This paper describes our participation in the Rest-Mex 2023 Sentiment Analysis Task. We proposed an ensemble of (i) a cascade of transformer-based two-class classifiers biased to lowering the Mean Average Error in Polarity, and (ii) multi-class transformer-based classifiers for the prediction of the Type and Location of the messages. Our system achieved a sentiment track score of 0.719.

Generative Artificial Intelligence has grown exponentially as a result of Large Language Models (LLMs). This has been possible because of the impressive performance of deep learning methods created within the field of Natural Language Processing (NLP) and its subfield Natural Language Generation (NLG), which is the focus of this paper. Within the growing LLM family are the popular GPT-4, Bard and more specifically, tools such as ChatGPT have become a benchmark for other LLMs when solving most of the tasks involved in NLG research. This scenario poses new questions about the next steps for NLG and how the field can adapt and evolve to deal with new challenges in the era of LLMs. To address this, the present paper conducts a review of a representative sample of surveys recently published in NLG. By doing so, we aim to provide the scientific community with a research roadmap to identify which NLG aspects are still not suitably addressed by LLMs, as well as suggest future lines of research that should be addressed going forward.

We present a data-driven approach to discover and extract patterns in textual genres with the aim of identifying whether there is an interesting variation of linguistic features among different narrative genres depending on their respective communicative purposes. We want to achieve this goal by performing a multilevel discourse analysis according to (1) the type of feature studied (shallow, syntactic, semantic, and discourse-related); (2) the texts at a document level; and (3) the textual genres of news, reviews, and children’s tales. To accomplish this, several corpora from the three textual genres were gathered from different sources to ensure a heterogeneous representation, paying attention to the presence and frequency of a series of features extracted with computational tools. This deep analysis aims at obtaining more detailed knowledge of the different linguistic phenomena that directly shape each of the genres included in the study, therefore showing the particularities that make them be considered as individual genres but also comprise them inside the narrative typology. The findings suggest that this type of multilevel linguistic analysis could be of great help for areas of research within natural language processing such as computational narratology, as they allow a better understanding of the fundamental features that define each genre and its communicative purpose. Likewise, this approach could also boost the creation of more consistent automatic story generation tools in areas of language generation.

Automatic Text Simplification (ATS) has emerged as a key area of research within the field of Natural Language Processing, aiming to improve access to information by reducing the linguistic complexity of texts. Simplification can be applied at various levels—lexical, syntactic, semantic, and stylistic—and must be tailored to meet the needs of different target audiences, such as individuals with cognitive disabilities, low-literacy readers, or non-native speakers. This work introduces a tool that automatically adapts Spanish texts into Easy-to-Read format, enhancing comprehension for people with cognitive or reading difficulties. The proposal addresses the need for accessible, well-documented solutions aligned with official guidelines, reinforcing the potential of text simplification as a strategy for inclusion.

Large language models have shown impressive performance in Natural Language Processing tasks, but their black box characteristics render the explainability of the model’s decision difficult to achieve and the integration of semantic knowledge. There has been a growing interest in combining external knowledge sources with language models to address these drawbacks. This paper, OntoLM, proposes a novel architecture combining an ontology with a pre-trained language model to classify biomedical entities in text. This approach involves constructing and processing graphs from ontologies and then using a graph neural network to contextualize each entity. Next, the language model and the graph neural network output are combined into a final classifier. Results show that OntoLM improves the classification of entities in medical texts using a set of categories obtained from the Unified Medical Language System. We can create more traceable natural language processing architectures using ontology graphs and graph neural networks.

Simple.Text es una herramienta de Procesamiento de Lenguaje Natural (PLN) que adapta automáticamente textos en español a formato de Lectura Fácil, facilitando la comprensión para personas con dificultades cognitivas o de lectura. Esta herramienta, fruto del proyecto de investigación ClearText, abarca transformaciones léxicas, ortotipográficas, sintácticas y semánticas. La interfaz de Simple.Text presenta dos cajas: en la primera se introduce el texto original y en la segunda se visualiza el texto transformado, ya sea resumido o simplificado. Dos botones permiten seleccionar entre resumen o simplificación/adaptación E2R (Easy-to-Read), resaltándose en color más intenso la opción activa. Además, el texto también puede ser adjuntado en lugar de copiado directamente. El texto transformado y cualquier glosario generado se muestran en una caja adicional, permitiendo su descarga en formato .txt. La API es responsive, es decir, se adapta a dispositivos móviles, tabletas y portátiles, y acepta diversos formatos de entrada como PDF, TXT y Word. Implementada en Python 3 bajo una arquitectura cliente-servidor, Simple.Text detecta fenómenos lingüísticos específicos mediante algoritmos avanzados y realiza las transformaciones definidas. Esta es la primera herramienta que adapta textos según las necesidades del usuario, abordando fenómenos lingüísticos y considerando discapacidades cognitivas, mejorando así la accesibilidad y comprensión de textos en español.

