📄 Ahasis Shared Task: Hybrid Lexicon-Augmented AraBERT Model for Sentiment Detection in Arabic Dialects

Summary

The paper describes a system for the Ahasis@RANLP-2025 shared task, which asks for three-way sentiment classification (positive, neutral, negative) of hotel reviews written in Saudi Arabic and Darija (Moroccan Arabic). Dialectal Arabic has no standard spelling, varies a lot by region, and has few resources. The task data is also small, with only 860 reviews.

The authors respond in two ways. First, they enlarge the training data with light token-level perturbations and with paraphrases generated by AraT5. Second, they derive a sentiment lexicon directly from the labelled training reviews. Lexicon-based features are projected through a small dense layer and joined with AraBERT contextual embeddings, and a multi-layer classification head makes the final decision.

On the test set, the hybrid system reached an F1 of 0.74, compared with 0.56 for an AraBERT-only baseline. The paper does not report an ablation that separates the effect of augmentation from the effect of the lexicon. Even so, the result suggests that explicit lexical signals combined with augmentation can help a pretrained encoder in small, dialect-heavy settings.

Key points

• Data: the Ahasis dataset has 860 hotel reviews, split evenly with 430 Saudi and 430 Darija, each labelled positive, neutral or negative. Stratified k-fold cross-validation keeps class proportions stable across splits.
• Augmentation: two methods were applied to disjoint, class-balanced subsets of the training data.
  • 20% received word deletion, swapping and noise injection.
  • 50% were paraphrased with AraT5 (the almarjeh t5-arabic-text-summarization checkpoint), with tuned beam width, repetition penalty and length limits.
• Preprocessing: ArabertPreprocessor removes Tashkeel and Tatweel, normalizes Alef variants, replaces URLs, mentions and emails with placeholder tokens, and cleans spacing and punctuation.
• Lexicon construction:
  • The stopword list was extended with dialectal stopwords.
  • Domain-frequent but neutral nouns such as "hotel", "room" and "restaurant" were dropped.
  • Only tokens seen at least 10 times were kept; thresholds of 5, 10 and 15 were tried, and 10 worked best.
  • Each token gets per-class proportions normalized to sum to one.
• Features and result: the lexical features are counts and ratios of positive and negative words, cumulative scores, distribution statistics and marker presence. These are fused with AraBERT by concatenation, giving F1 0.74 against 0.56 for the baseline.

Interesting ideas

• I1. A lexicon can be built from labelled data by recording, for each sufficiently frequent word, its normalized share across classes; the result is a transparent probability-like entry that can be inspected and tuned through a single frequency cutoff.
• I2. Hand-made, interpretable features can be injected into a neural model by passing them through a small projection layer and concatenating them with the contextual embeddings before the classifier, so explicit knowledge and learned representations sit side by side.
• I3. Augmentation intensity can be matched to how safely each method preserves meaning: noisy edits go to a small share of the data and meaning-preserving paraphrases to a larger share, on disjoint and class-stratified subsets.
• I4. Words that are frequent only because of the domain (topic nouns) are deliberately removed from a polarity resource, so attitude cues are not confused with subject matter.

For Mova

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

• I1 (world): evaluative words (good/bad, kind/cruel) could be scored against labelled school or fable material, so every value judgment in the world of the text traces back to visible counts.
• I2 (tree): gparse could receive compact lexical features, such as verb subcategorization hints from VerbNet classes, concatenated into its arc scorer as an explainable signal alongside transformer states.
• I3 (learn): when growing silver data, LLM-teacher paraphrases could be scaled up while token-level corruptions stay rare, and keeping the subsets disjoint lets each source be evaluated separately against negative controls.
• I4 (frames): when detecting speaker attitude or irony in prag, topical nouns that merely dominate a genre could be discounted, so stance is read from evaluative predicates and modifiers instead.