📄 AtlasIA at SemEval-2025 Task 11: FastText-Based Emotion Detection in Moroccan Arabic for Low-Resource Settings

Summary

The AtlasIA team tackles multi-label emotion detection for Moroccan Arabic (Darija) in SemEval-2025 Task 11. Darija is a hard case. It has no standard spelling, it varies by region, speakers mix in French, Berber and Spanish, and it lacks emotion lexicons and annotated corpora. The authors use the small Darija ('ary') subset of the task data, loaded through the Hugging Face datasets library. Every text is tagged with at least one of seven categories: joy, sadness, anger, fear, surprise, disgust, or a neutral label. Tweets are first cleaned with a short regex routine that strips mentions, links and hashtags and tidies whitespace and punctuation. A FastText classifier is then trained on the cleaned text, using character n-grams and a one-vs-all loss so that labels are predicted independently. The system reaches 46.44% validation accuracy. The authors present this as a modest baseline showing that cheap, subword-aware models are usable for dialects with few resources. Their error analysis points to code-switching, confusion between similar emotions, and the loss of secondary labels as the main weaknesses.

Key points

• Data: SemEval-2025 Task 11 Moroccan Arabic subset (from the BRIGHTER collection), with 1,608 training, 267 validation and 812 test samples.
• Model: FastText with learning rate 0.40, dimension 8, window 5 and 100 epochs. It uses character n-grams of length 3–6, word n-grams of 1, a One-vs-All (OVA) loss and 7 threads.
• Result: 46.44% validation accuracy, and accuracy is the only metric reported. The authors themselves admit this metric does not fully capture performance.
• Typical errors:
  • sentiment-bearing words in French or Berber are missed;
  • anger/surprise and fear/sadness get confused;
  • in multi-label samples, only the dominant emotion is predicted.
• Future directions: Darija-specific emotion lexicons, hybrids of FastText with AraBERT, cross-lingual transfer, larger datasets and richer metrics. Code is released at github.com/atlasia-ma/semeval-emotion-detection.

Interesting ideas

• I1. Subword pieces (character 3–6-grams) give the model a usable vector even for misspelled, transliterated or never-seen words, which keeps noisy informal text from falling off a cliff.
• I2. A one-vs-all loss treats each label as its own yes/no decision instead of forcing a single winner, which suits phenomena where several labels can hold at once.
• I3. The error analysis found that the model collapses onto the dominant label and drops secondary ones, a failure that plain accuracy hides.
• I4. The authors frame a deliberately cheap, fast model as a floor rather than a final system, and openly note that a single accuracy figure is insufficient.
• I5. Code-switching was a concrete failure source: emotion-bearing words in another language were simply not understood.

For Mova

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

• I1 (tokens): Character n-gram features could serve as a fallback signal in the en tagger and lemmatizer for OOV words, typos and nonce forms.
• I2 (frames): Speech-act and irony/implicature detectors in prag could score each act independently, so an utterance can be both a request and a complaint, with each decision explainable on its own.
• I3 (learn): Evaluation could include a dedicated check on multi-label items, counting recall of second and third labels, so collapse onto the majority reading is caught early.
• I4 (learn): A trivial bag-of-n-grams classifier could serve as a negative-control baseline. Any explainable Mova reader that fails to beat it on a task signals that the structured pipeline is not yet adding value there.
• I5 (tokens): Marking foreign-language spans at the token level (with a language tag in the morphology features) would let later stages handle embedded French or Latin phrases explicitly instead of misparsing them.