Julio Gonzalo

Julián Cendrero, Julio Gonzalo, Ivar Zapata

The Topics over Time (ToT) model captures thematic changes in timestamped datasets by explicitly modeling publication dates jointly with word co-occurrence patterns. However, ToT was not approached in a fully Bayesian fashion, a flaw that makes it susceptible to stability problems. To address this issue, we propose a fully Bayesian Topics over Time (BToT) model via the introduction of a conjugate prior to the Beta distribution. This prior acts as a regularization that prevents the online version of the algorithm from unstable updates when a topic is poorly represented in a mini-batch. The characteristics of this prior to the Beta distribution are studied here for the first time. Still, this model suffers from a difference in scale between the single-time observations and the multiplicity of words per document. A variation of BToT, Weighted Bayesian Topics over Time (WBToT), is proposed as a solution. In WBToT, publication dates are repeated a certain number of times per document, which balances the relative influence of words and timestamps along the inference process. We have tested our models on two datasets: a collection of over 200 years of US state-of-the-union (SOTU) addresses and a large-scale COVID-19 Twitter corpus of 10 million tweets. The results show that WBToT captures events better than Latent Dirichlet Allocation and other SOTA topic models like BERTopic: the median absolute deviation of the topic presence over time is reduced by $51\%$ and $34\%$, respectively. Our experiments also demonstrate the superior coherence of WBToT over BToT, which highlights the importance of balancing the time and word modalities. Finally, we illustrate the stability of the online optimization algorithm in WBToT, which allows the application of WBToT to problems that are intractable for standard ToT.

Eva Sánchez Salido, Roser Morante, Julio Gonzalo et al.

In this article we present UNED-ACCESS 2024, a bilingual dataset that consists of 1003 multiple-choice questions of university entrance level exams in Spanish and English. Questions are originally formulated in Spanish and translated manually into English, and have not ever been publicly released. A selection of current open-source and proprietary models are evaluated in a uniform zero-shot experimental setting both on the UNED-ACCESS 2024 dataset and on an equivalent subset of MMLU questions. Results show that (i) reasoning questions are challenging for models, (ii) smaller models perform worse than larger models and degrade faster in Spanish than in English and (iii) the performance gap between languages is negligible for the best models and grows up to 37% for smaller models. Model ranking on UNED-ACCESS 2024 is almost identical in English and Spanish, and has also a high correlation (0.98 Pearson) with ranking on MMLU, suggesting that a small dataset is sufficiently diverse and representative to measure performance by discipline.

Guillermo Marco, Julio Gonzalo, Ramón del Castillo et al.

It has become routine to report research results where Large Language Models (LLMs) outperform average humans in a wide range of language-related tasks, and creative text writing is no exception. It seems natural, then, to raise the bid: Are LLMs ready to compete in creative writing skills with a top (rather than average) novelist? To provide an initial answer for this question, we have carried out a contest between Patricio Pron (an awarded novelist, considered one of the best of his generation) and GPT-4 (one of the top performing LLMs), in the spirit of AI-human duels such as DeepBlue vs Kasparov and AlphaGo vs Lee Sidol. We asked Pron and GPT-4 to provide thirty titles each, and then to write short stories for both their titles and their opponent's. Then, we prepared an evaluation rubric inspired by Boden's definition of creativity, and we collected 5,400 manual assessments provided by literature critics and scholars. The results of our experimentation indicate that LLMs are still far from challenging a top human creative writer, and that reaching such level of autonomous creative writing skills probably cannot be reached simply with larger language models.

Guillermo Marco, Luz Rello, Julio Gonzalo

In this paper, we evaluate the creative fiction writing abilities of a fine-tuned small language model (SLM), BART-large, and compare its performance to human writers and two large language models (LLMs): GPT-3.5 and GPT-4o. Our evaluation consists of two experiments: (i) a human study in which 68 participants rated short stories from humans and the SLM on grammaticality, relevance, creativity, and attractiveness, and (ii) a qualitative linguistic analysis examining the textual characteristics of stories produced by each model. In the first experiment, BART-large outscored average human writers overall (2.11 vs. 1.85), a 14% relative improvement, though the slight human advantage in creativity was not statistically significant. In the second experiment, qualitative analysis showed that while GPT-4o demonstrated near-perfect coherence and used less cliche phrases, it tended to produce more predictable language, with only 3% of its synopses featuring surprising associations (compared to 15% for BART). These findings highlight how model size and fine-tuning influence the balance between creativity, fluency, and coherence in creative writing tasks, and demonstrate that smaller models can, in certain contexts, rival both humans and larger models.

