📄 LinguArabic at AraSentEval 2026: MARBERT for Multi-Dialect Arabic Sentiment Analysis

Summary

LinguArabic's entry to the sentiment classification track (Subtask 1) of AraSentEval 2026 labels dialectal Arabic sentences as positive, negative or neutral. Dialects are hard for this task because they have no standard spelling and vary widely in vocabulary and morphology. The training data is the MDS-3 corpus, built from hotel reviews rendered into four regional varieties and labelled by native speakers. The team fine-tuned MARBERTv2, a BERT-base encoder pretrained on a very large collection of Arabic tweets, with a linear softmax classifier on the [CLS] vector. Around the model they added three layers of processing:

• generic text cleaning,
• hand-written rules that map dialect words to standard equivalents,
• an inference rule that sends low-confidence predictions to the neutral class.

This enhanced setup scored about 0.83 Macro F1 against roughly 0.80 for a basic-preprocessing baseline. The official test score was 0.8333. Most remaining errors were confusions between neutral and positive, sarcasm, and unfamiliar dialect words. The work suggests that dialect-oriented pretraining combined with cheap normalization is a practical recipe for low-resource varieties.

Key points

• Data: MDS-3 has 2,043 training sentences in Moroccan, Egyptian, Jordanian and Saudi varieties, labelled with three sentiment classes. The test set has 312 reviews from the same dialect groups.
• Model and training: UBC-NLP/MARBERTv2 (12 layers, 12 heads, hidden size 768, pretrained on over 1B tweets).
  • Optimizer: AdamW, learning rate 3e-5, batch size 16, FP16.
  • Early stopping with patience 2 on validation loss.
• Two configurations:
  • System A: basic cleaning, maximum length 192 tokens.
  • System B: dialect normalization rules, maximum length 256 tokens, and confidence filtering.
  • The three changes in System B are evaluated together, not one at a time.
• Confidence filtering: if the top softmax probability is below τ, the label becomes neutral. τ = 0.7 was chosen on validation data from the range 0.5–0.9.
• Results: System B reached 0.8333 Macro F1, with accuracy, precision and recall also at 0.8333. System A reached 0.80.

Interesting ideas

• I1. Hand-written per-dialect mappings send regional variants of common words (for example, words for "this", "good" or "very") to one standard form, which shrinks vocabulary sparsity before a learned model sees the text.
• I2. Collapsing stretched letters and unifying letter variants is a cheap, transparent cleaning step that removes noise from user-generated spelling.
• I3. When the classifier is unsure, it does not guess; it falls back to the neutral class, using a cutoff tuned on held-out data.
• I4. The error analysis isolates a telling sarcasm pattern: strong praise followed by a conditional clause that describes a clearly unpleasant experience.
• I5. The analysis also shows that mildly positive words such as "acceptable" inside a hedged sentence push the model toward the wrong class.

For Mova

How these ideas could serve Mova, our small open language model — its files, stages and tests:

• I1 (tokens): A similar table could map informal English forms (e.g., "gonna", "y'all", "ur") to canonical lemmas while keeping the original surface form attached, so tagging and lemmatization become more stable without losing information.
• I2 (tokens): Normalizing elongations such as "soooo" or "goooood" into a base form plus an "emphasis" feature would keep the intensity signal available as an explainable cue for later stages.
• I3 (frames): Speech-act or irony decisions could abstain to a default "literal/neutral" reading below a calibrated confidence, and the explanation trace could record why the system declined to commit.
• I4 (frames): A rule could fire irony when a positive evaluative predicate is conditioned on enjoying a negative state (e.g., long waiting), using frame roles rather than word polarity.
• I5 (learn): Such sentences could form a negative-control set for sentiment and evaluation readers, checking that weak positive words are not treated as full praise.