NUHASHROXME/bangla-fake-news-interpretable

Discourse-Aware Psycholinguistic Bangla Fake News Detector

This model combines BERT embeddings with psycholinguistic and discourse features for interpretable Bangla fake news detection.

Model Details

• Architecture: BERT + 17 psycholinguistic features + 5 discourse features
• Base Model: sagorsarker/bangla-bert-base
• Language: Bengali (Bangla)
• Task: 4-class fake news detection
• Training Data: 42,380 samples from BanFakeNews-2.0 dataset

Performance

• Test F1-Score: 84.37%
• Test Accuracy: 85.73%
• Validation F1: 84.65%
• Classes: 4 categories (0-3)

Key Innovation

This is the first systematic integration of psycholinguistic theory with deep learning for Bangla fake news detection, enabling explainable predictions while maintaining state-of-the-art performance.

Features Extracted

Psycholinguistic Features (17):

• Emotional intensity markers (fear, anger, positive/negative sentiment)
• Uncertainty and hedging patterns
• Cognitive load indicators (repetition, disfluency)
• Deception-specific linguistic patterns (self-reference, present tense usage)

Discourse Features (5):

• Text coherence scores across paragraphs
• Argumentative structure analysis (claims vs evidence)
• Topic progression and transition patterns

Model Architecture

The model integrates:

1. BERT contextual embeddings (768 dimensions)
2. Psycholinguistic features (17 dimensions)
3. Discourse features (5 dimensions)
4. Feature fusion layer
5. Final classification head

Usage

# Note: This model requires custom feature extraction
from transformers import AutoTokenizer

tokenizer = AutoTokenizer.from_pretrained("NUHASHROXME/bangla-fake-news-interpretable")
# Custom model loading code required for interpretable features