This paper presents our submission to the SatiSPeech 2025 shared task at IberLEF, which focuses on satire detection using unimodal (text-only) and multimodal (text + audio) approaches. For the text classification task, we used TF-IDF features, BETO embeddings, and shallow models such as Logistic Regression, SVM, and XGBoost. In the multimodal setting, audio features extracted via MFCCs were processed by a CNN and fused with text features (TF-IDF + SVD and BETO) before being passed to classifiers including MLP and other shallow models. A key aspect of our system was the independent fine-tuning of the CNN, allowing it to act as a specialized expert on audio data within the final ensemble. We applied a voting ensemble to combine model predictions. Our system ranked 4th in the oficial evaluation phase and improved to 2nd place in the post-evaluation phase, highlighting the efectiveness of combining shallow and deep learning techniques, multimodal fusion, and targeted CNN optimization for satire detection in Spanish.
SatiSPeech at IberLEF 2025 [
The SatiSPeech 2025 competition consists of two subtasks. Task 1: Text Satire Detection focuses on classifying Spanish-language text segments as either satirical or non-satirical. This task requires models to identify linguistic patterns indicative of satire, such as wordplay, sarcasm, or implicit social critique [5, 6]. Task 2: Multimodal Satire Detection expands this challenge by incorporating audio data, enabling participants to exploit speech characteristics such as intonation, rhythm, and vocal emphasis in combination with textual content. This multimodal approach is particularly relevant, as satire often manifests through the interplay of spoken delivery and written language [7, 8].
For data preprocessing, we adopted distinct strategies for the text and audio modalities to best preserve the features relevant to satire detection.
For the textual data, we applied tokenization to segment each utterance into linguistically meaningful units, ensuring that important elements such as punctuation and word boundaries-often crucial for conveying irony or sarcasm-were retained. Following tokenization, we standardized the resulting feature vectors using a standard scaler. This normalization step was intended to enhance the stability and convergence of downstream machine learning models by ensuring that all features contributed equally during training.
In contrast, for the audio modality, we opted to use the original, unscaled audio features as input to our models. Satirical speech often relies on subtle prosodic and acoustic cues-such as intonation, rhythm, and emphasis-that can be diminished or lost through aggressive normalization or scaling. By preserving the raw audio features, we aimed to maintain the richness of these cues, allowing our models to fully exploit the nuanced characteristics of satirical speech. This approach ensured that both textual and acoustic information were optimally prepared for the subsequent classification tasks.
We explored two vectorization strategies: TF-IDF, which captures surface-level lexical patterns and serves as a strong baseline for linear models, and BETO, a pre-trained Spanish BERT model, which encodes deep contextualized features. These embeddings were then fed into diferent classifiers to study both linear and non-linear decision boundaries.
To improve performance and robustness, we experimented with multiple model combinations, optimized their hyperparameters using Optuna[9], and finally applied a simple ensemble strategy to aggregate predictions from the best-performing models.
We used two diferent strategies to convert input text into numerical representations: • TF-IDF (Term Frequency–Inverse Document Frequency): This classical vectorization method assigns a weight to each word based on its frequency across documents. It helps capture surfacelevel lexical cues that may be indicative of satire (e.g., rare or exaggerated word usage). The TF-IDF vectors were used with simple linear classifiers to test the generalizability of shallow features in satire detection. • BETO Embeddings: We used the Spanish pre-trained BETO [10] model to extract contextualized embeddings from the input text. Specifically, the token representation from the final hidden layer was used as a fixed-size sentence vector. These embeddings aim to capture deeper semantic patterns and contextual dependencies typical of satirical language.