Chatbids es un asistente virtual basado en modelos de inteligencia artificial para facilitar la comprensión y acceso simplificado a las licitaciones de contratos públicos, en fuentes de información de la administración pública de la provincia de Alicante. Este sistema integra tecnologías del tipo RAG (Generación Aumentada por Recuperación), las cuales permiten procesar, analizar e indexar documentos de manera optimizada. Así, dichos documentos pueden ser consultados mediante un asistente de diálogo generativo. Chatbids facilita la ingestión continua de contenido a través de un proceso de indexación semántica. La interacción entre la base de datos de licitaciones y el usuario emplea un modelo de lenguaje abierto interno, lo que asegura que todo el flujo de información permanezca seguro dentro del sistema sin recurrir a servicios externos.

AuTexTification is a shared task within the IberLEF workshop which aims to determine whether a text has been generated by an Artificial Intelligence (AI) or a human. The objective of this paper is to report the participation and results of the GPLSI team from the University of Alicante (Spain) in subtask 1: Human or Generated of the AuTexTification challenge for English and Spanish languages. We propose and experiment with different approaches based on Transfer Learning; Ensemble Learning; fine-tuning existing language models, such as RoBERTa or RemBERT; or relying on linguistic features. Our best models for both languages were trained through Transfer Learning techniques, obtaining the 6th and 8th position in the English and Spanish versions of this subtask, respectively. Results obtained in the Spanish-version were close to the top-performing team.

The detection of toxic content in social media has become a critical task in Natural Language Processing (NLP), particularly given its intersection with complex issues like subjectivity, implicit language, and cultural context. Among these challenges, bias in training data remains a central concern—especially as language models risk reproducing and amplifying societal inequalities. This paper investigates the interplay between toxicity and gender bias on Twitter/X by introducing a novel dataset of violent and non-violent Spanish tweets, annotated not only for violence but also for gender. We conduct an exploratory analysis of how biased data can distort toxicity classification and present algorithms to mitigate these effects through dataset balancing and debiasing. Our contributions include four new dataset splits—two balanced and two debiased—that aim to support the development of fairer and more inclusive NLP models. By foregrounding the importance of equity in data curation, this work lays the groundwork for more ethical approaches to automated violence detection and gender annotation.

ClearSim es un corpus paralelo en español compuesto por textos de la administración pública, tanto en su versión original como adaptados a Lectura Fácil y Lectura Facilitada. Este recurso incluye 15,000 textos originales, 10,000 textos simplificados con ChatGPT y revisados por humanos, y más de 4,000 textos adaptados por expertos siguiendo las directrices de Easy-to-Read. Los textos han sido seleccionados de los ayuntamientos más importantes de Alicante, específicamente de las secciones de deporte, cultura y ocio. Su adaptación y validación se han realizado en colaboración con APSA, una ONG especializada en accesibilidad textual. El corpus está alineado a nivel de documento y una parte de sus textos ha sido etiquetada y alineada a nivel de oración, lo que facilita su uso en tareas de aprendizaje automático y estudios lingüísticos. ClearSim es un recurso clave para la investigación en Procesamiento de Lenguaje Natural (PLN), ya que aborda la escasez de corpus en español centrados en accesibilidad. Su desarrollo impulsa el uso de tecnologías de PLN para generar textos accesibles, mejorando la accesibilidad cognitiva en español. Este corpus permitirá el entrenamiento de modelos más precisos y eficientes, favoreciendo una mayor automatización y personalización en la adaptación de textos. Como resultado, se mejorará la calidad de los sistemas de simplificación textual y se facilitará el acceso a la información para personas con dificultades de comprensión. ClearSim se distribuye bajo la licencia CC BY-NC (Reconocimiento - No Comercial), permitiendo su uso en investigación sin fines comerciales y con la posibilidad de futuras ampliaciones o anotaciones adicionales.

Los Parques Tecnológicos representan los ecosistemas de innovación con mayor potencial de cara al impulso de la innovación en el sector productivo, a la competitividad empresarial, a la vinculación y transferencia entre actores y al surgimiento de nuevas iniciativas emprendedoras. Argentina, si bien cuenta con iniciativas con más de dos décadas de andadura, es aún incipiente en términos generales en lo que a Parques Tecnológicos se refiere, lo que significa a su vez una gran oportunidad a nivel país, dado el destacado posicionamiento que tiene Argentina en términos de inversión y resultados en I+D. Este trabajo plantea un estudio heterogéneo de la situación de los Parques Tecnológicos en Argentina, que se ha desarrollado a partir de encuestas, entrevistas e información accesible. Además, se incluye un estudio comparativo con los Parques Tecnológicos españoles, con una trayectoria más consolidada en este campo. El estudio revela un menor número de iniciativas en Argentina, así como una distribución geográfica de carácter más asimétrico, entre otras particularidades propias de cada país. Igualmente, se detectan mayores dificultades para la obtención de datos de los ecosistemas de innovación, particularmente de métricas vinculadas con características endógenas. La limitación de acceso a información imposibilita la aplicación de técnicas que permitan el análisis, predicción y prescripción utilizando, entre otras, herramientas de inteligencia artificial, que sí se han podido aplicar al caso español en estudios previos. Este trabajo sugiere que la disponibilidad de información de calidad facilitaría la toma de decisiones fundamentadas en datos y la orientación de políticas públicas que podrían favorecer el desarrollo de los Parques Tecnológicos, con el consiguiente impacto socioeconómico asociado.

