yeniguno/nli-deberta-zero-shot-reviews-turkish-v1

Model Card for Model ID

This is a fine-tuned version of MoritzLaurer/deberta-v3-base-zeroshot-v1.1-all-33 on a custom Turkish dataset for zero-shot classification tasks. The model is trained to classify textual inputs as entailment or not_entailment for multiple categories like Memnuniyet (user satisfaction), Şikayet (complaints), and others.

The model leverages the strong multilingual capabilities of DeBERTa-v3-base and is designed for fine-grained understanding of customer feedback and textual sentiment in the Turkish language.

Model Details

Model Description

• Developed by: yeniguno
• Model type: Natural Language Inference (NLI) / Zero-shot classification
• Language(s) (NLP): Turkish
• License: MIT
• Finetuned from model: MoritzLaurer/deberta-v3-base-zeroshot-v1.1-all-33

Uses

• Analyzing customer feedback in Turkish.
• Zero-shot classification of user reviews into predefined categories.

Bias, Risks, and Limitations

The model is trained exclusively on Turkish data. It may not perform well for other languages. Biases from the training data (e.g., overrepresentation of certain categories) may affect performance. High reliance on proper hypothesis construction; ambiguous or irrelevant hypotheses may lead to suboptimal predictions.

How to Get Started with the Model

You can easily use the model for inference with the Hugging Face pipeline functionality. Below is an example to classify the review:

from transformers import pipeline

CANDIDATE_LABELS = [
    "Uygulama Performansı", 
    "Kullanıcı Arayüzü", 
    "Güncellemeler", 
    "Hata ve çökme", 
    "Reklam",
    "Satın Alımlar", 
    "Müşteri Desteği", 
    "Abonelik",	
    "Memnuniyet", 
    "Özellikler",
    "Şikayet"
    ]

# Load the pipeline
classifier = pipeline("zero-shot-classification", model="yeniguno/nli-zero-shot-reviews-MiniLM-turkish-v1")

text = "ChatGPT'nin abonelik ücreti çok yüksek ve müşteri desteği hiç yok, sorun yaşadığınızda yardım alabileceğiniz kimseyi bulamıyorsunuz."

response = pipe(text, CANDIDATE_LABELS, multi_label=True)

for label, score in zip(response["labels"], response["scores"]):
    print(f"{label}: {score:.3f}")

Training Details

Training Data

A Turkish dataset with 951,093 samples, including customer reviews and predefined hypotheses. 0: Neutral (813,408 examples) 1: Entailment (137,685 examples)

Each hypothesis was constructed using the following candidate labels:

Turkish Label	Description
Memnuniyet	User satisfaction
Şikayet	Complaints
Özellikler	Features
Satın Alımlar	Purchases
Kullanıcı Arayüzü	User interface
Uygulama Performansı	App performance
Hata ve Çökme	Bugs and crashes
Müşteri Desteği	Customer support
Abonelik	Subscriptions
Reklam	Advertisements
Güncellemeler	Updates

Training Hyperparameters

• Batch Size: 128
• Learning Rate: 2e-5
• Epochs: 3
• Label Smoothing Factor: 0.1
• Optimizer: AdamW
• Scheduler: Linear with warmup
• Loss Function: Weighted cross-entropy to address class imbalance.
• ...

Evaluation

Results

Step	Training Loss	Validation Loss	Accuracy	Precision	Recall	F1 Score
500	0.313	0.367	0.882	0.974	0.887	0.928
1000	0.291	0.268	0.884	0.981	0.881	0.929
2000	0.244	0.245	0.915	0.981	0.919	0.949
3000	0.220	0.244	0.923	0.982	0.927	0.954
5500	0.190	0.191	0.926	0.987	0.926	0.955

---

Key Observations

1. The model consistently achieved an F1 score of ~95.5%, indicating a strong balance between precision and recall.
2. Validation loss stabilized around 0.19, demonstrating good generalization and minimal overfitting.
3. The model showed robust performance on customer feedback across various categories in Turkish, making it well-suited for zero-shot classification tasks.

---

Evaluation Dataset

• Dataset Size: 951,093 samples
• Split: ~85% training, ~15% validation
• Label Distribution:
  • Entailment: 137,685 examples
  • Not Entailment: 813,408 examples
• Sequence Length: 95th percentile of token lengths used (max_length=116).