Active: Multilingual Transformer • v1.0
Academic Formulation: x = {T, V, C} → y ∈ {0,1}

Hate Speech Detection Platform

Explainable AI & Reward Optimized LLMs for Multilingual, Code-Mixed, and Multimodal Content Moderation.

Total Dataset Samples
325
Reward Optimization Runs
1
Human Feedback Audited
6
Experiments Completed
1/9
System Architecture Pipeline

Social Media Input (T, V, C) → Language & Modality Detection → Modality Preprocessing (Text Normalization, EasyOCR, Context Sanitization) → Feature Extraction & Cross-Attention Fusion → Hate Speech Classification → Grounded XAI Attribution → Human Feedback & Multi-Objective Reward Calculation → DPO Optimization.

English Hindi (Devanagari) Hinglish Code-Mixed Multimodal Memes Direct Preference Optimization Integrated Gradients / Perturbation
Core Research Mandates
  • Strict Binary Objective: 0 = Non-Hateful, 1 = Hateful.
  • No Hardcoding Rule: All metrics, predictions, explanations, and rewards are dynamically computed.
  • Multilingual Robustness: Devanagari script and Latin Romanized Hindi slang support without blind machine translation.
  • Multimodal Meme Analysis: EasyOCR extraction and Cross-Modal Cross-Attention fusion.
Quick Action Links

Hate Speech Detection Workspace

Interactive multimodal inference with grounded Explainable AI attributions and instant feedback collection.

Input Modalities: x = {T, V, C}
Load Academic Test Presets:

Dataset Laboratory

Universal dataset adapters, schema validation, class balance verification, and leakage-safe splits.

Registered Research Datasets
Dataset Name Identifier Samples Languages Modalities Status Actions

Preprocessing Studio

Inspect the actual transformation stages for text, multilingual code-mixing, memes, and context.

Text & Script Normalization

Applies Unicode NFKC normalization, URL neutralization ([URL]), mention anonymization ([MENTION]), camelCase hashtag decomposition, and Devanagari script integrity checking.

Raw: "Inko maro! #DeshBhakti https://t.co/xyz"
Normalized: "Inko maro ! Desh Bhakti [URL]"
Code-Mixed Hinglish Detection

Detects Latin vs Devanagari unicode characters and matches phonetic Romanized Hindi lexicon (e.g. kamine, bhai, desh, aukaat) to reliably identify Hinglish code-mixing without English translation loss.

Visual & EasyOCR Extraction

Loads meme graphics, resizes to [224, 224], extracts embedded text and spatial bounding boxes via EasyOCR, and normalizes pixel tensors for cross-modal cross-attention.

Context Sanitization

Sanitizes thread history and masks protected demographic references ([DEMOGRAPHIC_NEUTRAL]) to prevent models from learning identity-term biases.

Model Laboratory & Registry

Architectures, trained checkpoints, live performance metrics, and active model management.

Available Model Architectures
Model Name Identifier Version Type Accuracy / Macro F1 Status Action
Model Version Audit Trail (Rollback Supported)
Version ID Optimization Method Reward Score Macro F1 Created At

Explainability (XAI) Studio

Grounded token-level attribution, visual bounding saliency, and AOPC faithfulness validation.

Methodology: Grounded Perturbation & Integrated Gradients

Traditional attention maps frequently fail faithfulness tests. Our engine implements Leave-One-Out (LIME-style) perturbation attribution and verifies explanation faithfulness by measuring confidence drop (Δf(x)) when top-k attributed tokens are ablated.

Faithfulness Score Formula: AOPC_Drop = |f(x) - f(x \ top_k)| / f(x)
Positive Evidence Tokens

Tokens that elevate hateful logit probability (slurs, dehumanization, violent imperatives).

vermin [+0.42]
exterminate [+0.38]
kutto [+0.35]
Negative Evidence Tokens

Tokens that reduce hateful logit probability (counterspeech, civil terms, quotation critiques).

virodh [-0.31]
equality [-0.29]
peace [-0.25]

Reward Optimization Laboratory

Execute actual feedback-driven Direct Preference Optimization (DPO) and monitor reward metrics.

Documented Multi-Objective Reward Function
R(x, y, e, f) = α·R_cls + β·R_exp + γ·R_ctx + δ·R_fb - λ_err·R_err
Current Weights: α=0.35, β=0.25, γ=0.15, δ=0.25, λ_err=0.50

Rewards model accuracy, faithful explanations, and context awareness while heavily penalizing confident false positives.

Trigger Optimization Cycle
Optimization Run History
Run Name Method Preference Pairs Macro F1 (Before → After) Status Executed At
Recent Instance Rewards (Audit Log)
ID R_cls R_exp R_ctx R_fb R_err (Penalty) Total Reward

Bias and Fairness Laboratory

Disaggregated group metrics, False Positive Rate (FPR) parity, and Equalized Odds analysis.

Equalized Odds Disparity Gap
0.0%
Fairness Standard
EQUITABLE PARITY MET
Disaggregated Performance by Language Group
Language Subgroup Samples Accuracy Macro F1 False Positive Rate (FPR) False Negative Rate (FNR)

Research Experiment Laboratory

Run and verify Experiments 1 through 9 directly with real backend executions and metric calculations.

Research Reports & Academic Deliverables

Official ACL Paper Draft and Presentation Deck.

ACL Research Paper Draft (LaTeX Format)

Official ACL formatting following the core research problem statement.


          


          

            
Presentation Slide Deck

            
A 15-slide comprehensive viva presentation covering Motivation, Problem Formulation, Multimodal Architecture, XAI Faithfulness, and Reward Optimization.

            Open Interactive Presentation Slides →
          

        


        
        

          
System Settings & Hyperparameters

          
Centralized configuration without requiring source-code modifications.


          

            
Compute & Execution Settings

            

              

                
                
              

              

                
                
              

            

          


          

            
Multi-Objective Reward Weights