The unstoppable proliferation of news driven by the rise of digital media has intensified the challenge of news verification. Natural Language Processing (NLP) offers solutions, primarily through content and context analysis. Recognizing the vital role of linguistic analysis, this paper presents a multilingual study of linguistic markers for automated deceptive disinformation detection across English, Spanish, and Bulgarian. We compiled datasets in these languages to extract and analyze both general and specific linguistic markers. We then performed feature selection using the SelectKBest algorithm, applying it to various classification models with different combinations of general and specific linguistic markers. The results show that Logistic Regression and Support Vector Machine classification models achieved F1-scores above 0.8 for English and Spanish. For Bulgarian, Random Forest yielded the best results with an F1-score of 0.73. While these markers demonstrate potential for transferability to other languages, results may vary due to inherent linguistic characteristics. This necessitates further experimentation, especially in low-resource languages like Bulgarian. These findings highlight the significant potential of our dataset and linguistic markers for multilingual deceptive news detection.

The purpose of this article is to highlight the critical importance of language generation today. In particular, language generation is explored from the following three aspects: multi-modality, multilinguality, which play crucial role for NLG community. We present the activities conducted within the Multi3Generation COST Action (CA18231), as well as current trends and future perspectives for multitask, multilingual and multimodal language generation.

Recent advancements in Natural Language Processing (NLP) have allowed systems to address complex tasks involving cultural knowledge, multi-step reasoning, and inference. While significant progress has been made in text summarization guided by specific instructions or stylistic cues, the integration of pragmatic aspects like communicative intentions remains underexplored, particularly in non-English languages. This study emphasizes communicative intentions as central to summary generation, classifying Spanish product reviews by intent and using prompt engineering to produce intention-aligned summaries. Results indicate challenges for large language models (LLMs) in processing extensive document clusters, with summarization accuracy heavily dependent on prior model exposure to similar intentions. Common intentions such as complimenting and criticizing are reliably handled, whereas less frequent ones like promising or questioning pose greater difficulties. These findings suggest that integrating communicative intentions into summarization tasks can significantly enhance summary relevance and clarity, thereby improving user experience in product review analysis.

In the age of knowledge, the democratisation of information facilitated through the Internet may not be as pervasive if written language poses challenges to particular sectors of the population. The objective of this paper is to present an overview of research-based automatic text simplification tools. Consequently, we describe aspects such as the language, language phenomena, language levels simplified, approaches, specific target populations these tools are created for (e.g. individuals with cognitive impairment, attention deficit, elderly people, children, language learners), and accessibility and availability considerations. The review of existing studies covering automatic text simplification tools is undergone by searching two databases: Web of Science and Scopus. The eligibility criteria involve text simplification tools with a scientific background in order to ascertain how they operate. This methodology yielded 27 text simplification tools that are further analysed. Some of the main conclusions reached with this review are the lack of resources accessible to the public, the need for customisation to foster the individual’s independence by allowing the user to select what s/he finds challenging to understand while not limiting the user’s capabilities and the need for more simplification tools in languages other than English, to mention a few.

This paper addresses a key limitation in Natural Language Generation (NLG) systems: their struggle with commonsense reasoning, which is essential for generating contextually appropriate and plausible text. The study proposes an approach to enhance the commonsense reasoning abilities of NLG systems by integrating external knowledge framed in a constrained commonsense generation task. The paper investigates strategies for extracting and injecting external knowledge into pre-trained models, specifically BART and T5, in both base and large configurations. Experimental results show that incorporating external knowledge, extracted using a simple strategy, leads to significant performance improvements, with the models achieving 88% accuracy in generating plausible and correct sentences. When refined methods for knowledge extraction are applied, the accuracy further increases to 92%. These findings underscore the crucial role of high-quality external knowledge in enhancing the commonsense reasoning capabilities of NLG systems, suggesting that such integration is vital for advancing their performance in real-world applications.