Eva Sánchez Salido, Julio Gonzalo, Guillermo Marco

In LLM evaluations, reasoning is often distinguished from recall/memorization by performing numerical variations to math-oriented questions. Here we introduce a general variation method for multiple-choice questions that completely dissociates the correct answer from previously seen tokens or concepts, requiring LLMs to understand and reason (rather than memorizing) in order to answer correctly. Using this method, we evaluate state-of-the-art proprietary and open-source LLMs on two datasets available in English and Spanish: the public MMLU benchmark and the private UNED-Access 2024 dataset. Results show that all models experience remarkable accuracy drops under our proposed variation, with an average loss of 57% on MMLU and 50% on UNED-Access 2024, ranging from 10% to 93% across models. Notably, the most accurate model in our experimentation (OpenAI-o3-mini) is not the most robust (DeepSeek-R1-70B), suggesting that the best models in standard evaluations may not be the ones with better reasoning capabilities. Also, we see larger accuracy drops in public (vs private) datasets and questions posed in their original language (vs a manual translation), which are signs of contamination and also point to a relevant role of recall/memorization in current LLMs' answers.

Elena Alvarez-Mellado, Julio Gonzalo

Standard evaluation in NLP typically indicates that system A is better on average than system B, but it provides little info on how to improve performance and, what is worse, it should not come as a surprise if B ends up being better than A on outside data. We propose an evaluation methodology for sequence labeling tasks grounded on error analysis that provides both quantitative and qualitative information on where systems must be improved and predicts how models will perform on a different distribution. The key is to create test sets that, contrary to common practice, do not rely on gathering large amounts of real-world in-distribution scraped data, but consists in handcrafting a small set of linguistically motivated examples that exhaustively cover the range of span attributes (such as shape, length, casing, sentence position, etc.) a system may encounter in the wild. We demonstrate this methodology on a benchmark for anglicism identification in Spanish. Our methodology provides results that are diagnostic (because they help identify systematic weaknesses in performance), actionable (because they can inform which model is better suited for a given scenario) and predictive: our method predicts model performance on external datasets with a median correlation of 0.85.

Elena Alvarez-Mellado, Jordi Porta-Zamorano, Constantine Lignos et al.

This paper summarizes the main findings of ADoBo 2025, the shared task on anglicism identification in Spanish proposed in the context of IberLEF 2025. Participants of ADoBo 2025 were asked to detect English lexical borrowings (or anglicisms) from a collection of Spanish journalistic texts. Five teams submitted their solutions for the test phase. Proposed systems included LLMs, deep learning models, Transformer-based models and rule-based systems. The results range from F1 scores of 0.17 to 0.99, which showcases the variability in performance different systems can have for this task.

Andrés Fernández García, Javier de la Rosa, Julio Gonzalo et al.

The ability to summarize long documents succinctly is increasingly important in daily life due to information overload, yet there is a notable lack of such summaries for Spanish documents in general, and in the legal domain in particular. In this work, we present BOE-XSUM, a curated dataset comprising 3,648 concise, plain-language summaries of documents sourced from Spain's ``Boletín Oficial del Estado'' (BOE), the State Official Gazette. Each entry in the dataset includes a short summary, the original text, and its document type label. We evaluate the performance of medium-sized large language models (LLMs) fine-tuned on BOE-XSUM, comparing them to general-purpose generative models in a zero-shot setting. Results show that fine-tuned models significantly outperform their non-specialized counterparts. Notably, the best-performing model -- BERTIN GPT-J 6B (32-bit precision) -- achieves a 24\% performance gain over the top zero-shot model, DeepSeek-R1 (accuracies of 41.6\% vs.\ 33.5\%).

