University of Tripoli at AraSentEval: Fine-Tuning MARBERTv2 and CAMELBERT for Multi-Dialect Arabic Sentiment Analysis
The 7th Workshop on Open-Source Arabic Corpora and Processing Tools (OSACT7) with 5 Shared Tasks
Abstract
This paper presents our contribution to the AraSentEval 2026 shared task, specifically for Subtask 1: Arabic Dialect Sentiment Analysis, hosted at the OSACT7 workshop during LREC 2026. The task focuses on classifying the sentiment (positive, negative, neutral) of text written in four major Arabic dialects: Moroccan, Egyptian, Jordanian, and Saudi. We addressed this by fine-tuning several pre-trained language models, including MARBERTv2 and CAMELBERT, on the provided Multi-Dialect-Sent (MDS-3) dataset. Our best-performing system MARBERTv2, achieved a Macro F1-score of 84.29% on the official test set, securing fourth place among 13 participating teams. Our findings underscore the value of leveraging large pre-trained models tailored to dialectal Arabic for improved sentiment classification in this under-resourced domain.