Since the emergence of Transformers architecture, the Natural Language Generation (NLG) field has advanced at breakneck speed. Large language models (LLMs) have achieved remarkable results in the field of generative artificial intelligence (AI). Nevertheless, they also present some problems worth analysing: not only are they computationally non-viable to academia, but they also have other issues, such us not generating text in a fully controllable way or the phenomenon known as hallucination. Because of this, the purpose of this paper is to outline and set the ideas for a new PhD thesis research. This PhD thesis will aim at advancing the state of the art by discovering new cost-effective, efficient and high-performing approaches to controlled text generation that could perform well in the different NLG tasks. Therefore, the main objective of this PhD thesis is to design a novel and efficient task-agnostic architecture that could obtain equivalent performance of LLMs, while generating text in a controllable way and including external commonsense knowledge.

This paper presents the ongoing work conducted within the ClearText project, specifically focusing on the resource creation for the simplification of Spanish for people with cognitive disabilities. These resources include the CLEARSIM corpus and the Simple.Text tool. On the one hand, a description of the corpus compilation process with the help of APSA is detailed along with information regarding whether these texts are bronze, silver or gold standard simplification versions from the original text. The goal to reach is 18,000 texts in total by the end of the project. On the other hand, we aim to explore Large Language Models (LLMs) in a sequence-to-sequence setup for text simplification at the document level. Therefore, the tool’s objectives, technical aspects, and the preliminary results derived from early experimentation are also presented. The initial results are subject to improvement, given that experimentation is in a very preliminary stage. Despite showcasing flaws inherent to generative models (e.g. hallucinations, repetitive text), we examine the resolutions (or lack thereof) of complex linguistic phenomena that can be learned from the corpus. These issues will be addressed throughout the remainder of this project. The expected positive results from this project that will impact society are three-fold in nature: scientific-technical, social, and economic.

Large Language Models (LLMs) have shown remarkable capabilities, demonstrating their potential to transform the Natural Language Processing (NLP) field by achieving strong performance across the entire spectrum of tasks. In this context, there is growing interest in leveraging LLMs as automated crowdworkers to streamline the traditionally labor-intensive process of manual annotation and linguistic content generation. This paper specifically examines the feasibility of using LLMs to generate high-quality contextual information — an increasingly recognized element for enhancing generative systems — by producing linguistic contexts from given premise sentences. We evaluate four prominent LLMs — LLaMA 2, PaLM 2, Vicuna, and GPT-3.5 — by making them generate 180 contextual outputs, which are then compared to 60 contexts manually crafted by linguistic experts. To systematically assess the appropriateness of these generated contexts, we introduce CATS (Contextual Appropriateness in Texts for Spanish), a novel and adaptable evaluation metric for measuring the appropriateness of contextual information generated texts. CATS is rooted in established discourse theories and provides a robust framework for analyzing linguistic context quality. The implementation of CATS is made publicly available at https://github.com/gplsi/cats. The reliability of CATS is validated through a manual evaluation conducted by two independent human referees. The results indicate that the quality of contexts generated by LLMs is comparable to those produced by human specialists, as evidenced by both CATS scores (0.214 vs. 0.150) and human judgment. Our findings underscore the efficiency and cost-effectiveness of employing LLMs as alternatives to human crowdworkers for generating contextual information, offering promising implications for the scalability and advancement of generative systems.

El modelado de skipgrams es una técnica para la generación de términos multi-palabra que conserva parte de la secuencialidad y flexibilidad del lenguaje. Sin embargo, en algunos casos el número de skipgrams generados puede ser excesivo a medida que se aumenta la distancia entre palabras. Además, esta distancia no suele ser tenida en cuenta a la hora de valorar los términos que se generan. En este trabajo proponemos una técnica para la generación y filtrado eficientes de skipgrams y un esquema de pesado que tiene en cuenta la distancia entre los términos, dando más importancia a aquellos más cercanos. Aplicaremos y evaluaremos estas propuestas en la tarea de análisis de sentimientos.

Annotated corpora are indispensable tools to train computational models in Artificial Intelligence and Natural Language Processing. However, manual annotation is a costly, arduous, and time-consuming task, especially when the annotation is semantically complex. To address the problem, this work applies a methodology for semi-automatic annotation of datasets based on the Human-in-the-Loop paradigm. The methodology supports the building a resource, that benefits from a fine-grained annotation, to aid in the detection of Spanish violent messages sourced from social media (Twitter/X). After implementing the proposed methodology for semi-automatic violence annotation, a high quality resource was obtained (hereafter referred to as VILLANOS). The methodology consists of annotating the dataset incrementally, which delivers an increase in annotator efficiency, thereby validating the suitability of the proposal. Annotation time was reduced by 52% compared to manual annotation and performance, by training a model with the VILLANOS dataset, obtains an F 1 of 85.2%. These results demonstrate the efficiency and effectiveness of the methodology, evidencing its validity.

