Estimating the Arabic Level of Dialectness (ALDi) is a continuous regression task that quantifies how much a sentence diverges from Modern Standard Arabic (MSA). While auxiliary classification is often used to support regression, its effectiveness depends critically on how the auxiliary task is designed. We conduct a systematic study of auxiliary classification for ALDi estimation, showing that naive auxiliary supervision can degrade regression performance under skewed target distributions. We identify two key factors that determine when auxiliary tasks become beneficial: the granularity of discretization and the choice of loss function. Specifically, 7-bin discretization combined with focal loss substantially improves performance. Our approach achieves an RMSE of 0.1275 on the NADI 2024 blind test set, outperforming previous systems by over 9%, and the same design transfers to English toxicity regression. Detailed error analysis reveals how auxiliary classification reshapes model behavior across the Arabic dialect spectrum. We release our code and model predictions to facilitate future research

We describe our system for Track 1 (dialectgeneralization) of the ArA-DF-2026 Arabic speech deepfake detection shared task. Starting from the organizers’ XLS-R+AASIST baseline (14.65% EER on the development test set), we show that (i) raw-waveform augmentation (RawBoost) roughly halves the error rate, (ii) replacing the AASIST back-end with a simple mean-pooling head over a posttrained self-supervised frontend removes a hard performance ceiling that persists even after full convergence, and (iii) a large, architecturally diverse ensemble is the only lever that keeps improving once single-model performance saturates. Our final system, a 13-member weighted ensemble spanning three frontend sizes, two back-ends, and four training-data recipes, reaches 2.34% EER on the development-test leaderboard and 2.42% EER on the official held-out blind evaluation set (144,210 utterances, 9× larger than the development test). We also report two negative results of practical value: injecting a large (54k-clip) multi-system Arabic text-to-speech corpus and a vocoder-copy-synthesis pseudospoof corpus both degrade accuracy, indicating that indiscriminately adding synthetic speech is harmful when its generation artifacts do not match the target domain. Our code is publicly available at https://github.com/Body123/aradf-2026-saar-df.

We describe our submission to the AraSeg 2026 shared task on Arabic sentence segmentation (Paragraph-Aware, Closed track). We cast segmentation as binary token-boundary classification and build a system around three ideas: (i) task-adaptive pretraining (TAPT) of encoders on the task’s own training text, (ii) regularized fine-tuning with R-Drop and boundary smoothing, and (iii) a hand-designed cross-family ensemble of architecturally diverse Arabic encoders. Our final system reaches a boundary F1 of 94.8 on the blind test set (ranking 3rd of 12 teams, 0.5 points from the top-ranked team). Beyond the score, we report a practical finding backed by a series of blind submissions: on a small development set, a compact diversity-first ensemble generalizes better to unseen data than ensembles selected by exhaustively maximizing development F1, and TAPT helps only under-adapted backbones while hurting a heavily-pretrained one. We also report five techniques that did not help, to save future practitioners the effort. Our code is publicly available at https://github.com/Body123/araseg-2026-knights-st.

Recognizing the nuanced spectrum of dialectness in Arabic text poses a significant challenge for natural language processing (NLP) tasks. Traditional dialect identification (DI) methods treat the task as binary, overlooking the continuum of dialect variation present in Arabic speech and text. In this paper, we describe our submission to the NADI shared Task of ArabicNLP 2024. We participated in Subtask 2 - ALDi Estimation, which focuses on estimating the Arabic Level of Dialectness (ALDi) for Arabic text, indicating how much it deviates from Modern Standard Arabic (MSA) on a scale from 0 to 1, where 0 means MSA and 1 means high divergence from MSA. We explore diverse training approaches, including contrastive learning, applying a random weighted sampler along with fine-tuning a regression task based on the AraBERT model, after adding a linear and non-linear layer on top of its pooled output. Finally, performing a brute force ensemble strategy increases the performance of our system. Our proposed solution achieved a Root Mean Squared Error (RMSE) of 0.1406, ranking second on the leaderboard.