| --- |
| language: |
| - en |
| - hi |
| tags: |
| - hate-speech |
| - text-classification |
| - bilstm |
| - glove |
| - multilingual |
| - transfer-learning |
| - hinglish |
| - sequential-learning |
| datasets: |
| - tuklu/nprism |
| license: mit |
| model-index: |
| - name: hate-speech-multilingual-bilstm |
| results: |
| - task: |
| type: text-classification |
| name: Hate Speech Detection |
| dataset: |
| name: nprism |
| type: tuklu/nprism |
| metrics: |
| - type: f1 |
| value: 0.6419 |
| name: F1 Score (Best Strategy - Full Phase) |
| - type: accuracy |
| value: 0.6854 |
| name: Accuracy (Best Strategy - Full Phase) |
| - type: roc_auc |
| value: 0.7528 |
| name: ROC-AUC (Best Strategy - Full Phase) |
| --- |
| |
| # Multilingual Hate Speech Detection — GloVe + BiLSTM |
|
|
| **Task:** Binary text classification (Hate / Non-Hate) |
| **Languages:** English, Hindi, Hinglish (Hindi-English code-mixed) |
| **Architecture:** Bidirectional LSTM with frozen GloVe embeddings |
| **Best Strategy:** Hindi → English → Hinglish → Full (F1: 0.6419, AUC: 0.7528) |
|
|
| --- |
|
|
| ## Table of Contents |
| 1. [What This Project Does](#1-what-this-project-does) |
| 2. [The Dataset](#2-the-dataset) |
| 3. [Model Architecture](#3-model-architecture) |
| 4. [The Core Idea — Transfer Learning](#4-the-core-idea--transfer-learning) |
| 5. [The Experiment — Plan B](#5-the-experiment--plan-b) |
| 6. [Results & Best Model Selection](#6-results--best-model-selection) |
| 7. [Full Results by Strategy](#7-full-results-by-strategy) |
| 8. [All Model Checkpoints](#8-all-model-checkpoints) |
| 9. [How to Use](#9-how-to-use) |
|
|
| --- |
|
|
| ## 1. What This Project Does |
|
|
| This project investigates whether the **order of language exposure** during sequential transfer learning affects a model's ability to detect hate speech across three languages: English, Hindi, and Hinglish. |
|
|
| The key question: |
|
|
| > If you train a model on English first, then Hindi, then Hinglish — does it perform better or worse than training Hinglish first? |
|
|
| We ran all **6 possible orderings**, each followed by a final training pass on the complete shuffled dataset, and measured performance after every single phase. |
|
|
| --- |
|
|
| ## 2. The Dataset |
|
|
| Dataset: [tuklu/nprism](https://huggingface.co/datasets/tuklu/nprism) |
|
|
| | Split | Samples | |
| |---|---| |
| | Train | 17,704 | |
| | Validation | 2,950 | |
| | Test | 8,852 | |
| | **Total** | **29,505** | |
|
|
| | Language | Count | % | |
| |---|---|---| |
| | English | 14,994 | 50.8% | |
| | Hindi | 9,738 | 33.0% | |
| | Hinglish | 4,774 | 16.2% | |
|
|
| | Label | Count | % | |
| |---|---|---| |
| | Non-Hate (0) | 15,799 | 53.5% | |
| | Hate (1) | 13,707 | 46.5% | |
|
|
|  |
|
|
| The pie chart above shows the dataset is dominated by English (50.8%), with Hindi and Hinglish making up the rest. This imbalance is important — it means the model sees more English examples and GloVe embeddings are English-centric, which directly explains why English phase always achieves the highest accuracy. |
|
|
| --- |
|
|
| ## 3. Model Architecture |
|
|
| ``` |
| Input: Text sequence (max 100 tokens) |
| ↓ |
| GloVe Embedding Layer (vocab: 50,000 × 300d) — FROZEN |
| ↓ |
| Bidirectional LSTM (128 units) |
| → reads sentence left-to-right AND right-to-left |
