 10 (2024). doi:10. 7717/PEERJ-CS.1992. [11] R. Pan, J. A. García-Díaz, R. Valencia-García, Spanish mtlhatecorpus 2023: Multi-task learning for hate speech detection to identify speech type, target, target group and intensity, Computer Standards & Interfaces 94 (2025) 103990. URL: https://www.sciencedirect.com/science/article/pii/ S0920548925000194. doi:https://doi.org/10.1016/j.csi.2025.103990.