Newest papers

Benchmarking Large Language Models for Detecting Satirical Code-Mixed Sentences

Răzvan-Alexandru Smădu, Ioana Buhnila, Daniela Tănase, Dumitru-Clementin Cercel, Florin Pop, Mihaela-Claudia Cercel · 2026 · arXiv (Cornell University)

DOI: 10.48550/arxiv.2610.29907 · cited by 0 works (OpenAlex) · open-access PDF

French

Le changement de code est un phénomène linguistique naturel, mais son impact sur les tâches de raisonnement de haut niveau reste largement inexploré dans les langues sous-ressourcées telles que le roumain. Les tâches de traitement du langage naturel, telles que la détection de satire, reposent fortement sur des indices culturels et syntaxiques nuancés que les grands modèles de langue luttent pour capturer. Dans ce document, nous explorons quatre méthodes (c'est-à-dire la substitution aléatoire, la traduction de phrase, le partage de dépendance et les grands modèles de langue) pour générer des phrases roumaines synthétiques mélangées de code avec l'anglais, l'allemand, le français et l'italien

Machine translation (Meta M2M100); the English original follows.

Abstract

Code-switching is a natural linguistic phenomenon, yet its impact on high-level reasoning tasks remains largely unexplored in under-resourced languages like Romanian. Natural language processing tasks, such as satire detection, rely heavily on nuanced cultural and syntactic cues that large language models struggle to capture. In this paper, we investigate four methods (i.e., random replacement, phrase translation, dependency parsing, and large language models) for generating synthetic code-mixed Romanian sentences with English, German, French, and Italian. We start with an existing dataset for satire detection, compiled from Romanian news articles and annotated at the sentence level by humans, and augment it with over 600k samples, namely CS-SeLeRoSa. We provide an in-depth analysis of the dataset generation process. Then, we evaluate several multilingual large language models across zero-shot, chain-of-thought, and fine-tuning regimes on both the synthetic corpus and a human-adjudicated gold benchmark, namely CS-SeLeRoSa-Gold. While older open models suffer substantial degradation (dropping to 37%-55% macro-F1), newer models like Gemma 4 31B achieve 68.1% zero-shot, rivaling frontier models like GPT-5 mini (69.5%). Furthermore, fine-tuning significantly reduces formatting errors and improves the results, increasing Gemma 4 31B to 77.5% on CS-SeLeRoSa and 72.6% on the gold standard, outperforming frontier zero-shot baselines.

Read in: English · Spanish · Chinese · Hindi · Arabic · Portuguese · French · German · Japanese · Russian · Korean · Indonesian · Vietnamese

Topics: Computer science, Natural language processing, Artificial intelligence, Language model, Sentence, Phrase

Sources: OpenAlex (CC0), arXiv, Crossref. Agent Exchange