| → captures context from both directions |
| ↓ |
| Dropout (0.5) — randomly disables 50% of neurons during training |
| → prevents memorising training data (overfitting) |
| ↓ |
| Dense Layer (64 neurons, ReLU activation) |
| ↓ |
| Output Layer (1 neuron, Sigmoid) |
| → outputs probability 0.0 to 1.0 |
| → > 0.5 = Hate Speech |
| → ≤ 0.5 = Not Hate Speech |
| ``` |
|
|
| **Why GloVe?** |
| GloVe (Global Vectors) is a pre-trained word embedding trained on 6 billion tokens. Each word becomes a 300-number vector that captures semantic meaning — "hate" and "violence" end up close together in this 300-dimensional space. We freeze it (don't update during training) to preserve this general knowledge and only train the layers on top. |
|
|
| **Why BiLSTM?** |
| A regular LSTM reads text left to right. A BiLSTM reads it both ways and combines the results. The sentence *"I don't hate you"* needs both directions to understand the negation — the word "don't" only makes sense in context of what comes after it. |
|
|
| **Training config:** |
| - Optimizer: Adam |
| - Loss: Binary Cross-Entropy |
| - Epochs per phase: 8 |
| - Batch size: 32 (64 for full phase) |
| - Max sequence length: 100 tokens |
|
|
| --- |
|
|
| ## 4. The Core Idea — Transfer Learning |
|
|
| **Transfer learning** = the model keeps what it learned from one task when starting the next one. |
|
|
| Think of it like a student who already knows French — learning Spanish is faster because both share Latin roots. The vocabulary, grammar intuitions, and reading skills transfer. |
|
|
| In our case: train on English → the model learns what "hate speech patterns" look like in a language GloVe understands well → then fine-tune on Hindi → the model adapts those patterns to Hindi → then Hinglish → the model adapts again using everything it knows. |
|
|
| ### The Bug That Was Fixed |
|
|
| The original code was reinitialising the model inside the loop — meaning **every language got a brand new, untrained model**. That is not transfer learning at all. |
|
|
| ```python |
| # WRONG — model reset every iteration, no knowledge transfer |
| for lang in languages: |
| model = Sequential() # ← destroys all previous learning |
| model.fit(X_lang, ...) |
| |
| # CORRECT — model built once, weights carry forward |
| model = build_model() # ← built once before the loop |
| for lang in languages: |
| model.fit(X_lang, ...) # ← each fit continues from where previous left off |
| ``` |
|
|
| This single fix is the entire point of the experiment. |
|
|
| --- |
|
|
| ## 5. The Experiment — Plan B |
|
|
| We tested all 6 permutations of [English, Hindi, Hinglish], each ending with a full shuffled dataset phase: |
|
|
| | # | Training Order | |
| |---|---| |
| | 1 | English → Hindi → Hinglish → Full | |
| | 2 | English → Hinglish → Hindi → Full | |
| | 3 | Hindi → English → Hinglish → Full | |
| | 4 | Hindi → Hinglish → English → Full | |
| | 5 | Hinglish → English → Hindi → Full | |
| | 6 | Hinglish → Hindi → English → Full | |
|
|
| **After each phase**, the model is immediately evaluated on **that specific language's test subset**. So for strategy `English → Hindi → Hinglish → Full`: |
|
|
| ``` |
| Train on English → evaluate English test set → save metrics + plots |
| Train on Hindi → evaluate Hindi test set → save metrics + plots |
| Train on Hinglish → evaluate Hinglish test set → save metrics + plots |
