The rise of toxic content on digital platforms has intensified the demand for automatic moderation tools. While English has benefited from large-scale annotated corpora, Spanish remains under-resourced, particularly for nuanced cases of toxicity such as irony, sarcasm, or indirect aggression. We present an extended version of the NECOS-TOX corpus, comprising 4,011 Spanish comments collected from 16 major news outlets. Each comment is annotated across three levels of toxicity (Non-Toxic, Slightly Toxic, and Toxic), following an iterative annotation protocol that achieved substantial inter-annotator agreement (k = 0.74). To reduce annotation costs while maintaining quality, we employed a human-in-the-loop active learning strategy, with manual correction of model pre-labels. We benchmarked the dataset with traditional machine learning (ML) methods, domain-specific transformers, and instruction-tuned large language models (LLMs). Results show that compact encoder models (e.g., RoBERTa-base-bne, 125M parameters) perform on par with much larger models (e.g., LLaMA-3.1-8B), underscoring the value of in-domain adaptation over raw scale. Our error analysis highlights persistent challenges in distinguishing subtle forms of toxicity, especially sarcasm and implicit insults, and reveals entity-related biases that motivate anonymization strategies. The dataset and trained models are released publicly.

Question Answering (QA) over Tabular Data has been traditionally a challenging task, but LLMs have recently shown the ability to respond to questions related to this type of structured data. However, current tabular QA datasets are skewed toward Wikipedia tables and SQL-style answers composed of human-crafted question–answer pairs. This limits the evaluation of LLMs on this task to a narrow genre of data and language, while also requiring extensive human effort for dataset or benchmark creation. To address this, we introduce SynTabQA, a methodology for the automatic generation of synthetic question–answer pairs from any unannotated table. SynTabQA defines a detailed question typology, enabling fine-grained evaluation and facilitating the creation of diverse QA datasets. Our approach not only provides an automated test bed for any tabular dataset but can also be used in few-shot settings to supply LLMs with tailored examples, improving their focus and accuracy. We validate SynTabQA on two large, manually constructed tabular QA benchmarks of distinct nature.

Question Answering on Tabular Data (or Table Question Answering) has seen tremendous advances with the coming of new generation Large Language Models (LLMs). Despite this, significant challenges still remain to be solved if we are to develop robust enough approaches for general usage. One of these is ambiguity in question answering, which historically has not merited much attention due to the previously limited capabilities of LLMs. In this work, we outlay the main types of ambiguousness inherent to tabular data. Then, we discuss how they are influenced by the way our models interact with the information stored in the tables, and we test the capabilities of some LLMs in detecting them. This work provides an initial ground for a deeper discussion on how to approach ambiguity in Tabular Data in the age of LLMs.

We introduce the findings and results of SemEval-2025 Task 8: Question Answering over Tabular Data. We featured two subtasks, DataBench and DataBench Lite. DataBench consists on question answering over tabular data, and DataBench Lite small comprising small datasets that might be easier to manage by current models by for example fitting them into a prompt. The task was open for any approach, but their answer has to conform to a required typing format. In this paper we present the task, analyze a number of system submissions and discuss the results. The results show how approaches leveraging LLMs dominated the task, with larger models exhibiting a considerably superior performance compared to small models.

Language models exhibit inherent security vulnerabilities, which may be related to several factors, among them the malicious alteration of the input data. Such weaknesses compromise the robustness of language models, which is more critical when adversarial attacks are stealthy and do not require high computational resources. In this work, we study how vulnerable English language models are to adversarial attacks based on subtle modifications of the input of pretrained English language models. We claim that the attack may be more effective if it is targeted to the most salient words for the discriminative task of the language models. Accordingly, we propose a new attack built upon a two-step approach: first, we use a posteriori explainability methods to identify the most influential words for the classification task, and second, we replace them with contextual synonyms retrieved by a small language model. Since the attack has to be as stealthy as possible, we also propose a new evaluation measure that combines the effectiveness of the attack with the number of modifications performed. The results show that pretrained English language models are vulnerable to minimal semantic changes, which makes the design of countermeasure methods imperative.

Language models are susceptible to vulnerability through adversarial attacks, using manipulations of the input data to disrupt their performance. Accordingly, it represents a cibersecurity leak. Data manipulations are intended to be unidentifiable by the learning model and by humans, small changes can disturb the final label of a classification task. Hence, we propose a novel attack built upon explainability methods to identify the salient lexical units to alter in order to flip the classification label. We asses our proposal on a disinformation dataset, and we show that our attack reaches high balance among stealthiness and efficiency.

This paper describes the participation of the SINAI research group at WASSA 2022 (Empathy and Personality Detection and Emotion Classification). Specifically, we participate in Track 1 (Empathy and Distress predictions) and Track 2 (Emotion classification). We conducted extensive experiments developing different machine learning solutions in line with the state of the art in Natural Language Processing. For Track 1, a Transformer multi-output regression model is proposed. For Track 2, we aim to explore recent techniques based on Zero-Shot Learning models including a Natural Language Inference model and GPT-3, using them in an ensemble manner with a fine-tune RoBERTa model. Our team ranked 2nd in the first track and 3rd in the second track.