Detecting spans of hallucination in LLM-generated answers is crucial for improving factual consistency. This paper presents a span-level hallucination detection framework for the SemEval-2025 Shared Task, focusing on English and Arabic texts. our approach integrates Semantic Role Labeling (SRL) to decompose the answer into atomic roles, which are then compared with a retrieved reference context obtained via question-based LLM prompting. Using a DeBERTa-based textual entailment model, we evaluate each role’s semantic alignment with the retrieved context. The entailment scores are further refined through token-level confidence measures derived from output logits, and the combined scores are used to detect hallucinated spans. Experiments on the Mu-SHROOM dataset demonstrate competitive performance. Additionally, hallucinated spans have been verified through fact-checking by prompting GPT-4 and LLaMA. Our findings contribute to improving hallucination detection in LLM-generated responses.

Hate Speech is an increasingly common occurrence in verbal and textual exchanges on online platforms, where many users, especially those from vulnerable minorities, are in danger of being attacked or harassed via text messages, posts, comments, or articles. Therefore, it is crucial to detect and filter out hate speech in the various forms of text encountered on online and social platforms. In this paper, we present our work on the shared task of detecting hate speech in dialectical Arabic tweets as part of the OSACT shared task on Fine-grained Hate Speech Detection. Normally, tweets have a short length, and hence do not have sufficient context for language models, which in turn makes a classification task challenging. To contribute to sub-task A, we leverage MARBERT’s pre-trained contextual word representations and aim to improve their semantic quality using a cluster-based approach. Our work explores MARBERT’s embedding space and assess its geometric properties in-order to achieve better representations and subsequently better classification performance. We propose to improve the isotropic word representations of MARBERT via clustering. we compare the word representations generated by our approach to MARBERT’s default word representations via feeding each to a bidirectional LSTM to detect offensive and non-offensive tweets. Our results show that enhancing the isotropy of an embedding space can boost performance. Our system scores 81.2% on accuracy and a macro-averaged F1 score of 79.1% on sub-task A’s development set and achieves 76.5% for accuracy and an F1 score of 74.2% on the test set.

Poetry generation tends to be a complicated task given meter and rhyme constraints. Previous work resorted to exhaustive methods in-order to employ poetic elements. In this paper we leave pre-trained models, GPT-J and BERTShared to recognize patterns of meters and rhyme to generate classical Arabic poetry and present our findings and results on how well both models could pick up on these classical Arabic poetic elements.