alighabusaleh/SARF-MARBERTv2-Arabic-Sentiment

SARF (صرف): Sentiment Analysis via Root-based Fusion for Multi-Dialectal Arabic

SARF is a 3-class (negative / neutral / positive) sentiment classifier for dialectal Arabic. It reads each sentence in three morphological views: the normalized surface form, a light stem, and the consonantal root. All three go through one shared MARBERTv2 encoder and are fused with cross-attention.

This is the exact model behind the TTLab submission to the AraSentEval 2026 shared task (OSACT7 @ LREC 2026). It reached a macro-F1 of 0.9263 on the official test set and placed 2nd of 15 teams.

📄 Paper: TTLab at AraSentEval: SARF (صرف) Sentiment Analysis via Root-based Fusion for Multi-Dialectal Arabic 💻 Code: github.com/aliabusaleh/ArabicSentimentAnalysis_AraSentEval_2026

Model details

Developed by	Ali Abusaleh, Bhuvanesh Verma, Alexander Mehler (Text Technology Lab, Goethe University Frankfurt)
Model type	Multi-view BERT + cross-attention + CNN/LSTM classifier
Base model	UBC-NLP/MARBERTv2
Language	Arabic: Egyptian, Saudi, Jordanian and Moroccan (Darija) dialects, plus MSA
Domain	Hotel / tourism reviews
Labels	0: negative, 1: neutral, 2: positive
Parameters	~167.5M

Architecture

            ┌─ surface ─┐      ┌─ stem ─┐       ┌─ root ─┐
 sentence ─▶│ normalize │ ───▶ │ Snowball│ ───▶ │Tashaphyne│
            └─────┬─────┘      └───┬────┘       └───┬────┘
                  ▼                ▼                ▼
              MARBERTv2 ════ shared weights ═══ MARBERTv2 (×3 passes)
                  │ H_s            │ H_st           │ H_r
                  │      ┌─────────┴────────────────┘
                  │      ▼  shared 8-head cross-attention (Q = H_s)
                  │   A_st = Attn(H_s, H_st),  A_r = Attn(H_s, H_r)
                  │      │
                  │   F = (H_s + A_st + A_r) / 3
                  ▼      ▼
      CNN (k=3,4,5 × 200) on H_s    LSTM (128) on F → mean-pool
                  └──────── concat (728) ────────┘
                              ▼
                     Dropout(0.3) → Linear → 3 classes

• Surface view: URLs and digits removed, diacritics and tatweel stripped, repeated characters collapsed, and Alef/Hamza variants normalized (أ إ آ → ا, ؤ → و, ئ → ي).
• Stem view: Snowball Arabic stemmer (PyStemmer).
• Root view: Tashaphyne ArabicLightStemmer.get_root().

How to use

The model ships custom code, so pass trust_remote_code=True. The morphological preprocessing needs three small extra packages:

pip install torch transformers PyStemmer tashaphyne pyarabic

from transformers import AutoModel, AutoTokenizer

repo = "alighabusaleh/SARF-MARBERTv2-Arabic-Sentiment"
tokenizer = AutoTokenizer.from_pretrained(repo, trust_remote_code=True)
model = AutoModel.from_pretrained(repo, trust_remote_code=True).eval()

texts = [
    "الفندق كان رائع والخدمة ممتازة",      # positive
    "الغرفة كانت وسخة وما عجبتني",          # negative
]
print(model.predict(texts, tokenizer))
# ['positive', 'negative']

print(model.predict(texts[0], tokenizer, return_probs=True))
# [{'label': 'positive', 'scores': {'negative': 7e-05, 'neutral': 7e-05, 'positive': 0.9999}}]

For custom batching, model.encode(texts, tokenizer) returns the six input tensors: input_ids / attention_mask for the surface, stem and root views. Pass them straight to model(**batch):

import torch
batch = model.encode(texts, tokenizer)
with torch.no_grad():
    logits = model(**batch).logits
labels = [model.config.id2label[i] for i in logits.argmax(-1).tolist()]

⚠️ Always pad to 128 tokens. The LSTM branch mean-pools over all 128 positions, including padding, exactly as in training. encode() already pads to max_length=128. If you tokenize yourself with dynamic padding, the predictions will differ from the reported results.

The released weights were checked against the original training checkpoint. They are bit-identical, and they reproduce the leaderboard submission exactly (312/312 test predictions match).

Training

Training ran in two stages, with the same hyperparameters (found by a W&B sweep):

Hyperparameter	Value
Optimizer	AdamW, weight decay 0.02
Learning rate	1.24e-4 (stage 1), 1.24e-5 (stage 2)
Batch size	128
Max sequence length	128
CNN filters / kernels	200 / {3, 4, 5}
LSTM hidden size	128 (1 layer, unidirectional)
Dropout	0.3
Loss	Cross-entropy

1. Stage 1: transfer from MSA hotel reviews. 10 epochs on the Arabic hotel reviews from SemEval-2016 Task 5 ABSA (srinivasbilla/semeval-2016-absa-reviews-arabic). Sentence-level labels, de-duplicated by sentence.
2. Stage 2: dialectal adaptation. 10 epochs at a 10× lower learning rate on the AraSentEval 2026 training set: 1,731 sentences, balanced across Darija, Saudi, Jordanian and Egyptian (657 negative / 609 positive / 465 neutral).

Evaluation

Split	Macro-F1
AraSentEval 2026 official test set	0.9263 (2nd / 15 teams)

See the paper for the full results and analysis.

Limitations

• Trained only on hotel and tourism reviews. Expect lower accuracy on other domains such as politics, social media or product reviews.
• Covers four dialects: Egyptian, Saudi, Jordanian and Moroccan. Other dialects (e.g. Iraqi, Sudanese, Gulf varieties beyond Saudi) and Arabizi (Latin-script Arabic) are not represented in the training data.
• Neutral is the hardest class. Most errors are neutral sentences predicted as positive or negative.
• The root and stem extractors are rule-based and designed for MSA. They are noisy on dialectal words; the model learns to weight them through cross-attention.
• Every input runs the encoder three times, so inference costs about 3× a single MARBERTv2 forward pass.

Citation

If you use this model, please cite our paper:

@inproceedings{abusaleh-etal-2026-ttlab,
    title = "{TTL}ab at {A}ra{S}ent{E}val: {SARF}( صرف) Sentiment Analysis via Root-based Fusion for Multi-Dialectal {A}rabic",
    author = "Abusaleh, Ali  and
      Verma, Bhuvanesh  and
      Mehler, Alexander",
    editor = "Al-Khalifa, Hend  and
      El-Haj, Mo  and
      Ezzini, Saad",
    booktitle = "The 7th Workshop on Open-Source {A}rabic Corpora and Processing Tools ({OSACT}7) with 5 Shared Tasks",
    month = may,
    year = "2026",
    address = "Palma, Mallorca (Spain)",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2026.osact-1.35/",
    doi = "10.63317/4wj6s3ys5osk",
    pages = "262--268"
}

Please also cite MARBERT (Abdul-Mageed et al., ACL 2021), which SARF builds on.