We employed three classification configurations: • TF-IDF + Logistic Regression: This setup serves as a strong and interpretable baseline to evaluate whether surface-level lexical cues are suficient for satire classification. By applying logistic regression to TF-IDF features, the model learns a linear decision boundary based on word frequency patterns. While this approach lacks semantic understanding, it ofers insight into how much information can be captured from basic lexical statistics alone. It is also computationally eficient and robust to small training data, making it suitable for initial prototyping and benchmarking. • BETO + Logistic Regression: This configuration combines the power of contextualized word embeddings with a simple linear classifier. BETO model—a BERT variant pre-trained on a large Spanish corpus—we aim to capture deeper semantic and syntactic features inherent in satirical expressions. Logistic regression on top of BETO embeddings helps to test whether these learned representations are linearly separable and whether a simple classifier can exploit the richness of deep contextual cues without introducing model complexity. • BETO + XGBoost: This configuration is designed to model the non-linear and complex relationships embedded within the high-dimensional BETO features. Unlike logistic regression, which assumes a linear boundary, XGBoost leverages an ensemble of decision trees to capture intricate patterns and interactions that are often present in satirical language—such as implicit irony, multi-level semantic cues, and context-dependent humor. The use of XGBoost enables the model to adapt to subtle and non-obvious characteristics that are critical for efective satire detection, particularly in cases where linear models may underperform due to oversimplified assumptions.
As such, this setup serves as the most expressive model in our pipeline.
Hyperparameters for each model were optimized using Optuna, a hyperparameter tuning framework based on Bayesian optimization. Each configuration was trained and validated independently.
Finally, we applied a soft voting ensemble at figure 1 to combine the predictions of the three models. Specifically, we assigned a weight of 0.2 to the TF-IDF + Logistic Regression model and a weight of 0.4 to each of the two BETO-based models: BETO + Logistic Regression and BETO + XGBoost. This weighted strategy allows the ensemble to place greater emphasis on the deep contextual features learned by BETO, while still retaining the complementary lexical cues captured by TF-IDF. The ensemble improved performance and robustness on the validation set by leveraging the strengths of both shallow and deep representations, as well as combining linear and non-linear decision boundaries.
For Task 2, we follow a multimodal pipeline at figure 2 that processes audio and text in parallel. Audio features are extracted and embedded through a CNN-based module, while textual features are obtained via embedding techniques. The resulting representations are concatenated into a joint feature vector and passed through a multimodal classification block. Final predictions are produced using a Hard-voting ensemble to enhance overall performance.
For textual representation, we employed two encoding strategies to transform the input into fixed-size vectors suitable for multimodal integration: • TF-IDF + SVD: This method uses Term Frequency-Inverse Document Frequency (TF-IDF) to capture the importance of words across the corpus. As TF-IDF typically yields high-dimensional sparse vectors, we apply Singular Value Decomposition (SVD) to reduce the dimensionality to 300. This step not only improves computational eficiency but also ensures compatibility with downstream neural components such as the MLP, which perform better with dense, fixed-length inputs. • BETO Embeddings: We utilize BETO, a BERT-based pretrained language model for Spanish.
Sentence-level embeddings are obtained by mean pooling over token representations from the ifnal hidden layer, resulting in 768-dimensional dense vectors that encode rich contextual and semantic information.
These vector representations are later concatenated with the audio embeddings in the multimodal pipeline for final classification.
To capture relevant acoustic cues for satire detection, we employed a Convolutional Neural Network (CNN) architecture at figure 3 to process Mel Frequency Cepstral Coeficients (MFCCs), a widely adopted representation of speech signals. MFCCs ofer a perceptually grounded, compact encoding of the short-term spectral envelope of audio, efectively capturing timbral and prosodic nuances that are potentially informative for detecting satirical tone.
Our CNN architecture is designed to extract robust feature representations from the MFCC matrices, treating them as 2D inputs over time and frequency. The network consists of two convolutional blocks, each composed of a 2D convolutional layer followed by batch normalization and ReLU activation. The ifrst block uses 16 filters while the second uses 32, allowing the model to progressively learn more abstract and complex patterns in the spectro-temporal domain. These blocks are followed by an adaptive average pooling layer, which reduces the spatial dimensions to a fixed-size 4 ×4 output, regardless of the input length — a crucial design choice to accommodate variable-length utterances during inference.
