kubi565/emotion-berturk-turkish

Model Card for Model ID

Emotion BERTurk is a fine-tuned language model developed specifically for emotion classification in Turkish text. It is based on the BERTurk architecture, which itself is a BERT model trained on a large corpus of Turkish data. It is based on a survey with 5000 participants asked to write stories about their strongest ex perienced memories for the basic six emotions (Happiness,Fear,Anger,Sadness,Disgust,Surprise). The model can be used as-is for emotion classification tasks in Turkish or further fine-tuned for domain-specific emo tion recognition needs. It provides a reliable foundation for sentiment-aware applications such as chatbots, mental health analysis, and user feedback categorization, specifically tailored to the linguistic structure and emotional expression of Turkish speakers.

Model Details

This model is fine-tuned with the TREMO dataset for 6 different emotion classification in Turkish. 93% of accuracy is achieved on the validation set.

Model Description

It has 2 main outputs: the label (indicates the emotion) and polarity score.

The output and its corresponding emotion:

LABEL_0: Happy LABEL_1: Fear LABEL_2: Anger LABEL_3: Sadness LABEL_4: Disgust LABEL_5: Surprised

Example usage:

model = AutoModelForSequenceClassification.from_pretrained("kubi565/emotion-berturk-turkish")
tokenizer = AutoTokenizer.from_pretrained("kubi565/emotion-berturk-turkish")
sentiment_analyzer = pipeline("text-classification", model=model, tokenizer=tokenizer)

result = sentiment_analyzer(text, truncation=True)[0] 
label = result['label'] 
score = result['score']

• Developed by: [Kubilay Kürtür]
• Model type: [Text Classification]
• Language(s) (NLP): [Turkish]
• License: [More Information Needed]
• Finetuned from model [BERT]: [savasy/bert-base-turkish-sentiment-cased]

Training Details

The Emotion BERTurk model was fine-tuned using the TREMO dataset, which includes six basic emotions: happiness, sadness, anger, fear, surprise, and disgust. The dataset was split into 70% training and 30% validation. Fine-tuning was done with the Hugging Face Transformers library, and the model achieved 93% accuracy on the validation set.

Training Data

Dataset link: https://www.kaggle.com/datasets/mansuralp/tremo