Abstract We launched ASBN-MT5, the system for Arabic Sentiment Swap, which performs the task of inverting the sentiment of a sentence while keeping the meaning intact. This is a sequence-to-sequence task. We demonstrate ASBN-MT5: mT5, which is a MultiLingual T5 model, fine-tuned on the provided dataset of the AraSentEval 2026 Shared Task. We describe the data as the first of its kind for the Arabic language, as MAAKS is the first manually composed, parallel, cross-linguistic corpus for the Arabic language. With the preliminary results of Sentiment Flip for the task of Sentiment Inversion, we have recorded a rate of 59.5% for positive to negative conversions and 58.5% for negative to positive conversions, while maintaining an average similarity to the original sentences of 0.955. We present the Arabic prompts and a neuro-developmental (Deep Learning) recipe. Due to the evaluation criteria which include Exact Match, Flip Success, Surface Similarity, and Quality of Output, we restrict the use of Prepended Negation as the main technique and recommend the use of LLMs designed for the Arabic language in the near future. Keywords: mT5, sequence-to-sequence, AraSentEval 2026, Arabic NLP, Text Style Transfer, Sentiment Swap

Cross-topic stance detection in Arabic is the task of identifying whether a text expresses a pro, against, or neutral position toward a given issue, and it is particularly challenging under topic shifts and class imbalance. In Subtask B of the StanceNakba 2026 shared task on Arabic cross-topic stance detection, we are given a Levantine Arabic sentence and one of two topics: “Normalization with Israel” or “Refugee/Immigrant Presence in Jordan,” and we must classify the expressed stance. A central difficulty is the systematic failure of standard fine-tuning to recognize the minority neutral class, driven by majority-class dominance in cross-entropy training and accuracy-based checkpoint selection. To address this, we combine random oversampling with class-weighted cross-entropy loss, and we build an ensemble of four Arabic pre-trained transformers MARBERT, AraBERT Large, XLM-RoBERTa Base, and CAMeL-BERT Mix each trained using Stratified 5-Fold cross-validation. Our final system achieves a macro-F1 of 0.9777 and an accuracy of 97.79% on the evaluation set.

We describe the CodeZone Research Group system for NADI 2026 Shared Task 2, spo ken Arabic dialect identification over Modern Standard Arabic and 19 country-level dialects under cross-domain evaluation. Our system ensembles an ECAPA-TDNN language identi fier fine-tuned on the shared task data with a Whisper-large-v3 dialect classifier that we fur ther adapt using waveform- and spectrogram level augmentation, fusing their log-softmax scores. This ensemble reaches 94.35% accu racy on in-domain validation but initially re tained only 25.40% on the official blind test, a collapse that affected every participating team. We show that about a fifth of this gap is a cali bration failure rather than a modelling failure: discarding per-class bias terms fit in domain raised official accuracy from 0.2540 to 0.3918 and lowered the shared task’s average cost from 0.2077 to 0.1515, and a softening temperature reduced cost further to 0.1469; neither change touched either model. Our submission ranked 10th of 16 teams, and we show analytically why uniform additive score shifts cannot affect this metric while temperature scaling can.

Dialectal Arabic MT error detection asks a sys tem to locate error spans in a machine trans lation and label each with a linguistic error category. It is hard because the dialects lack a standardised orthography and the error cate gories are severely imbalanced. In Subtask 3 of the AlexandriaX 2026 shared task, the in put is an English sentence and its machine translation into one of five Arabic dialects, and the output is character-level error spans with their LQM categories. The central difficulty is collapse onto the majority sociolinguistics class, 58% of training spans, leaving rare cat egories undetected. To address this, we fine tune NileChat-3B-Base with QLoRA on a sin gle 8 GB consumer GPU, casting the task as schema-prompted JSON generation and com bining character-position annotations that let the model copy offsets rather than count them, a category guide with one worked example per error type, and self-consistency decoding merging five samples by character-level union voting. Our system scores 0.3484 Overall on the official test set. We report two negative results: early stopping on validation loss de grades performance, since rare categories are acquired only in later epochs, and fine-tuning on the skewed distribution lowers macro-F1 be low the un-fine-tuned baseline until the cate gory guide is added