The pooled feature maps are then flattened and passed through a fully connected (linear) layer with 64 hidden units, acting as a feature bottleneck. Finally, a classification head maps the 64-dimensional representation to a two-class output (satire vs. non-satire). This design efectively balances model complexity and computational eficiency, enabling the system to learn discriminative audio features without overfitting.
By learning directly from low-level MFCC inputs, the CNN is capable of capturing non-trivial acoustic cues — such as exaggerated prosody, irregular rhythm, or tonal patterns — that may signal satirical intent. These audio-based features were later integrated with text-based features in our multimodal system to further enhance satire detection performance.
The final multimodal feature vector, formed by concatenating audio and text embeddings, is passed to a classification model to predict whether an utterance is satirical or not. We explored three classification approaches: • Multi-Layer Perceptron (MLP): This model consists of a two-layer feedforward neural network at figure 4. Each hidden layer includes a linear transformation, followed by Batch Normalization and ReLU activation. The final layer is a linear projection to the output space. This architecture, shown in Figure 2.3.3, is designed to capture non-linear relationships between the multimodal features. • Logistic Regression: As a simpler alternative, we also experimented with a linear logistic regression classifier directly applied to the concatenated multimodal vector. This model serves as a lightweight yet efective approach in cases where the feature space is already well-structured. • Support Vector Machine with RBF Kernel: To model more complex decision boundaries, we employed a non-linear SVM with a radial basis function (RBF) kernel. This method is particularly suited for handling the fused feature space where interactions between modalities may not be linearly separable.
To improve robustness and overall performance, we employed a hard voting ensemble that combines three classifiers: a Multi-Layer Perceptron (MLP), Logistic Regression, and a Support Vector Machine with RBF kernel. All use the same multimodal input—concatenated text embeddings from BETO and audio embeddings from a CNN.
A key design choice is the training duration of the CNN. While the MLP is trained for 20 epochs, the CNN is fine-tuned for 100 epochs to produce more stable and refined audio features. This extended training aims to improve the quality of audio embeddings and strengthen their influence within the ensemble.
Final predictions are made via hard voting, where each model votes and the majority class is selected. This strategy stabilizes results and outperformed our expectations on the evaluation set.
To support these tasks, the organizers compiled a diverse dataset of annotated audio-text pairs from various Spanish-language sources, including satirical programs and news broadcasts [11]. The data was segmented and transcribed using state-of-the-art diarization and speech recognition systems [12, 13]. Despite growing interest in satire and irony detection, most prior studies have focused on unimodal text analysis [4, 5, 6]. Recent works have shown the promise of leveraging large language models and multimodal fusion techniques for satire detection, but research in Spanish—especially with spoken content—remains limited [8].
The dataset includes 6,000 training samples, 384 validation samples, and 2,000 unlabeled test samples, with a nearly balanced class distribution, as shown in Table 1.
For Task 1, we used the oficial datasets provided by the organizers to train our models. To facilitate a comprehensive understanding of the data, we present both the distribution summary and a sequence length analysis. Table 1 shows the data split across training, validation, and test sets. The task involves classifying each sample as either Satire or Non-satire. The training set contains 2,832 satire and 3,168 non-satire examples; the validation set contains 178 satire and 206 non-satire examples. The test set comprises 2,000 unlabeled samples.
Although a mild class imbalance exists, both categories are suficiently represented to enable efective model training. These distributions serve as a foundation for both initial training and hyperparameter tuning.
Figure 6 illustrates the distribution of sequence lengths (i.e., number of words per sample) across the two classes. The violin plots highlight density distributions, with both classes showing central concentration between 40–70 words. However, the Satire class exhibits a wider range, extending up to 220 words. This suggests that satirical utterances may involve longer or more elaborate expressions than factual ones. The median sequence lengths remain comparable across both classes, but the broader spread in the satire class indicates higher variability.
For Task 2, we extended the satire detection task into a multimodal setting by incorporating raw audio alongside the text transcripts. While the dataset splits and class distributions remained identical to Task 1 (see Table 1), each sample was paired with an audio recording.