| Train on Full data → evaluate full test set → save metrics + plots |
| ``` |
|
|
| This gives us 4 snapshots per strategy — letting us see exactly how the model evolves as it learns each new language. |
|
|
| --- |
|
|
| ## 6. Results & Best Model Selection |
|
|
| ### Full Phase Results (Final Model Performance) |
|
|
| | Strategy | Accuracy | Balanced Acc | Precision | Recall | Specificity | F1 | ROC-AUC | |
| |---|---|---|---|---|---|---|---| |
| | **Hindi → English → Hinglish → Full** | 0.6854 | **0.6802** | 0.6810 | 0.6070 | 0.7534 | **0.6419** | 0.7528 | |
| | Hindi → Hinglish → English → Full | **0.6865** | 0.6801 | 0.6900 | 0.5905 | 0.7698 | 0.6364 | 0.7507 | |
| | Hinglish → Hindi → English → Full | 0.6845 | 0.6775 | 0.6918 | 0.5786 | 0.7764 | 0.6301 | **0.7548** | |
| | English → Hinglish → Hindi → Full | 0.6813 | 0.6740 | 0.6899 | 0.5703 | 0.7776 | 0.6244 | 0.7535 | |
| | Hinglish → English → Hindi → Full | 0.6778 | 0.6718 | 0.6768 | 0.5866 | 0.7570 | 0.6285 | 0.7521 | |
| | English → Hindi → Hinglish → Full | 0.6796 | 0.6678 | 0.7243 | 0.5010 | 0.8346 | 0.5923 | 0.7599 | |
|
|
| ### Why Hindi → English → Hinglish → Full is the Best Model |
|
|
| **F1 Score is the most important metric here.** For hate speech detection, we need to balance two things: |
| - **Precision** — don't falsely flag innocent content as hate |
| - **Recall** — don't miss actual hate speech |
|
|
| F1 is the harmonic mean of both. A model that misses half the hate speech (low recall) or flags everything as hate (low precision) is useless in practice. |
|
|
| Look at `English → Hindi → Hinglish → Full` — it has the highest ROC-AUC (0.7599) but an F1 of only 0.5923. Why? Its Recall is only 0.5010 — it misses **half of all hate speech**. High ROC-AUC can be misleading when threshold calibration is off. |
|
|
| `Hindi → English → Hinglish → Full` has: |
| - Best F1 (0.6419) — best balance of precision and recall |
| - Best Balanced Accuracy (0.6802) — most fair across both classes |
| - Recall of 0.607 — catches significantly more hate speech than alternatives |
|
|
| **Why does Hindi-first work better?** |
|
|
| Hindi is the hardest language for this model (GloVe has limited Hindi coverage). Training on Hindi *first* forces the model to develop general hate-speech-detection features that aren't dependent on GloVe's English-centric embeddings. It learns to detect patterns from context and sequence rather than relying on word meanings alone. When English comes next, the model improves dramatically and carries robust features forward. English-first strategies give the model an easy start but it never develops the robustness needed for low-resource languages. |
|
|
| ### Best Model Training Curves (Hindi → English → Hinglish → Full) |
|
|
| **Phase 1: Train on Hindi** |
|
|
|  |
|
|
| The model starts cold on Hindi. Accuracy is low (~55-57%) and validation loss is unstable — this is expected. GloVe doesn't cover Hindi well so the model is learning purely from sequential patterns. The struggle here is valuable — it forces the model to build language-agnostic features. |
|
|
| **Phase 2: Train on English** |
|
|
|  |
|
|
| Dramatic improvement. The model jumps to ~77-78% accuracy. GloVe embeddings now align well with the input language. Notice that it doesn't start from scratch — the Hindi training gave it a base of sequential hate-speech patterns, and now with English vocabulary the model improves rapidly. |
