This paper describes a method based on data from Wikipedia for the automatic semantic tagging of common and proper nouns in context. We first predict the semantic category of each Wikipedia entry using a rule-based method that detects definition patterns, and then we generalize from there using a statistical model that associates semantic categories with elements of the entry. The evaluation of proper and common nouns in Spanish reveals a general precision of .82 and a recall of .77. One feature of the method is its conceptual simplicity and computational efficiency. The implementation is offered as open-source code and the data used in the study is in the public domain.
This work focuses on the automatic recognition of location entities in Spanish clinical reports, using the MEDDOPLACE challenge (IberLEF 2023) as the experimental framework. We evaluated both general-domain pre-trained models and biomedical-specific models. Furthermore, we explored data augmentation techniques via back-translation and LLM-based paraphrase generation. Our results outperform previous state-of-the-art approaches, demonstrating the effectiveness of combining these data augmentation strategies with pre-trained clinical domain models.
Early detection of mental health risks on digital platforms is essential for enabling timely interventions. However, existing metrics such as Early Risk De-tection Error (ERDE) assess detection latency by message count rather than real time, misaligning with clinical priorities that depend on the temporal dynamics of risk. We propose Time-Aware Early Risk Detection Error (T-ERDE), a metric that replaces message-count penalties with temporal cost functions based on days elap-sed since the first user interaction. T-ERDE preserves compatibility with existing metrics while offering a more clinically meaningful and equitable evaluation. Expe-riments on the mentalRiskES (Spanish) and eRisk (English) datasets demonstrate that T-ERDE provides more accurate and relevant assessments, better capturing intervention urgency and reducing biases linked to user activity volume.
Despite advances in Natural Language Processing (NLP), figurative language remains a challenge for Large Language Models (LLMs). We evaluate and compare the performance of LLMs ChatGPT-4o, DeepSeek-V3, Gemini 2.5 Pro Experimental, and Grok 3 Preview in figurative language processing tasks. Drawing on the Conventional Figurative Language Theory (CFLT) proposed by Dobrovol'skij and Piirainen (2021), we compiled CLUE, an English-Spanish dataset of metaphors, idioms, similes, and proverbs. Using this dataset, we designed and conducted tests on figurative language detection, interpretation, and generation. The models exhibited strong performance in detection (0.94) and interpretation (0.91) tasks in both languages. However, with the exception of Gemini, their performance dropped significantly in the generation task (0.70), particularly in Spanish. We conclude that LLMs are effective in figurative language comprehension tasks but remain limited in their ability to produce it.
Media bias manifests through subtle editorial, discursive, and linguistic mechanisms that shape public perception without explicit falsehoods. Research on automatic media bias detection has focused largely on English resources and on binary, document-level labels, overlooking the hierarchical and perspectivist nature of bias. This article presents the Media Bias Bias-Mitigated Dataset (MBBMD), new corpus designed to address these limitations. MBBMD integrates three annotation levels: binary and multilab el document-level bias, and fine-grained sentence level manifestations. The annotation process combines a perspectivist document level scheme, preserving annotator disagreement, with a deterministic sentence-level procedure. The dataset also incorporates systematic Counterfactual Data Augmentation (CDA), enabling analyses of how such factors influence perceived bias. The sentence-level component comprises 2,348 annotated sentences covering six linguistic manifestations of media bias.
Subjectivity Detection (SD) is a fundamental NLP task that distinguishes between factual content and opinion. Its analysis provides profound linguistic knowledge, useful for tasks such as disinformation detection. Our objective is to develop and validate a hybrid annotation framework for SD in Spanish. Our methodology includes: (1) a pilot study comparing human annotation against Machine Learning (ML) and Large Language Models (LLMs); (2) an error analysis to refine the framework; and (3) the application of the consolidated framework. Results show ML achieving higher precision, but overlooking complex structures (e.g., subjunctive). The consolidated hybrid methodology is tested in three annotation setups (human [A1], unrevised automatic [A2], consolidated semi-automatic [A3]). A3 achieves the best balance between quality and time consumption (only 19% of the manual annotation effort).
Detecting machine-generated Spanish text remains challenging across domains and generators. Pre-trained models like RoBERTa provide strong contextual embeddings but often underperform on human-authored texts and are sensitive to domain shifts. In this work, we integrate linguistic features from PUCP-Metrix-covering lexical, syntactic, semantic, psycholinguistic, and cohesion properties-with pre-trained models. We evaluate feature-based classifiers, fine-tuned RoBERTa, hybrid models, and ensembles on the AuTexTification dataset. Hybrid models improve human-text detection (F1 65.49 vs. 60.74 for RoBERTa) and machine-text classification (F1 81.76), while a voting ensemble achieves the highest macro-F1 (74.75) and strongest robustness. Analyses indicate linguistic features provide stable, interpretable anchors, reducing overfitting and enhancing generalization across LLM outputs. Results demonstrate that combining linguistic and pre-trained models yields a robust solution for Spanish machine-generated text detection.