Text preprocessing followed the same pipeline as in Task 1, including tokenization and normalization. For audio, we extracted features such as Mel-frequency cepstral coeficients (MFCCs) directly from the raw waveform to retain nuanced prosodic cues, such as exaggerated intonation or sarcastic pacing—features essential for capturing satirical speech.
Model performance was evaluated using precision, recall, and macro F1-score. As in Task 1, macro F1 was used as the primary metric due to its ability to equally account for both classes under slight imbalance. This consistent evaluation protocol allows meaningful comparisons between unimodal and multimodal approaches.
For the text modality, we used two types of vector representations: BETO embeddings with 768 dimensions and TF-IDF vectors reduced to 300 dimensions using Singular Value Decomposition (SVD). These representations were used consistently across all text-based models.
For shallow classifiers such as Logistic Regression and RBF-SVM, we employed Optuna to perform automatic hyperparameter optimization on the validation set.
For the audio modality, the CNN used to generate audio embeddings was trained for 20 epochs with a batch size of 16. The downstream MLP classifier was trained for 20 epochs with a batch size of 32. Both the CNN and MLP were optimized using the Adam optimizer, with a fixed learning rate of 0.001.
In the ensemble setup described for Task 2, we further fine-tuned the CNN independently for 100 epochs—five times longer than the MLP training schedule—to obtain more refined and stable audio feature representations. This extended training was motivated by the need to enhance the contribution of the audio stream within the final hard voting ensemble.
Model Logistic Regression (TF-IDF) SVM (TF-IDF) Logistic Regression (BETO) XGBoost (BETO) Final Ensemble (method 1+3+4 with Soft Voting)
As shown in Table 4.1, the use of TF-IDF embeddings combined with Logistic Regression achieved an F1-score of 0.8047, outperforming SVM (0.7929). This result suggests that even simple linear classifiers, when combined with sparse lexical representations, can capture a significant portion of satire-related patterns. However, these models may still be limited in capturing nuanced contextual information, which is often essential in satire.
Where BETO embeddings were used, we observed an improvement in performance. Logistic Regression with BETO achieved the highest F1-score of 0.8291, while XGBoost slightly lagged at 0.8053. This indicates that the BETO embeddings encode useful semantic and syntactic signals for satire detection, and that a simple linear decision boundary is already efective when applied to rich contextual representations. The relatively smaller gain from using XGBoost suggests that BETO embeddings are already linearly separable to a large extent.
Finally, the table presents the F1-scores of the three individual models alongside the result of their soft voting ensemble. The ensemble method combines BETO + Logistic Regression, BETO + XGBoost, and TF-IDF + Logistic Regression with respective weights of 0.4, 0.4, and 0.2. The final ensemble achieved the best overall performance with an F1-score of 0.8345, outperforming all individual models. This confirms that the models capture complementary patterns—BETO contributes deep semantic information, while TF-IDF captures surface-level lexical features that may still be informative in certain satirical contexts. The ensemble thus balances linear and non-linear decision boundaries, as well as shallow and deep features, leading to a more robust and generalizable classifier.
Among the individual models, the combination of BETO embeddings with an MLP classifier yielded the highest performance, achieving an F1-score of 0.8359. Building upon this, the final hard voting ensemble—which integrates predictions from BETO-based models and the independently fine-tuned CNN—achieved the best overall F1-score of 0.86164, demonstrating enhanced robustness and generalization through multimodal model aggregation.
This paper presented our approach to the SatiSPeech 2025 shared task on satire detection in Spanish. For Task 1, which focused on text-based satire classification, we adopted a combination of shallow learning techniques and pretrained language models. Specifically, we used BETO embeddings with XGBoost and Logistic Regression to capture deep contextual semantics, while TF-IDF combined with Logistic Regression was used to model broader, surface-level lexical patterns. This dual strategy allowed us to exploit both fine-grained and general textual cues.
