This paper describes our submission to SemEval-2026 Task 9 (Subtask 1) on Spanish online polarisation detection. We investigate whether sentiment-adapted pretrained language models provide an advantage over general-purpose multilingual models for binary polarisation classification. Under a controlled training setup, we compare a base XLM-RoBERTa model, an emotion-adapted model, and a sentiment-adapted XLM-R model trained on Twitter data. To mitigate overfitting in the relatively small training dataset, we additionally apply back-translation as a data augmentation strategy. Experimental results show that the sentiment-adapted checkpoint consistently outperforms the alternative pretrained models under identical conditions. When combined with back-translation augmentation, the final system achieves a macro-averaged F1 score of 0.743 on the preliminary competition leaderboard. These findings suggest that prior adaptation to affective signals in social media can provide beneficial inductive bias for polarisation detection.

The paper presents a taxonomy-guided approach to role classification in news articles using Large Language Models (LLMs). Instead of traditional model training, the system employs zero-shot and few-shot prompting strategies, leveraging structured taxonomies and contextual cues for classification. The study evaluates hierarchical and single-step classification approaches, finding that a unified, single-step model with contextual preprocessing achieves the best performance. The research underscores the importance of input structuring and classification strategy in optimizing LLM performance for real-world applications.

The automatic identification of hate speech constitutes an important task, playing a relevant role towards inclusivity. In these terms, the shared task on Climate Activism Stance and Hate Event Detection at CASE 2024 proposes the analysis of Twitter messages related to climate change activism for three subtasks. Subtasks A and C aim at detecting hate speech and establishing the stance of the tweet, respectively, while subtask B seeks to determine the target of the hate speech. In this paper, we describe our approach to the given subtasks. Our systems leverage transformer-based multi-task learning. Additionally, since the dataset contains a low number of tweets, we have studied the effect of adding external data to increase the learning of the model. With our approach we achieve the fourth position on subtask C on the final leaderboard, with minimal difference from the first position, showcasing the strength of multi-task learning.