Recent multilingual large language models enable richer access to scientific information, but their use in low-and mid-resource settings remains under-explored. We evaluate whether domain-adapted generative and embedding models can improve multilingual scientific information retrieval across Catalan, Spanish, and English. Two tasks are addressed: multilingual query parsing and cross-lingual semantic search through adapted sentence embeddings. Our results show that compact multilingual models, when tuned with domain-specific research data, provide accurate and language-agnostic access to open research information.
The objective of this paper is to propose and evaluate an approach for detecting factual hallucinations in Spanish texts by combining linguistic analysis with fact verification from structured knowledge sources-in our case, Wikidata. The results demonstrate that the proposed approach can identify contradictions and fabricated facts with high precision (around 90%), providing a pioneering and competitive tool compared to similar frameworks for factual verification in Spanish.
The proliferation of fake news and conspiracy narratives has become a significant challenge within the Natural Language Processing (NLP) research field, since there is an increasing circulation of misinformation across digital platforms. The spread of these false narratives poses a potential threat for the preservation and promotion of the information trustworthiness, especially in online media communication. However, most relevant NLP models and datasets are still focused on the English language, leaving the Spanish language relatively underrepresented. This survey paper addresses the linguistic and computational challenges of disinformation and conspiracy detection in Spanish-language contexts. We review current and relevant NLP methods and techniques, from traditional machine learning methods to Large Language Models (LLMs) and transformer-based architectures, presenting the most effective experiments in identifying disinformation and conspiracy discourses in Spanish. Furthermore, we highlight the challenges of the multimodal approaches in NLP to include the analysis of visual and audio data in online media and suggest the need for culturally informed annotation schemes and domain-specific corpora. In this regard, we review the research advances that are contributing to the development of more precise models, capable of capturing the subtle and subjective rhetorical and ideological features that characterize misinformation in the Spanish-speaking digital sphere.
This paper is a summary of the PhD dissertation Lexical borrowing detection as a sequence labeling task. Data, modeling and evaluation methods for anglicism retrieval in Spanish by Elena Alvarez Mellado, conducted under the supervision of Julio Gonzalo (UNED) and Constantine Lignos (Brandeis University) within the PhD program in Intelligent Systems at UNED School of Computer Science. The defense took place in Madrid on May 27th, 2025. The PhD committee was formed by Iria da Cunha (UNED), Javier de la Rosa (National Library of Norway) and Mariona Taule (University of Barcelona). The thesis was awarded the cum laude distinction and the international mention.
Automatic summarization of political news requires preserving factual content, journalistic style, and editorial orientation. However, research in Spanish remains limited due to the lack of ideologically diverse resources and outlet-sensitive evaluations. We present Spanish PoliSUM-2025, a corpus of 94,832 Spanish political news articles from multiple outlets with different editorial viewpoints and their original summaries. Using this resource, we compare mBART with three instruction-tuned large language models. Results show that instruction-tuned models consistently outperform the encoder-decoder baseline and that model size is the main determinant of quality. Linguistic analysis indicates that LLMs preserve semantic content but diverge from the concise, information-dense journalistic style. Substantial outlet-level variability appears in ROUGE-L but not in BERTScore, while ideological orientation produces no detectable differences under these metrics.
This work presents a comparative analysis of labeling in sexism detection using ambiguous Spanish-language data selected with the Think Twice method. A subset of examples with content related to sexism was annotated by Mexican women from diverse sociocultural backgrounds and contrasted with labels produced by an LLM. The results show low agreement both among humans and between humans and the LLM, reflecting the interpretative variability inherent in subjective tasks. Despite this variability, the model tends to approximate the average human judgment. These findings highlight the need for annotation schemes and classification approaches that account for cultural and linguistic diversity rather than forcing a single correct interpretation in sensitive tasks such as sexism detection.
This paper presents the results from the 2024 ODESIA Challenge, a public competition aimed at benchmarking natural language processing (NLP) systems in Spanish across ten discriminative tasks using a standardized methodology based on private, held-out test sets. Results show the winning system (Qwen2.5-14B) prevailed due to structural advantages in extractive Question Answering, whereas encoders outperformed LLMs in other tasks such as sequence labeling and soft classification. We conclude that, while generative models may dominate reasoning-heavy tasks involving long contexts, encoder architectures obtain on-par or even better performance in many other discriminative scenarios, challenging the assumption that massive scale universally supersedes specialized architectural design.