In Task 2, which incorporated both audio and text modalities, we used a CNN architecture to extract features from MFCC representations of the audio signal. These were then fused with text features and passed to an MLP for classification. Notably, we took the same CNN used for feature extraction and fine-tuned it independently on audio data for 100 epochs—five times longer than the default MLP schedule. This extended training aimed to enhance the audio representation, and the resulting model was added to the ensemble, contributing significantly to performance.
By combining shallow models, deep embeddings, multimodal fusion, and a custom-trained CNN audio encoder, our system achieved strong performance—ranking 4th in the oficial evaluation and climbing to 2nd in the post-evaluation phase. The improvement was largely driven by our novel idea of introducing a separately fine-tuned CNN dedicated to the audio stream in the ensemble. This result demonstrates the power of tailored multimodal architectures and ensemble learning in complex tasks like satire detection.
During the preparation of this work, the author(s) used ChatGPT and Grammarly in order to improve writing style, perform grammar and spelling checks, and provide formatting assistance. Specifically, these tools were employed to ofer suggestions for sentence structure and word choice to enhance overall flow, to identify and correct grammatical errors and typographical mistakes that may have been overlooked, and to ensure the paper adheres to the specific formatting guidelines required by the conference proceedings. 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.
This research was supported by The VNUHCM-University of Information Technology’s Scientific Research Support Fund. Languages Evaluation Forum (IberLEF 2025), co-located with the 41st Conference of the Spanish Society for Natural Language Processing (SEPLN 2025), CEUR-WS.org, 2025. [3] T. Jiang, H. Li, Y. Hou, Cultural diferences in humor perception, usage, and implications, Frontiers in Psychology 10 (2019). [4] M. Salas-Zárate, G. Alor-Hernández, J. L. Sánchez-Cervantes, M. Paredes-Valverde, J. L. GarcíaAlcaraz, R. Valencia-García, Review of English literature on figurative language applied to social networks, Knowledge and Information Systems 62 (2020) 2105–2137. [5] L. Li, O. Levi, P. Hosseini, D. Broniatowski, A multi-modal method for satire detection using textual and visual cues, in: Proceedings of the 3rd NLP4IF Workshop on NLP for Internet Freedom, 2020, pp. 33–38. [6] R. Ortega-Bueno, P. Rosso, J. E. M. Pagola, Multi-view informed attention-based model for irony and satire detection in Spanish variants, Knowledge-Based Systems 235 (2022) 107597. [7] H. Bredin, A. Laurent, End-to-end speaker segmentation for overlap-aware resegmentation, in:
Proceedings of Interspeech 2021, 2021, pp. 3111–3115. [8] G. Wick-Pedro, C. F. da Silva, M. L. Inácio, O. A. Vale, H. de Medeiros Caseli, Using large language models for identifying satirical news in Brazilian Portuguese, in: Proceedings of the 16th International Conference on Computational Processing of Portuguese, 2024, pp. 156–167. [9] T. Akiba, S. Sano, T. Yanase, T. Ohta, M. Koyama, Optuna: A next-generation hyperparameter optimization framework, in: Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, 2019, pp. 2623–2631. [10] J. Cañete, G. Chaperon, R. Fuentes, J.-H. Ho, H. Kang, J. Pérez, Spanish pre-trained BERT model and evaluation data, arXiv preprint arXiv:2308.02976 (2023). [11] J. A. García-Díaz, R. Valencia-García, Compilation and evaluation of the Spanish SatiCorpus 2021 for satire identification using linguistic features and transformers, Complex & Intelligent Systems 8 (2022) 1723–1736. [12] H. Bredin, R. Yin, J. M. Coria, G. Gelly, P. Korshunov, M. Lavechin, D. Fustes, H. Titeux, W. Bouaziz, M.-P. Gill, pyannote.audio: Neural building blocks for speaker diarization, in: Proceedings of the IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), IEEE, 2020, pp. 7124–7128. [13] A. Radford, J. W. Kim, T. Xu, G. Brockman, C. McLeavey, I. Sutskever, Robust speech recognition via large-scale weak supervision, in: Proceedings of the International Conference on Machine Learning (ICML), 2023, pp. 28492–28518.