|
|
| **Phase 3: Train on Hinglish** |
|
|
|  |
|
|
| Hinglish is code-mixed — it borrows from both languages the model already knows. Training accuracy climbs to ~68-69%. The model adapts its existing knowledge to handle the mixed vocabulary. |
|
|
| **Phase 4: Train on Full Dataset** |
|
|
|  |
|
|
| Final fine-tuning on all 17,704 shuffled training samples. Training and validation accuracy converge, loss stabilises. This phase consolidates all language knowledge into the final model. |
|
|
| ### Best Model Evaluation Charts |
|
|
| **Confusion Matrix:** |
|
|
|  |
|
|
| Shows actual vs predicted counts. A well-balanced confusion matrix means the model is not biased toward one class. True Positives (hate correctly identified) and True Negatives (non-hate correctly identified) should both be high. |
|
|
| **ROC Curve (AUC = 0.7528):** |
|
|
|  |
|
|
| The ROC curve shows the trade-off between True Positive Rate (catching hate speech) and False Positive Rate (wrongly flagging non-hate). AUC of 0.7528 means the model has a 75.3% chance of correctly ranking a hate speech example higher than a non-hate example — significantly better than random (0.5). |
|
|
| **Precision-Recall Curve:** |
|
|
|  |
|
|
| Shows the trade-off between precision and recall at different thresholds. The curve staying high across recall values means the model maintains good precision even as it catches more hate speech. Useful for choosing the operating threshold based on deployment requirements. |
|
|
| **F1 vs Threshold Curve:** |
|
|
|  |
|
|
| Shows F1 score at every possible decision threshold. The peak is near 0.5 confirming our threshold choice is well-calibrated. If deploying in a high-recall scenario (catch all hate speech even at cost of false positives), lower the threshold; for high-precision (only flag certain hate speech), raise it. |
|
|
| --- |
|
|
| ## 7. Full Results by Strategy |
|
|
| ### Strategy 1: English → Hindi → Hinglish → Full |
|
|
| | Phase | Accuracy | F1 | ROC-AUC | |
| |---|---|---|---| |
| | English | 0.7701 | 0.7696 | 0.8504 | |
| | Hindi | 0.5507 | 0.0000 | 0.5689 | |
| | Hinglish | 0.6780 | 0.5155 | 0.6691 | |
| | Full | 0.6796 | 0.5923 | 0.7599 | |
|
|
| **Note on the Hindi phase row** — Precision=0, Recall=0, F1=0, Specificity=1.0. This is not a data error. After training only on English, the model predicted **zero hate speech** for every Hindi test sample — it classified everything as non-hate. This means: |
| - Specificity = 1.0 ✓ (no false positives — because it never predicts hate at all) |
| - Recall = 0.0 (catches zero actual hate speech) |
| - F1 = 0.0 (completely useless for Hindi at this stage) |
|
|
| This is the strongest evidence that English-first is the wrong order — the model becomes so tuned to English patterns that it cannot generalise to Hindi at all. |
|
|
| | Phase | Training Curves | Confusion Matrix | ROC | PR | F1 Curve | |
| |---|---|---|---|---|---| |
| | English |  |  |  |  |  | |
| | Hindi |  |  |  |  |  | |
| | Hinglish |  |  |  |  |  | |
| | Full |  |  |  |  |  | |
|
|
| --- |
|
|
| ### Strategy 2: English → Hinglish → Hindi → Full |
|
|
| | Phase | Accuracy | F1 | ROC-AUC | |
| |---|---|---|---| |
| | English | 0.7721 | 0.7743 | 0.8525 | |
