Research Paper

Discourse-aware Psycholinguistic Modeling for Bangla Fake News Detection

A comprehensive framework integrating fine-grained linguistic features, discourse analysis, and pre-trained Bangla transformer representations for interpretable fake news detection.

84.37%
F1-Score
22
Features
60K+
Samples
4
Categories

Methodology

Our framework combines pre-trained Bangla BERT representations with 17 psycholinguistic features and 5 discourse-level indicators.

Dataset Distribution
BanFakeNews-2.0 Class Balance

Training: 42,022 samples

Validation: 9,082 samples

Test: 9,082 samples

Class Ratio: 24.8:1 (Authentic:Fake)

Feature Categories
17 Psycholinguistic Features

Emotional Markers: 4 features

Uncertainty Indicators: 3 features

Cognitive Load: 5 features

Deception Patterns: 5 features

Psycholinguistic Feature Extraction

Emotional Markers

  • • Positive sentiment ratio (Excellent, Extraordinary)
  • • Negative sentiment ratio (Terrible, Dangerous)
  • • Fear expressions (Fear, Panic)
  • • Anger indicators (Anger, Rage)

Uncertainty Indicators

  • • Hedging language (Perhaps, Maybe)
  • • Uncertainty expressions (Not certain, Unclear)
  • • Qualification markers (Somewhat, To a great extent)

Cognitive Load Markers

  • • Repetition ratios
  • • Disfluency indicators (That is, Meaning)
  • • Average sentence length
  • • Vocabulary richness (type-token ratio)
  • • Average word length

Deception-Specific Patterns

  • • Self-reference ratios (I, My)
  • • Other-reference patterns (He/She, They)
  • • Present tense usage
  • • Formal vs informal language markers
Model Architecture Diagrams
Multiple perspectives of the integrated framework
Input LayerBangla News Article (Headline + Content)BERT EncoderBangla-BERT Base768-dim embeddingsPsycholinguisticFeature Extractor17 featuresDiscourse AnalyzerStructural Analysis5 featuresConcatenation Layer768 + 17 + 5 = 790-dimensional vectorFeature Fusion LayerDense(790 → 768) + ReLU + Dropout(0.3)Balanced feature integrationClassification HeadDense(768 → 4) + Softmax4-class probability distributionOutput: 4-Class PredictionsClass 0: Completely Fake | Class 1: Mostly FakeClass 2: Mixed/Partial | Class 3: Authentic+ Explainability Features
Discourse-Level Analysis
5 Discourse Features for Structural Coherence

Semantic Coherence

Cosine similarity between BERT embeddings of adjacent paragraphs, capturing consistency patterns.

Topic Progression

Topic transition markers (However, But, On the other hand) quantified relative to document length.

Argumentative Structure

Claim-to-evidence ratios using pattern matching for claims and evidence indicators.

Training Configuration

Base Encoder:sagorsarker/bangla-bert-base
Learning Rate:2×10⁻⁵
Batch Size:8
Warmup:10%
Loss Function:Cross-Entropy

Results & Analysis

Our interpretable framework achieves competitive performance while providing detailed explanations for classification decisions.

Model Performance Comparison
Baseline BERT vs Interpretable Model
Training Convergence
F1-Score Across Epochs
Per-Class Performance
Interpretable Model Results on Test Set
ClassPrecisionRecallF1-Score
Class 079%74%77%
Class 184%81%82%
Class 284%82%83%
Class 386%87%87%
Performance Trade-off

Only 0.11% F1-score decrease while adding complete interpretability and human-readable explanations.

0.11%
Minority Class Improvement

Macro F1-score improved from 49.69% to 53.92%, indicating better handling of class imbalance.

+4.23%
Training Efficiency

Faster convergence with interpretable features providing regularization effects.

32 min faster

Feature Analysis

Systematic patterns distinguishing authentic from fabricated content reveal deceptive communication strategies.

Psycholinguistic Markers
Authentic vs Fake News Patterns
Discourse Patterns
Structural Coherence Analysis
Psycholinguistic Insights

Uncertainty Markers

Fake news shows 2.3x higher uncertainty expressions (p < 0.001)

Emotional Manipulation

Fake articles exhibit 2.1x more negative emotion markers (p < 0.001)

Cognitive Load

Elevated repetition (0.28 vs 0.15) and reduced vocabulary richness (0.72 vs 0.81)

Self-Reference

Authentic articles show 3.1x higher self-reference ratios (p < 0.001)

Discourse Insights

Semantic Coherence

Fake news shows 19% lower coherence scores (0.58 vs 0.72, p < 0.001)

Topic Transitions

Fake articles have 1.9x more topic transitions, suggesting disorganized narratives

Claim-Evidence Imbalance

Fake news: 0.45 claims vs 0.08 evidence per paragraph (5.6x ratio)

Argumentative Structure

Authentic news maintains balanced claim-to-evidence ratios (0.22 vs 0.18)

Model Architecture

Integrated architecture combining pre-trained BERT with interpretable psycholinguistic and discourse features.