Despite the importance of performance of large language models (LLMs) in multilingual and low-resource settings, most work on stance detection has focused on supervised text classification in English. This paper revisits the task through in-context learning (ICL) method, introducing a novel method called modified Chain-of-Stance (mCoS) and assessing whether open-source LLMs can serve as competitive, zero-training alternatives to supervised systems. This work evaluates four instruction-tuned LLMs on Spanish and Catalan tweets from the Catalonia Independence Corpus (CIC). It also evaluates them on Basque (Euskera) data from VaxxStance and English tweets from SemEval-2016. The study compares five prompting strategies: zero-shot, few-shot, chain-of-thought, modified Chain-of-Stance, and the COLA multi-agent framework, and the results are compared against strong supervised baselines. The paper concludes that mCoS achieves state-of-the-art performance in all languages evaluated for generative models, but remains below fine-tuned and statistical baselines in the multilingual settings of Spanish, Catalan and Basque. Additional analysis shows that annotation noise and language-linked biases for the CIC dataset drive many errors, highlighting persistent challenges in multilingual stance detection and the need for higher-quality and fairness-oriented evaluation resources.
Speech has driven the development of Automatic Speech Recognition (ASR) models like Whisper, capable of converting spoken language into written text. However, these models require large amounts of data (corpora), which limits their performance in languages or variants with scarce resources, such as the Colombian Spanish, whose accents and regionalisms are underrepresented. Likewise, the use of recordings often includes sensitive information, such as names or IDs, which makes the collection and sharing of these corpora difficult. This work proposes the development of a model based on the Whisper arquitecture and WhisperX's pipeline, for anonymized speech transcription in Colombian Spanish, with temporal annotation and speaker diarization. With models achieving a 7,60 % transcription word error rate (WER), an F1-score of 60,81 % for named entity recognition, and an F1-score of 76,10% for anonymization, it contributes to closing the gap between existing models and Colombian dialects, ensuring robust performance even in low-resource settings.
Trustworthy intent detection is limited by deep learning opacity. While dynamical systems theory has emerged as a powerful tool for interpreting Recurrent Neural Networks (RNNs), its application has been unexplored in high-intent, large scale scenarios common to real-world products. We extend this analytical framework to benchmarks with up to 150 intents. We find RNNs trained on these tasks still converge to an interpretable geometric solution, forming robust, intentspecific clusters in their hidden space. We show this space's intrinsic dimensionality grows sub-linearly with task complexity. Building on this, we introduce Functional Dimensionality (FD), a novel, task-aware metric that quantifies the minimum dimensionality required to preserve this semantic structure. Our analysis reveals FD is remarkably low, suggesting RNNs solve complex tasks via an efficient, highly organized subspace. We show this subspace is structured for inference, with clusters aligning strongly with their corresponding readout vectors. These findings offer a scalable framework for auditing and interpreting high-intent dialogue systems.
This study computationally models metaphorical utterances based on a manually annotated corpus extracted from Platero y yo by Juan Ram & oacute;n Jim & eacute;nez. The metaphorical formalization begins with a directed acyclic graph (DAG), whose nodes and edges are classified according to a lexical, syntactic-logical-semantic, and conceptual typology. Following the analysis of topological metrics and community detection, a depth-first search (DFS) algorithm is employed as a segmentation operator on the network, from which the graphs corresponding to the corpus utterances are derived. Transition curves are then constructed between cognitive node classes, sequentially ordered according to the perceptual linearity of the discourse. Structural, semantic, and cognitive-perceptual-sequential features, discretely encoded, are integrated with dense contextual embeddings as input to a graph neural network (GNN), employed as a representational encoder, according to the Graph Networks theoretical framework. The evaluation conducted using unsupervised clustering techniques, semantic cohesion metrics, and an ablation study indicates that the proposed model generates greater cohesion and separability in the analyzed vector representation space than exclusively statistical-distributional models. Furthermore, it highlights the relevance of the relational categorization employed. This finding opens new lines of research in the computational modeling and processing of metaphor.
This paper presents the final outcomes of the TAN-IBE project, whose central objective has been to develop Neural Machine Translation (NMT) systems specifically adapted for the Romance languages of the Iberian Peninsula, with a particular focus on the most under-resourced among them: Asturian, Aragonese, and Aranese. We present the developed corpora and the final neural machine translation systems trained.
PhD thesis in Computer Science focussed on irony detection in social media, written by Reynier Ortega Bueno under the supervision of Prof. Paolo Rosso, at the Universitat Polite`cnica de Valencia. This thesis investigates irony detection as a multifaceted linguistic, computational, and social challenge, addressing multilingual variation, multimodality, and corpus bias. The work introduces an attentive LSTM architecture integrating linguistic and deep features for Spanish irony and satire detection, and proposes an end-to-end model combining textual and visual transformers for multimo dal irony detection in social media content. This work further analyses topic bias in irony corpora, demonstrating its detrimental impact on model generalisation and showing gains achieved through bias identification and mitigation. The defense took place in Valencia, Spain, on July 25(th), 2025. The doctoral committee was composed by Rafael Berlanga Llavori, (Universitat Jaume I), Els Lefever, (Ghent University), and Tony Veale, (University College Dublin). The thesis received an international mention, an excellent qualification, and the distinction of Cum Laude.