The rapid evolution of Large Language Models (LLMs) has produced diverse model families with multiple versions differing in parameter size and release chronology. While this ongoing progress has driven substantial advancements in Natural Language Generation (NLG), several critical challenges remain unmet. This paper presents a cross-sectional analysis of 24 model versions from five prominent LLM families, evaluating their performance across five persistent generation gaps. For each gap, we define three targeted tasks to systematically assess model behaviour, with three prompt variants per task to analyse them with comprehensive examples, yielding 45 distinct prompts and 1080 model responses. Human expert qualitative analysis reveals that gpt-4o and gpt-40-mini accurately respond to 75.5% and 46.6% of prompts respectively, while models as llama-3.2-3b-instruct and llama-3.2-1b-instruct achieve only 11% and 10% accuracy respectively. Our study also highlights a key and unexpected insight: in several cases, earlier or smaller LLM versions sometimes match or exceed the performance of their latest, larger counterparts. These findings invite a timely discussion within the NLG community about the current trajectory of scaling models versus focusing on the optimisation and adaptation of existing architectures. We hope this analysis contributes valuable perspectives for more informed and strategic development of future LLMs in the field of NLG.

Millions of people worldwide face barriers in accessing and understanding complex written information due to limited literacy. Automatic text simplification (ATS) addresses this challenge by transforming complex texts into simpler, more accessible versions. However, most existing ATS research focuses on English, leaving Spanish, a language spoken by over 500 million people, underrepresented. This paper fills this gap by introducing large-scale sentence-aligned simplification resources for Spanish, developed from the Newsela and ClearSim corpora. We propose detailed guidelines for manual alignment, evaluate a wide range of automatic sentence alignment algorithms, and present the first systematic exploration of LLM-based monolingual sentence alignment in Spanish. Our analysis incorporates comprehensive quantitative and qualitative evaluation, supported by statistical significance testing, and reveals clear differences in the structural simplification patterns across corpora. In addition, we train and release baseline ATS models using the new aligned datasets, demonstrating their practical utility for downstream simplification. All alignment code, trained models, and evaluation scripts will be publicly released to ensure transparency and reproducibility. Together, these contributions substantially advance the resources and methodology for Spanish-language ATS.

Recent advances in generative models have skyrocketed the popularity of conversational chatbots and have revolutionized the way we interact with artificial intelligence. At the same time, research has shown that machine learning models can unconsciously reflect and amplify human biases. This is particularly dangerous for generative models given the huge popularity of such technologies. Specifically, a fundamental source of bias of such technologies is the resources on which the models are trained. To address this issue, this paper proposes a methodology to analyze intrinsic gender bias in Natural Language Generation (NLG). Some works already propose metrics and approaches to measure bias in the Natural Language processing field. However, there is a lack of standard methodology to measure gender bias in NLG. Therefore, adapting the Bias Score approach, our proposal involves three sequential stages applied to individual texts to detect intrinsic gender bias on NLG effectively. Those steps are as follows: (i) word scoring; (ii) word filtering; and (iii) generative-word analysis. This methodology is applied to recent datasets and pre-trained models widely used for the generation of text with common sense. In particular, this paper analyzes the potential gender bias in CommonGen and C2 Gen datasets and the SimpleNLG and T5 models. The results show the ability of the proposed methodology to detect gender bias in word distributions, presenting a strong correlation with the words typically associated with a specific gender. Results indicate that both tested datasets are intrinsically gender-biased, and therefore, tested models fine-tuned with those datasets also are.

The CLEAR.TEXT project (TED2021-130707B-I00) researches how natural language processing technology can support the authoring of accessible content in Spanish for people with cognitive disabilities. Our main objective is to research, implement, deploy, evaluate, and ultimately provide robust technologies for natural language processing to support the authoring of accessible Spanish content for public sector organisations (at local, regional and national level) that is intelligible to people with cognitive disability, thereby widening their inclusion and empowerment in Europe. It is expected to impact positively the quality of life of people with cognitive disabilities, facilitating their access to educational, vocational, cultural, and social opportunities in public sector organisations.

This paper presents the participation of the GPLSI team in the Quantum CLEF (QClef) Lab at CLEF 2025, focusing on Task 2 - Instance Selection and Task 3 - Clustering. The QClef Lab explores the applicability of quantum and quantum-inspired techniques to core AI tasks, emphasizing optimization efficiency and data reduction. In Task 2, we propose three multi-paradigm approaches for selecting representative training instances for sentiment classification, leveraging sentiment-aware pairing, local set-based criteria, and classical heuristics. In Task 3, we introduce a single quantum-inspired clustering framework that integrates four distinct pivot selection strategies for document grouping in embedding space. Our methods achieved competitive performance across both tasks. In particular, our LocalSets method achieved the highest effectiveness in Task 2 while substantially reducing the training set, and our FPS-Medoids approach delivered the best results for Task 3 in terms of nDCG@10. Overall, our findings support the potential of annealing-based techniques to deliver effective trade-offs between performance and computational efficiency in realistic machine learning pipelines.