Guillermo Marco, Julio Gonzalo, Víctor Fresno

Recent studies comparing AI-generated and human-authored literary texts have produced conflicting results: some suggest AI already surpasses human quality, while others argue it still falls short. We start from the hypothesis that such divergences can be largely explained by genuine differences in how readers interpret and value literature, rather than by an intrinsic quality of the texts evaluated. Using five public datasets (1,471 stories, 101 annotators including critics, students, and lay readers), we (i) extract 17 reference-less textual features (e.g., coherence, emotional variance, average sentence length...); (ii) model individual reader preferences, deriving feature importance vectors that reflect their textual priorities; and (iii) analyze these vectors in a shared "preference space". Reader vectors cluster into two profiles: 'surface-focused readers' (mainly non-experts), who prioritize readability and textual richness; and 'holistic readers' (mainly experts), who value thematic development, rhetorical variety, and sentiment dynamics. Our results quantitatively explain how measurements of literary quality are a function of how text features align with each reader's preferences. These findings advocate for reader-sensitive evaluation frameworks in the field of creative text generation.

Guillermo Marco Remón, Julio Gonzalo

In recent years, several systems of automated metric analysis of Spanish poetry have emerged. These systems rely on complex methods of syllabification and stress assignment, which use PoS-tagging libraries, whose computational cost is high. This cost increases with the calculation of metric ambiguities. Furthermore, they do not consider determining issues in syllabic count such as the phenomena of compensation between hemistichs of verses of more than eleven syllables. However, it is possible to carry out an informative and accurate metric analysis without using these costly methods. We propose an algorithm that performs accurate scansion (number of syllables, stress pattern and type of verse) without syllabification. It addresses metric ambiguities and takes into account the hemistichs compensation. Our algorithm outperforms the current state of the art by 2% in fixed-metre poetry, and 25% in mixed-metre poetry. It also runs 21 and 25 times faster, respectively. Finally, a desktop application is offered as a tool for researchers of Spanish poetry.

Enrique Amigó, Julio Gonzalo, Stefano Mizzaro et al.

In Ordinal Classification tasks, items have to be assigned to classes that have a relative ordering, such as positive, neutral, negative in sentiment analysis. Remarkably, the most popular evaluation metrics for ordinal classification tasks either ignore relevant information (for instance, precision/recall on each of the classes ignores their relative ordering) or assume additional information (for instance, Mean Average Error assumes absolute distances between classes). In this paper we propose a new metric for Ordinal Classification, Closeness Evaluation Measure, that is rooted on Measurement Theory and Information Theory. Our theoretical analysis and experimental results over both synthetic data and data from NLP shared tasks indicate that the proposed metric captures quality aspects from different traditional tasks simultaneously. In addition, it generalizes some popular classification (nominal scale) and error minimization (interval scale) metrics, depending on the measurement scale in which it is instantiated.

Enrique Amigó, Julio Gonzalo, Javier Artiles et al.

Many Artificial Intelligence tasks cannot be evaluated with a single quality criterion and some sort of weighted combination is needed to provide system rankings. A problem of weighted combination measures is that slight changes in the relative weights may produce substantial changes in the system rankings. This paper introduces the Unanimous Improvement Ratio (UIR), a measure that complements standard metric combination criteria (such as van Rijsbergen's F-measure) and indicates how robust the measured differences are to changes in the relative weights of the individual metrics. UIR is meant to elucidate whether a perceived difference between two systems is an artifact of how individual metrics are weighted. Besides discussing the theoretical foundations of UIR, this paper presents empirical results that confirm the validity and usefulness of the metric for the Text Clustering problem, where there is a tradeoff between precision and recall based metrics and results are particularly sensitive to the weighting scheme used to combine them. Remarkably, our experiments show that UIR can be used as a predictor of how well differences between systems measured on a given test bed will also hold in a different test bed.

Arkaitz Zubiaga, Damiano Spina, Enrique Amigó et al.

This paper explores the real-time summarization of scheduled events such as soccer games from torrential flows of Twitter streams. We propose and evaluate an approach that substantially shrinks the stream of tweets in real-time, and consists of two steps: (i) sub-event detection, which determines if something new has occurred, and (ii) tweet selection, which picks a representative tweet to describe each sub-event. We compare the summaries generated in three languages for all the soccer games in "Copa America 2011" to reference live reports offered by Yahoo! Sports journalists. We show that simple text analysis methods which do not involve external knowledge lead to summaries that cover 84% of the sub-events on average, and 100% of key types of sub-events (such as goals in soccer). Our approach should be straightforwardly applicable to other kinds of scheduled events such as other sports, award ceremonies, keynote talks, TV shows, etc.