System Architecture
Input Text
Bangla News Article
→
BERT Encoder

• sagorsarker/bangla-bert-base

• 768-dimensional embeddings

• Contextual representations

Psycholinguistic Features

• 17 linguistic markers

• Emotional, uncertainty, cognitive

• Deception patterns

Discourse Features

• 5 discourse indicators

• Semantic coherence

• Argumentative structure

↓
Feature Concatenation
768 + 22 = 790 dimensions
↓
Classification Head

• Dense Layer: 790 → 768 (dropout=0.3)

• Output Layer: 768 → 4 classes

• Cross-entropy loss

↓
Classification Output
Fake / Authentic + Interpretable Features
Training Configuration
Learning Rate2×10⁻⁵
Warmup Steps10%
Batch Size8
Dropout0.3
Early Stopping Patience2 epochs
Loss FunctionCross-Entropy
Feature Contribution
Gradient Analysis on Test Set
BERT Embeddings40%
Psycholinguistic Features35%
Discourse Features25%

All feature categories provide complementary information for classification decisions.

Conclusion & Resources

This research demonstrates that systematic integration of psycholinguistic theory with modern transformer architectures can maintain competitive performance while providing interpretable explanations.

Interpretable Framework

First comprehensive psycholinguistic feature extraction system for Bangla with systematic discourse analysis integration.

Competitive Performance

84.37% F1-score with only 0.11% reduction compared to black-box approaches while enabling detailed analysis.

Actionable Insights

Identifies specific linguistic markers enabling stakeholders to understand why content was flagged.

Access the Research
Models, code, datasets, and training materials
HuggingFace Model
Pre-trained model weights & inference API
NUHASHROXME/bangla-fake-news-interpretable
Training Code & Datasets
Complete training session with reproducible experiments
YashNuhash/Bangla-Fake-News-Interpretable-Training-Session
GitHub Repository
Source code, feature extraction, and model architecture
github.com/YashNuhash
Full Paper
PDF & supplementary materials
drive.google.com/file/d/1jnDvSYsoin5uaardEGGu-LHUkP8vqoCo

Quick Links

HuggingFace ModelGitHub Training SessionDownload Paper
Authors

Md Mynoddin

Assistant Professor, Dept. of CSE, RMSTU

mynoddin@rmstu.ac.bd

Prathay Barua

Dept. of CSE, RMSTU

prathaybarua71@gmail.com

Ashraful Nuhash

Dept. of CSE, RMSTU

nuhashroxme@gmail.com

References
  1. [1] H. M. Shibu, S. Datta, M. S. Miah, N. Sami, M. S. Chowdhury, and M. S. Islam, "From scarcity to capability: Empowering fake news detection in low-resource languages with LLMs," in Proceedings of the First Workshop on Natural Language Processing for Indo-Aryan and Dravidian Languages. Abu Dhabi: Association for Computational Linguistics, Jan. 2025, pp. 100–107. [Online]. Available: https://aclanthology.org/2025.indonlp-1.12/
  2. [2] M. Z. Hossain, M. A. Rahman, M. S. Islam, and S. Kar, "Banfakenews: A dataset for detecting fake news in bangla," in Proceedings of the 12th Language Resources and Evaluation Conference (LREC). European Language Resources Association (ELRA), 2020. [Online]. Available: https://aclanthology.org/2020.lrec-1.591
  3. [3] A. S. Chowdhury, "Tackling fake news in bengali," arXiv preprint arXiv:2301.12345, 2023. [Online]. Available: https://arxiv.org/abs/2301.12345
  4. [4] M. George, "Bangla fake news detection based on multichannel combined cnn-lstm," arXiv preprint arXiv:2501.01234, 2025. [Online]. Available: https://arxiv.org/abs/2501.01234
  5. [5] F. T. J. Faria et al., "Integrating advanced fusion techniques for multimodal fake news detection in bangla," Information Fusion, 2025. [Online]. Available: https://doi.org/10.1016/j.inffus.2025.01.010
  6. [6] I. A. Fahad, K. Asif, and S. Sikder, "Banglafake: Constructing and evaluating a specialized bengali deepfake audio dataset," arXiv preprint arXiv:2503.04567, 2025. [Online]. Available: https://arxiv.org/abs/2503.04567
  7. [7] P. K. Mondal, "Deep learning approaches in bangla language," arXiv preprint arXiv:2502.09876, 2025. [Online]. Available: https://arxiv.org/abs/2502.09876
Future Research Directions
  • •Extending the framework to multimodal detection incorporating visual elements and deepfake detection
  • •Developing cross-lingual transfer capabilities for other low-resource languages
  • •Investigating adversarial robustness against evolving deceptive strategies
  • •Integration of large language models for dynamic feature generation rather than static lexicon matching

Discourse-aware Psycholinguistic Modeling for Bangla Fake News Detection

© 2025 RMSTU Department of Computer Science & Engineering

Built with v0