Language produced by Public Administrations has crucial implications in citizens’ lives. However, its syntactic complexity and the use of legal jargon, among other factors, make it difficult to be understood for laypeople and certain target audiences. The NLP task of Automatic Text Simplification (ATS) can help to the necessary simplification of this technical language. For that purpose, specialized parallel datasets of complex-simple pairs need to be developed for the training of these ATS systems. In this position paper, an on-going project is presented, whose main objectives are (a) to extensively analyze the syntactical, lexical, and discursive features of the language of English-speaking ombudsmen, as samples of public administrative language, with special attention to those characteristics that pose a threat to comprehension, and (b) to develop the OmbudsCorpus, a parallel corpus of complex-simple supra-sentential fragments from ombudsmen’s case reports that have been manually simplified by professionals and annotated with standardized simplification operations. This research endeavor aims to provide a deeper understanding of the simplification process and to enhance the training of ATS systems specialized in administrative texts.

Numerical information in written texts impacts their readability and is considered complex for people with cognitive disabilities by the Easy-to-Read guidelines. This paper presents Simple.Text, a rule-based system designed to automatically simplify all numerical expressions deemed complex, with a focus on Rules 19-25 from Section 6.2 of the Easy-to-Read guidelines. The results from the evaluation indicate a high precision and accuracy in numerical phenomena detection and transformation, although with some limitations. This system proves to be an efficient and cost-effective tool for the simplification of numerical expressions.

This paper presents CLEARS, a shared task organized within the framework of the evaluation campaign for Natural Language Processing systems in Spanish and other Iberian languages, IberLEF 2025. The main objective of CLEARS is to investigate and develop automatic techniques for adapting Spanish texts to plain language and easy-to-read formats. Two subtasks are proposed: the first focuses on adapting texts to plain language format, while the second aims to produce easy-to-read texts. CLEARS attracted participation from four teams in Subtask 1 and five teams in Subtask 2. Each team submitted their results and approaches, which are presented below.

The increasing use of Artificial Intelligence by non-experts underscores the need to address biases in its outputs, particularly in gendered languages like Spanish, where linguistic features reflect gender. This paper proposes a validated methodology to quantify gender bias in text generated by Large Language Models (LLMs) in Spanish. The approach involves creating gendered seed-word lists, building a Spanish-specific corpus using curated prompts, and analyzing gender polarity and co-occurrences within the generated text. Validated using five state-of-the-art LLMs—GPT-3.5, GPT-4o, Llama 3, Gemini 1.5, and Mixtral8x7b—this study provides a systematic framework for bias detection in Spanish and highlights differences in model performance. By addressing the challenges of bias in gendered languages, this research aims to contribute to the development of equitable AI systems and advances methodologies for bias quantification. The findings highlight gender disparities in LLMs, with masculine and feminine biases shifting depending on the discourse structure. Specifically, the analysis reveals a prevalence of masculine biases in non-leading prompts and feminine biases in leading prompts, underscoring the nuanced ways in which prompt positioning can influence gender representation in model outputs.

XatDocs es un software informático tipo RAG (Retrieval Augmented Generation) que procesa, analiza e indexa documentos de forma optimizada y semántica para posteriormente interrogar dichos documentos a través de una asistente de diálogo generativo. XatDocs permite ingerir contenido continuamente y a través de un proceso de indexación semántico. Respecto a la interacción entre la base de datos y el usuario se utiliza un modelo de lenguaje abierto local, lo cual garantiza que todo el flujo de información se conserve de forma segura dentro del sistema sin utilizar servicios de terceros.

This paper introduces a specialized knowledge discovery framework designed to process health technical documents and extract knowledge. The framework improves existing technology, known as LETO, through the integration of CARMEN, a multilingual entity classification system capable of infusing health-related semantics into the initial versatile approach. This collaborative approach enables the generation of domain-specific knowledge graphs for two languages, Spanish and English. Additionally, this provides a valuable means by which to explore relationships within the health domain that could otherwise remain undiscovered. The resulting technology is subjected to an evaluation procedure using standard metrics employed in knowledge discovery tasks, illustrating how CARMEN contributes to an augmentation in the knowledge discovered in LETO. Thus, the generated knowledge graph can be leveraged for the creation of explanatory representation techniques, facilitating a more comprehensive articulation of human knowledge and potentially serving, among other purposes, as an educational resource.

Recent research has shown that the unaware automation of high-risk decision-making tasks can result in unfair decisions being made. The most common approaches to address this problem adopt definitions of fairness based on protected attributes. Precise annotation of protected attributes enables the application of bias mitigation techniques to commonly unlabeled kinds of data (e.g., images, text, etc.). This paper proposes a framework to automatically annotate protected attributes in data collections. The framework focuses on providing a single interface to annotate protected attributes of different types (e.g., gender, race, etc.) and from different kinds of data. Internally, the framework coordinates multiple sensors to produce the final annotation. Several sensors for textual data are proposed. An optimization search technique is designed to tune the framework to specific domains. Additionally, a small dataset of movie reviews —annotated with gender and sentiment— was created. The evaluation in datasets of texts from diverse domains shows the quality of the annotations and their effectiveness to be used as a proxy to estimate fairness in datasets and machine learning models. The source code is available online for the research community.