| | Hinglish | 0.6631 | 0.5460 | 0.6899 | |
| | Hindi | 0.5810 | 0.4444 | 0.5975 | |
| | Full | 0.6813 | 0.6244 | 0.7535 | |
|
|
| | Phase | Training Curves | Confusion Matrix | ROC | PR | F1 Curve | |
| |---|---|---|---|---|---| |
| | English |  |  |  |  |  | |
| | Hinglish |  |  |  |  |  | |
| | Hindi |  |  |  |  |  | |
| | Full |  |  |  |  |  | |
|
|
| --- |
|
|
| ### Strategy 3: Hindi → English → Hinglish → Full ⭐ BEST MODEL |
|
|
| | Phase | Accuracy | F1 | ROC-AUC | |
| |---|---|---|---| |
| | Hindi | 0.5662 | 0.2860 | 0.5748 | |
| | English | 0.7780 | 0.7830 | 0.8549 | |
| | Hinglish | 0.6880 | 0.5641 | 0.7172 | |
| | **Full** | **0.6854** | **0.6419** | **0.7528** | |
|
|
| Starting with the hardest language (Hindi) builds robustness. Despite the rough start, the model recovers strongly and achieves the best final F1. |
|
|
| | Phase | Training Curves | Confusion Matrix | ROC | PR | F1 Curve | |
| |---|---|---|---|---|---| |
| | Hindi |  |  |  |  |  | |
| | English |  |  |  |  |  | |
| | Hinglish |  |  |  |  |  | |
| | Full |  |  |  |  |  | |
|
|
| --- |
|
|
| ### Strategy 4: Hindi → Hinglish → English → Full |
|
|
| | Phase | Accuracy | F1 | ROC-AUC | |
| |---|---|---|---| |
| | Hindi | 0.5779 | 0.3898 | 0.5972 | |
| | Hinglish | 0.6986 | 0.5289 | 0.7109 | |
| | English | 0.7780 | 0.7816 | 0.8563 | |
| | Full | 0.6865 | 0.6364 | 0.7507 | |
|
|
| | Phase | Training Curves | Confusion Matrix | ROC | PR | F1 Curve | |
| |---|---|---|---|---|---| |
| | Hindi |  |  |  |  |  | |
| | Hinglish |  |  |  |  |  | |
| | English |  |  |  |  |  | |
| | Full |  |  |  |  |  | |
|
|
| --- |
|
|
| ### Strategy 5: Hinglish → English → Hindi → Full |
|
|
| | Phase | Accuracy | F1 | ROC-AUC | |
| |---|---|---|---| |
| | Hinglish | 0.6652 | 0.5119 | 0.6692 | |
| | English | 0.7716 | 0.7829 | 0.8484 | |
| | Hindi | 0.5638 | 0.2466 | 0.5982 | |
| | Full | 0.6778 | 0.6285 | 0.7521 | |
|
|
| | Phase | Training Curves | Confusion Matrix | ROC | PR | F1 Curve | |
| |---|---|---|---|---|---| |
| | Hinglish |  |  |  |  |  | |
| | English |  |  |  |  |  | |
| | Hindi |  |  |  |  |  | |
| | Full |  |  |  |  |  | |
|
|
| --- |
|
|
| ### Strategy 6: Hinglish → Hindi → English → Full |
|
|
| | Phase | Accuracy | F1 | ROC-AUC | |
| |---|---|---|---| |
| | Hinglish | 0.6837 | 0.5369 | 0.6929 | |
| | Hindi | 0.5924 | 0.4656 | 0.5964 | |
| | English | 0.7765 | 0.7811 | 0.8534 | |
| | Full | 0.6845 | 0.6301 | 0.7548 | |
|
|
| | Phase | Training Curves | Confusion Matrix | ROC | PR | F1 Curve | |
| |---|---|---|---|---|---| |
| | Hinglish |  |  |  |  |  | |
| | Hindi |  |  |  |  |  | |
| | English |  |  |  |  |  | |
| | Full |  |  |  |  |  | |
|
|
| --- |
|
|
| ## 8. All Model Checkpoints |
|
|
| All 6 trained models are available as archives in the `models/` folder of this repo. Each filename encodes the training order. |
|
|
| | File | Strategy | Final F1 | Final AUC | |
| |---|---|---|---| |
| | `model.h5` | Hindi → English → Hinglish → Full ⭐ | 0.6419 | 0.7528 | |
| | `models/planB_hindi_to_english_to_hinglish_Full.h5` | Hindi → English → Hinglish → Full | 0.6419 | 0.7528 | |
| | `models/planB_hindi_to_hinglish_to_english_Full.h5` | Hindi → Hinglish → English → Full | 0.6364 | 0.7507 | |
| | `models/planB_hinglish_to_hindi_to_english_Full.h5` | Hinglish → Hindi → English → Full | 0.6301 | 0.7548 | |
