KEMALGUNAY/STANCE_DETECTION_ANIMAL_LAW_DISCUSSION

Turkish Stance Detection Model for Street Animal Law Discussions

This model is a Turkish text classification model fine-tuned to detect stance in discussions about the street animal law debate in Turkey. It classifies texts into two categories:

• SUPPORT: the text supports the law/policy
• OPPOSE: the text opposes the law/policy

The model is intended for research on online public discourse, political communication, digital debate, and issue polarization in Turkish-language discussions.

Model description

This is a binary stance detection model trained on Turkish texts related to public debate around the street animal law in Turkey. The model aims to identify whether a given text expresses support for or opposition to the law.

This model should be understood as a stance classification system, not as a fact-checking, sentiment analysis, or legal interpretation tool. It predicts the likely stance expressed in a text based on patterns learned from the training data.

Labels

• 0 = OPPOSE
• 1 = SUPPORT

Intended uses

This model can be used for:

• stance detection in Turkish social media or online discussions
• research on public debate and polarization
• issue-specific discourse analysis
• large-scale analysis of support/opposition patterns in user-generated texts

Out-of-scope uses

This model should not be used for:

• legal decision-making
• law enforcement or surveillance targeting individuals
• automated moderation without human review
• determining the moral worth, intent, or ideological identity of a person
• high-stakes decisions affecting individuals or groups

Training data

The model was trained on a custom Turkish-language annotated dataset focused on discussions about the street animal law debate in Turkey.

• Training set size: 1,084 texts
• Test set size: 271 texts

The dataset is issue-specific and domain-specific. Performance may decrease on texts outside this topic, outside Turkish, or outside the discourse style represented in the training data.

Training procedure

The model was fine-tuned for 5 epochs on the custom annotated dataset.

Evaluation by epoch

Epoch 1

• Training loss: 0.5709
• Accuracy: 0.7970

Epoch 2

• Training loss: 0.2978
• Accuracy: 0.8339

Epoch 3

• Training loss: 0.1634
• Accuracy: 0.8413

Epoch 4

• Training loss: 0.0629
• Accuracy: 0.8524

Epoch 5

• Training loss: 0.0407
• Accuracy: 0.8450

Test performance

Best observed test performance was achieved at Epoch 4:

• Accuracy: 0.8524
• Macro F1: 0.85
• Weighted F1: 0.85

Classification report at best epoch:

              precision    recall  f1-score   support

      OPPOSE       0.86      0.85      0.85       136
     SUPPORT       0.85      0.86      0.85       135

    accuracy                           0.85       271
   macro avg       0.85      0.85      0.85       271
weighted avg       0.85      0.85      0.85       271