Aspect Based Sentiment Analysis with Turkish 🇹🇷 Data

This model performs Aspect-Based Sentiment Analysis (ABSA) 🚀 for Turkish text. It predicts sentiment polarity (Positive, Neutral, Negative) towards specific aspects within a given sentence.

This model is fine-tuned from the dbmdz/bert-base-turkish-cased pretrained BERT model. It is trained on the Turkish-ABSA-Wsynthetic dataset, which contains Turkish restaurant reviews annotated with aspect-based sentiments. The model is capable of identifying the sentiment polarity for specific aspects (e.g., "servis," "fiyatlar") mentioned in Turkish sentences.

• Developed by: Sengil
• Language(s): Turkish 🇹🇷
• License: Apache-2.0
• Finetuned from model: dbmdz/bert-base-turkish-cased
• Number of Labels: 3 (Negative, Neutral, Positive)

• Notebook: ABSA_Turkish_BERT_Based_Small

This model can be used directly for analyzing aspect-specific sentiment in Turkish text, especially in domains like restaurant reviews.

It can be fine-tuned for similar tasks in different domains (e.g., e-commerce, hotel reviews, or customer feedback analysis).

• Not suitable for tasks unrelated to sentiment analysis or Turkish language.
• May not perform well on datasets with significantly different domain-specific vocabulary.

• May struggle with rare or ambiguous aspects not covered in the training data.
• May exhibit biases present in the training dataset.

!pip install -U transformers

Use the code below to get started with the model:

from transformers import AutoTokenizer, AutoModelForSequenceClassification

# Load the model and tokenizer
tokenizer = AutoTokenizer.from_pretrained("Sengil/ABSA-Turkish-bert-based-small")
model = AutoModelForSequenceClassification.from_pretrained("Sengil/ABSA-Turkish-bert-based-small")

# Example inference
text = "Servis çok yavaştı ama yemekler lezzetliydi."
aspect = "servis"
formatted_text = f"[CLS] {text} [SEP] {aspect} [SEP]"

inputs = tokenizer(formatted_text, return_tensors="pt", padding="max_length", truncation=True, max_length=128)
outputs = model(**inputs)
predicted_class = outputs.logits.argmax(dim=1).item()

# Map prediction to label
labels = {0: "Negative", 1: "Neutral", 2: "Positive"}
print(f"Sentiment for '{aspect}': {labels[predicted_class]}")

Training Data The model was fine-tuned on the Turkish-ABSA-Wsynthetic.csv dataset. The dataset contains semi-synthetic Turkish sentences annotated for aspect-based sentiment analysis.

• Training Procedure
• Optimizer: AdamW
• Learning Rate: 2e-5
• Batch Size: 16
• Epochs: 5
• Max Sequence Length: 128

The model achieved the following scores on the test set:

• Accuracy: 95.48%
• F1 Score (Weighted): 95.46%

@misc{absa_turkish_bert_based_small,
  title={Aspect-Based Sentiment Analysis for Turkish},
  author={Sengil},
  year={2024},
  url={https://huggingface.co/Sengil/ABSA_Turkish_BERT_Based_Small}
}

For any questions or issues, please open an issue in the repository or contact LinkedIN.