The incorporation of machine learning algorithms into high-risk decision-making tasks has raised some alarms in the scientific community. Research shows that machine learning-based technologies can contain biases that cause unfair decisions for certain population groups. The fundamental danger of ignoring this problem is that machine learning methods can not only reflect the biases present in our society but could also amplify them. This article presents the design and validation of a technology to assist the fair automation of classification problems. In essence, the proposal is based on taking advantage of the intermediate solutions generated during the resolution of classification problems through using Auto-ML tools, in particular, AutoGOAL, to create unbiased/fair classifiers. The technology employs a multi-objective optimization search to find the collection of models with the best trade-offs between performance and fairness. To solve the optimization problem, we introduce a combination of Probabilistic Grammatical Evolution Search and NSGA-II. The technology was evaluated using the Adult dataset from the UCI repository, a common benchmark in related research. Results were compared with other published results in scenarios with single and multiple fairness definitions. Our experiments demonstrate the technology's ability to automate classification tasks while incorporating fairness constraints. Additionally, our method achieves competitive results against other bias mitigation techniques. A notable advantage of our approach is its minimal requirement for machine learning expertise, thanks to its Auto-ML foundation. This makes the technology accessible and valuable for advancing fairness in machine learning applications. The source code is available online for the research community.

Automatic text evaluation metrics are widely used to measure the performance of a Natural Language Generation (NLG) system. However, these metrics have several limitations. This article empirically analyses the problem with current evaluation metrics, such as their lack of ability to measure the semantic quality of a text or their high dependence on the texts they are compared against. Additionally, traditional NLG systems are compared against more recent systems based on neural networks. Finally, an experiment with GPT-4 is proposed to determine if it is a reliable source for evaluating the validity of a text. From the results obtained, it can be concluded that with the current automatic metrics, the improvement of neural systems compared to traditional ones is not so significant. On the other hand, if we analyse the qualitative aspects of the texts generated, this improvement is reflected.

The role played by science and technology parks (STPs) in technology transfer, industrial innovation, and economic growth is examined in this paper. The accurate monitoring of their evolution and impact is hindered by the lack of uniformity in STP models or goals, and the scarcity of high-quality datasets. This work uses existing terminologies, definitions, and core features of STPs to conduct a multidimensional data analysis that explores and evaluates the 21 core features which describe the key internal factors of an STP. The core features are gathered from a reliable and updatable dataset of Spanish STPs. The methodological framework can be replicated for other STP contexts and is based on descriptive techniques and machine-learning tools. The results of the study provide an overview of the general situation of STPs in Spain, validate the existence and characteristics of three types of STPs, and identify the typical features of STPs. Moreover, the prototype STP can be used as a benchmark so that other STPs can identify the features that need to be improved. Finally, this work makes it possible to carry out classifications of STPs, in addition to prediction and decision making for innovation ecosystems.

Considering the potential of tools such as ChatGPT or Gemini to generate texts in a similar way to a human would do, having reliable detectors of AI –AI-generated content (AIGC) is vital to combat the misuse and the surrounding negative consequences of those tools. Most research on AIGC detection has focused on the English language, often overlooking other languages that also have tools capable of generating human-like texts, such is the case of the Spanish language. This paper proposes a novel multilingual and multi-task approach for detecting machine vs. human-generated text. The first task classifies whether a text is written by a machine or by a human, which is the research objective of this paper. The second task consists in detect the language of the text. To evaluate the results of our approach, this study has framed the scope of the AuTexTification shared task and also we have collected a different dataset in Spanish. The experiments carried out in Spanish and English show that our approach is very competitive concerning the state of the art, as well as it can generalize better, thus being able to detect an AI-generated text in multiple domains.

Disinformation is an important problem facing society nowadays. Given the rapid and easy access to information, news stories quickly go viral, the vast majority of which are misleading and with no prospect of verification. Specifically, the headline of a correctly designed news item must correspond to a summary of the main information of that news item and it should be neutral. However, many headlines circulating on the Internet use false or distorted information, seeking to confuse or mislead the reader. Misleading headlines indicate a dissonance between the headline and the content of the news story. From a computational perspective, this problem is being tackled as a Stance Detection problem between the headline and the body text of the news item. This paper contributes to the fight against the spread of misleading information by presenting a generic and flexible multi-level hierarchical classification. The approach is based on two stages that enable the detection of the stance between the news headline and the body text. The proposed architecture, called HeadlineStanceChecker+ uses the headline and only the essential information of the news item (not the full body text) as inputs. To extract this essential information, different summarization approaches (extractive and abstractive) are analyzed in order to determine the most relevant information for the task. The experimentation has been carried out using the Fake News Challenge (FNC-1) dataset. A 94.49% accuracy was obtained using extractive summaries, which were more helpful than abstractive ones. HeadlineStanceChecker+ improves the accuracy results of existing state-of-the-art systems. In conclusion, using automatic extractive summaries together with the two-stage generic architecture is an effective solution to the problem.

