A fine-tuned Turkish BERT model for extracting product entities from e-commerce texts: product names, brands, prices, colors, sizes, materials, and gender targets.

This model is part of the Çarşı project, a dual-model NER demo comparing fine-tuned BERT against zero-shot LLM extraction.

Model Overview

Property	Value
Base Model	`dbmdz/bert-base-turkish-cased`
Language	Turkish (tr)
Task	Token Classification / NER (15 BIO labels)
Parameters	~111M
Inference Speed	~10–50 ms on CPU
License	Apache 2.0

Supported Entity Types (7)

Entity	Tag	Description	Example
PRODUCT	B/I-PRODUCT	Product name	koşu ayakkabısı, t-shirt, cep telefonu
BRAND	B/I-BRAND	Brand name	Nike, Samsung, Zara, LC Waikiki
PRICE	B/I-PRICE	Price amount	2.500 TL, 1.299,99 TL
COLOR	B/I-COLOR	Color	kırmızı, lacivert, siyah, bordo
SIZE	B/I-SIZE	Size / shoe number	M, XL, 42, 38 numara
MATERIAL	B/I-MATERIAL	Material / fabric	deri, pamuklu, polyester, kaşmir
GENDER	B/I-GENDER	Target gender	kadın, erkek, unisex

Full BIO label set (15):

O, B-PRODUCT, I-PRODUCT, B-BRAND, I-BRAND, B-PRICE, I-PRICE, B-COLOR, I-COLOR, B-SIZE, I-SIZE, B-MATERIAL, I-MATERIAL, B-GENDER, I-GENDER

Quick Start

With Transformers Pipeline

python

from transformers import pipeline

ner = pipeline(
    "ner",
    model="cihatyldz/carsi-bert-turkish-ecommerce-ner",
    aggregation_strategy="simple",
)

results = ner("Nike Air Max 90 kırmızı 42 numara 3.499 TL")
for entity in results:
    print(f"  [{entity['entity_group']}] {entity['word']} ({entity['score']:.1%})")

# [BRAND] Nike (99.8%)
# [PRODUCT] Air Max 90 (99.2%)
# [COLOR] kırmızı (99.5%)
# [SIZE] 42 numara (98.9%)
# [PRICE] 3.499 TL (99.7%)

Manual Inference

python

from transformers import AutoTokenizer, AutoModelForTokenClassification
import torch

model_name = "cihatyldz/carsi-bert-turkish-ecommerce-ner"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForTokenClassification.from_pretrained(model_name)
model.eval()

text = "Zara kadın deri ceket siyah M beden 2.999 TL"
inputs = tokenizer(text, return_tensors="pt", return_offsets_mapping=True)
offset_mapping = inputs.pop("offset_mapping")[0]

with torch.no_grad():
    outputs = model(**inputs)

predictions = torch.argmax(outputs.logits, dim=-1)[0]

for idx, (pred_id, (start, end)) in enumerate(zip(predictions, offset_mapping)):
    if start == 0 and end == 0:
        continue
    label = model.config.id2label[pred_id.item()]
    if label != "O":
        print(f"  [{label}] {text[start:end]}")

Training Details

Dataset

Source: cihatyldz/carsi-turkish-ecommerce-ner
Size: ~8,000 synthetic Turkish e-commerce sentences
Generation Method: Template-based synthesis with 25 sentence patterns, 39 brands, 50+ products, 25 colors, and randomized parameters
Labeling: Automatic BIO tagging with character-level entity position tracking — no manual annotation needed
Split: 80% train / 10% validation / 10% test

Training Configuration

Hyperparameter	Value
Base model	`dbmdz/bert-base-turkish-cased`
Epochs	5
Batch size	16
Learning rate	3e-5
Weight decay	0.01
Warmup ratio	0.1
Max sequence length	128
FP16	Enabled
Early stopping	Patience = 2 (metric: F1)

Infrastructure

Hardware: NVIDIA Tesla T4 (Google Colab)
Training time: ~10 minutes
Framework: Hugging Face Transformers + Trainer API
Evaluation: seqeval (entity-level metrics)

Evaluation Results

Evaluated on the held-out test split (~800 examples) using seqeval:

Entity	Precision	Recall	F1
PRODUCT	~0.99	~0.99	~0.99
BRAND	~0.99	~0.99	~0.99
PRICE	~0.99	~0.99	~0.99
COLOR	~0.99	~0.99	~0.99
SIZE	~0.99	~0.99	~0.99
MATERIAL	~0.99	~0.99	~0.99
GENDER	~0.99	~0.99	~0.99
micro avg	~0.99	~0.99	~0.99

Note: High scores reflect the synthetic/template-based training data. Real-world e-commerce text (user reviews, messy listings) would yield lower but still useful performance.

Live Demo

Try the model in the Çarşı Space, where it runs side-by-side with a zero-shot LLM (Qwen2.5-7B):

🚀 Çarşı Demo

	BERT NER (this model)	LLM Zero-shot (Qwen2.5-7B)
Speed	~10–50 ms (CPU)	~1–3 s (Inference API)
Approach	Fine-tuned, specialized	Zero-shot, flexible
New entities	Requires retraining	Just add to prompt
Cost	Free (CPU)	Free (HF Inference API)

Use Cases

Product catalog parsing: Extract structured data from unstructured product listings
Search enrichment: Auto-tag products with brand, color, size for better search
Price monitoring: Extract prices from competitor listings
Review analysis: Identify which products, brands, and attributes customers mention
Chatbot integration: Parse user queries like "kırmızı Nike ayakkabı 42 numara" into structured filters

Limitations

Synthetic training data: Trained on template-generated sentences, not real marketplace listings. Performance on noisy real-world text may vary.
Turkish only: Designed for Turkish e-commerce text.
Fixed entity set: Supports 7 predefined entity types. Custom entities require retraining.
Brand coverage: Trained on ~39 major brands. Lesser-known brands may not be recognized.

Project Portfolio

Project	Domain	Architecture	Link
🪙 Akçe	Banking	Fine-tuned LLM (LoRA)	Space
🐪 Kervan	Logistics	RAG	Space
🏥 Şifahane	Healthcare	Classifier vs LLM	Space
⚖️ Mizan	Media	RAG Fact-Checker	Space
🏪 Çarşı	E-Commerce	NER (Token Classification)	Space

Citation

bibtex

@misc{yildiz2025carsi,
  author       = {Cihat Yıldız},
  title        = {Çarşı: Turkish E-Commerce Named Entity Recognition with Fine-tuned BERT},
  year         = {2025},
  publisher    = {Hugging Face},
  url          = {https://huggingface.co/cihatyldz/carsi-bert-turkish-ecommerce-ner}
}

Contact

Developer: Cihat Yıldız
Hugging Face: @cihatyldz
Demo: Çarşı Space