Persian sentiment analysis (polarity) for short user-generated text, trained on the SnappFood review corpus. A hashed-bigram linear textcat with no vectors and no transformer: 85.92 macro F on the held-out test split at 7,868 docs/second on an i5-7200U (2 cores). Fine-tuned ParsBERT scores 87.31 on the same split at 229 docs/second on a Tesla T4, so this model gives up 1.4 F for a 34x throughput gain and no GPU requirement. Labels: HAPPY, SAD.
| Feature |
Description |
| Name |
fa_persian_sentiment_snappfood_bow |
| Version |
0.1.0 |
| spaCy |
>=3.8.14,<3.9.0 |
| Default Pipeline |
textcat |
| Components |
textcat |
| Vectors |
0 keys, 0 unique vectors (0 dimensions) |
| Sources |
SnappFood sentiment corpus (Farahani et al. (ParsBERT), via ParsiAI) |
| License |
MIT |
| Author |
Fazel |
Label Scheme
View label scheme (2 labels for 1 components)
| Component |
Labels |
textcat |
HAPPY, SAD |
Accuracy
| Type |
Score |
CATS_SCORE |
86.08 |
CATS_MICRO_P |
86.12 |
CATS_MICRO_R |
86.12 |
CATS_MICRO_F |
86.12 |
CATS_MACRO_P |
86.61 |
CATS_MACRO_R |
86.16 |
CATS_MACRO_F |
86.08 |
CATS_MACRO_AUC |
92.64 |
TEXTCAT_LOSS |
4219.38 |
Trained from configs/textcat_bow.cfg. No vectors are bundled, so the wheel is ~2 MB and inference is pure CPU. Tokenisation is spaCy's blank Persian tokenizer with no normalisation, which is deliberate: hashed bigrams absorb ZWNJ and elongation variants as distinct features rather than collapsing them.