Los modelos de aprendizaje automático están siendo ampliamente utilizados en múltiples áreas de la vida humana. Tradicionalmente, se han aplicado en reconocimiento de voz, detección de rostros, clasificación de imágenes, sistemas de recomendación, etc. Con la reciente revolución de los modelos generativos, la popularidad de los chatbots conversacionales se ha disparado. Esto ha dado lugar a que los modelos de aprendizaje automático se utilicen cada vez más para abordar tareas para las que no estaban específicamente capacitados. El prompt engineering ha permitido que personas no expertas en aprendizaje automático (que comúnmente tampoco están familiarizadas con los problemas subyacentes al uso de modelos de aprendizaje automático para hacer predicciones) automaticen ciertas tareas. La incorporación de algoritmos de aprendizaje automático en tareas de toma de decisiones de alto riesgo ha levantado algunas alertas en la comunidad científica. Las tareas de toma de decisiones de alto riesgo denotan aquellas tareas que pueden tener un gran impacto en las vidas de las personas sobre quienes se toman las decisiones. Por ejemplo, se han utilizado modelos para decidir si una persona es contratada o no, si se le concede un préstamo, si se acepta una solicitud de cobertura ampliada de seguridad sanitaria y para predecir la probabilidad de reincidencia en un delito. Estudios han demostrado que la automatización inconsciente de este tipo de tareas contiene sesgos, lo cual provoca que decisiones injustas sean tomadas sobre determinados grupos de población. El peligro fundamental de ignorar este problema es que los métodos de aprendizaje automático podrían no sólo reflejar los sesgos presentes en nuestra sociedad, sino que también amplificarlos. Esta tesis presenta el diseño y validación de una tecnología para asistir la automatización justa de problemas de clasificación. En esencia, la propuesta se basa en diseñar una tecnología que saque provecho de las soluciones intermedias generadas durante la resolución de problemas de clasificación mediante el uso de herramientas de Auto-ML, en particular, AutoGOAL, con el propósito de crear clasificadores imparciales y justos. Cuatro componentes fundamentales conforman la propuesta: (I) una componente de Auto-ML, encargada de generar colecciones de modelos con hipótesis diferentes entre sí, cada uno debiendo ser capaz de transformar los datos de entrada a los de salida; (II) una componente de ensemble, responsable de combinar múltiples soluciones para producir una más robusta, exacta y justa, según criterios definidos por funciones de pérdida; (III) una componente de cuantificación de sesgos, encargada de medir la calidad de un modelo según varias definiciones de sesgo y equidad; y (IV) una componente de anotación automática de atributos protegidos, para permitir estimar los sesgos en situaciones donde normalmente no sería posible. Los resultados discutidos en este documento prueban la capacidad de la tecnología propuesta para automatizar tareas de clasificación a la par que se controla su equidad. Además, la experimentación muestra la competitividad de la propuesta respecto a otras alternativas para controlar los sesgos en el modelo de aprendizaje automático y colecciones de datos. La propuesta tiene la ventaja adicional de requerir poco o ningún conocimiento sobre arquitecturas de aprendizaje automático dada las componentes de Auto-ML que incorpora. Todo esto lo convierte en una tecnología valiosa y conveniente de explotar. Más allá de los resultados obtenidos sobre la línea principal de investigación, se obtuvieron algunos resultados secundarios de gran valor. Por un lado, se pudo demostrar que ciertas componentes de la tecnología pueden utilizarse para extender, de forma más robusta, corpus de texto con entidades y relaciones anotadas, asumiendo que se cuente con varias versiones de anotación que se quieran combinar; esta condición suele ser viable dada la existencia de competencias internacionales donde varios participantes compiten por generar la mejor solución. Muy ligado a esto, otro resultado secundario es la generación del eHealth-KD 2019 ensembled corpus a partir de ensamblar las soluciones participantes de la competencia eHealth-KD 2019. Otro resultado secundario se tiene en la componente de anotación automática de atributos protegidos, la cual puede funcionar independiente del flujo end-to-end de automatización justa, y por tanto puede utilizarse como fase de preprocesamiento para otros sistemas de cuantificación o mitigación de sesgos. En ese sentido, se publicó también el corpus Reviews' Gender} diseñado para auxiliar el entrenamiento y evaluación de modelos de anotación de atributos protegidos con el fin de usar sus predicciones como estimadores de equidad. Todos los recursos desarrollados en esta investigación están a disposición pública de la comunidad científica.