| | `models/planB_english_to_hinglish_to_hindi_Full.h5` | English → Hinglish → Hindi → Full | 0.6244 | 0.7535 | |
| | `models/planB_hinglish_to_english_to_hindi_Full.h5` | Hinglish → English → Hindi → Full | 0.6285 | 0.7521 | |
| | `models/planB_english_to_hindi_to_hinglish_Full.h5` | English → Hindi → Hinglish → Full | 0.5923 | 0.7599 | |
|
|
| --- |
|
|
| ## 9. How to Use |
|
|
| ```python |
| import json |
| import numpy as np |
| import tensorflow as tf |
| from tensorflow.keras.preprocessing.text import tokenizer_from_json |
| from tensorflow.keras.preprocessing.sequence import pad_sequences |
| from huggingface_hub import hf_hub_download |
| |
| # Load tokenizer |
| tokenizer_path = hf_hub_download(repo_id="tuklu/SASC", filename="tokenizer.json") |
| with open(tokenizer_path) as f: |
| tokenizer = tokenizer_from_json(f.read()) |
| |
| # Load best model |
| model_path = hf_hub_download(repo_id="tuklu/SASC", filename="model.h5") |
| model = tf.keras.models.load_model(model_path) |
| |
| # Predict |
| texts = ["I hate all of them", "Have a great day!"] |
| sequences = tokenizer.texts_to_sequences(texts) |
| padded = pad_sequences(sequences, maxlen=100) |
| probs = model.predict(padded).flatten() |
| |
| for text, prob in zip(texts, probs): |
| label = "Hate Speech" if prob > 0.5 else "Non-Hate" |
| print(f"{label} ({prob:.3f}): {text}") |
| ``` |
|
|
| --- |
|
|
| ## Explainability — SHAP Analysis |
|
|
| We applied **SHAP (SHapley Additive exPlanations)** to all 6 trained models to understand which words drive hate speech predictions. A `GradientExplainer` runs on the BiLSTM sub-model (embedding layer bypassed — embeddings pre-computed as floats), with 200 background training samples. Each model is evaluated on all 4 test sets (English, Hindi, Hinglish, Full). |
|
|
| > Full methodology, all plots, and detailed word tables: **[SHAP_REPORT.md](SHAP_REPORT.md)** |
|
|
| ### Best Model (Hindi → English → Hinglish) — Top SHAP Words |
|
|
| | Eval | Top Hate Words | Top Non-Hate Words | |
| |---|---|---| |
| | English | credence, bj, ghazi, eni | plain, stranger, sarcasm, rubbish | |
| | Hindi | कॉल, भूमिपूजन, मूर्ख | मैसेज, पुलिसकर्मी, जाएगी | |
| | Hinglish | bacchi, bull, srk, behan | madrassa, gdp, bech | |
| | Full | skua, brut, cleansing, baar | taraf, directory, quran | |
|
|
|  |
|
|
|  |
|
|
| ### Cross-Model Comparison (Full Test Set) |
|
|
| Words appearing in the top-10 of at least 3 models — shows which signals are consistent vs strategy-specific: |
|
|
|  |
|
|
| ### Key Takeaways |
|
|
| - **Hindi SHAP values are 10× smaller** than English/Hinglish — confirms GloVe has near-zero Hindi coverage; the model relies on positional patterns, not semantics |
| - **"online" and "rajya"** are consistent non-hate signals across all 6 models — informational/political discussion context |
| - **Accusatory verbs** (`blame`, `blaming`, `criticized`) and **violence language** (`massacres`, `cleansing`) are the most coherent English hate markers |
| - **Spurious correlations visible** (`syntax`, `skua`, `ahh`) — expected limitation of non-contextual GloVe embeddings |
|
|
| --- |
|
|
| ## Citation |
|
|
| ``` |
| @misc{sasc2026, |
| title={Multilingual Hate Speech Detection via Sequential Transfer Learning}, |
| author={tuklu}, |
| year={2026}, |
| publisher={HuggingFace}, |
| url={https://huggingface.co/tuklu/SASC} |
| } |
| ``` |
|
|
