An end-to-end deep learning engine for aspect-level sentiment analysis in restaurant reviews
This project implements a novel joint model that simultaneously performs:
- Aspect Term Extraction (ATE) — Identifies opinion targets in a sentence (e.g., "pasta", "service")
- Aspect Sentiment Classification (ASC) — Classifies the sentiment towards each extracted aspect (positive, negative, neutral, conflict)
Unlike pipeline approaches that treat these as separate tasks, our architecture performs both tasks in a single forward pass, leveraging shared representations for better accuracy and efficiency.
- ModernBERT backbone — Latest-generation encoder with RoPE positional embeddings, replacing legacy BERT
- Relational Graph Attention (RGAT) — Encodes syntactic dependency structure (nsubj, amod, obj, etc.) to capture long-range aspect-opinion relationships
- Joint BIO + Sentiment heads — Unified architecture with token-level aspect extraction and span-level sentiment classification
┌───────────────────────┐
│ Raw Text Input │
│ "The pasta was great" │
└───────────┬───────────┘
│
┌─────────────────┼─────────────────┐
│ │ │
┌───────▼───────┐ ┌───────▼───────┐ ┌───────▼───────┐
│ ModernBERT │ │ SpaCy NLP │ │ Tokenizer │
│ Tokenizer │ │ Dep Parse │ │ Alignment │
└───────┬───────┘ └───────┬───────┘ └───────┬───────┘
│ │ │
│ ┌───────▼───────┐ │
│ │ Adjacency │ │
│ │ Tensor [R] │ │
│ └───────┬───────┘ │
│ │ │
┌───────▼─────────────────▼─────────────────▼───────┐
│ │
│ ModernBERT Encoder (22 layers) │
│ ══════════════════════════════ │
│ RoPE + Flash Attention │
│ (8 layers unfrozen for fine-tuning) │
│ │
└─────────────────────────┬─────────────────────────┘
│
┌───────▼───────┐
│ RGAT │
│ Layer │
│ (7 relations) │
└───────┬───────┘
│
┌─────────────────┴─────────────────┐
│ │
┌───────▼───────┐ ┌───────▼───────┐
│ ATE Head │ │ ASC Head │
│ Token-level │ │ Span-level │
│ BIO Tagger │ │ Classifier │
│ ─────────── │ │ ────────── │
│ O / B-ASP / │ │ pos / neg / │
│ I-ASP │ │ neu / conflict│
└───────┬───────┘ └───────┬───────┘
│ │
▼ ▼
Extracted Aspects Sentiment Labels
["pasta", "service"] [positive, negative]
RGAT Relation Types:
| Relation | Description | Example |
|---|---|---|
nsubj |
Nominal subject | pasta ← was |
amod |
Adjectival modifier | great → pasta |
obj |
Direct object | loved → pasta |
advmod |
Adverbial modifier | terribly → slow |
neg |
Negation | not → good |
compound |
Compound words | butter → chicken |
conj |
Conjunction | pasta and service |
| Dataset | ATE F1 (Strict) | ATE F1 (Partial) | ASC Accuracy | ASC Macro-F1 | Combined F1 |
|---|---|---|---|---|---|
| SemEval 2014 | 78.35% | 86.45% | 76.10% | 56.73% | 67.54% |
| SemEval 2015 | 67.96% | 74.94% | 83.84% | 59.54% | 63.75% |
| SemEval 2016 | 67.20% | 77.90% | 87.79% | 65.28% | 66.24% |
| Class | 2014 | 2015 | 2016 |
|---|---|---|---|
| Positive | 87.88% | 90.08% | 92.58% |
| Negative | 65.42% | 78.02% | 84.21% |
| Neutral | 53.61% | 10.53% | 19.05% |
| Conflict | 20.00% | — | — |
| Model | ATE F1 | ASC Accuracy | Year |
|---|---|---|---|
| BERT-SPC | — | 76.96% | 2019 |
| AEN-BERT | — | 76.31% | 2019 |
| LCF-BERT | — | 78.47% | 2019 |
| DualGCN-BERT | — | 75.92% | 2021 |
| Ours (ModernBERT-RGAT) | 78.35% | 76.10% | 2026 |
Note: Most SOTA models perform ATE and ASC as separate tasks. Our model is one of few doing joint extraction + classification in a single pass, which is a harder but more practical task.
| Component | Technology |
|---|---|
| Backbone | ModernBERT-base (149M params) |
| Graph Network | Relational Graph Attention Network (RGAT) |
| Tokenization | HuggingFace Tokenizers with BIO alignment |
| Dependency Parse | spaCy en_core_web_sm |
| Training | Mixed-precision (FP16), gradient accumulation, differential LR |
| Loss | Joint weighted CrossEntropy (ATE + ASC) with label smoothing |
| Regularization | Dropout (0.4), weight decay (0.05), gradient clipping |
| Demo | Gradio interactive web app |
# Clone the repository
git clone https://github.com/karthik0306/SOTA-ModernBERT-RGAT-Joint-Aspect-Sentiment-Extraction-for-Food-Tech-Reviews.git
cd SOTA-ModernBERT-RGAT-Joint-Aspect-Sentiment-Extraction-for-Food-Tech-Reviews
# Create environment
conda create -n modernbert_rgat python=3.10 -y
conda activate modernbert_rgat
# Install dependencies
pip install -r requirements.txt
python -m spacy download en_core_web_smfrom src.inference import load_predictor
# Load a trained model
predictor = load_predictor(year='2014', device='cpu')
# Predict
text = "The pasta was delicious but the service was terrible."
results = predictor.predict(text)
for r in results:
print(f" {r.aspect} → {r.sentiment} ({r.confidence:.0%})")
# Output:
# pasta → positive (84%)
# service → negative (91%)python app.py
# Opens Gradio interface at http://localhost:7860Run the notebooks in order:
notebooks/01_eda_profiling.ipynb # Data exploration
notebooks/02_data_pipeline.ipynb # Data preprocessing
notebooks/03_model_architecture.ipynb # Model design & smoke tests
notebooks/04_training.ipynb # Training (requires GPU)
notebooks/05_evaluation.ipynb # Evaluation & benchmarking
notebooks/06_inference_demo.ipynb # Interactive inference demo
├── configs/
│ └── config.yaml # All hyperparameters & settings
│
├── Data/
│ ├── Raw_Data/ # SemEval XML files (2014, 2015, 2016)
│ └── Processed_Data/ # Cleaned CSVs
│
├── src/ # Core source code
│ ├── model.py # ModernBERT-RGAT architecture
│ ├── dataset.py # Dataset & preprocessing (multi-aspect BIO)
│ ├── data_pipeline.py # Data loading, splits, class weights
│ ├── trainer.py # Training loop, scheduling, checkpointing
│ ├── losses.py # Joint loss function (ATE + ASC)
│ ├── evaluator.py # Metrics computation & benchmarking
│ └── inference.py # Inference pipeline
│
├── notebooks/ # Step-by-step notebooks (01–06)
├── checkpoints/ # Saved model weights (.pt)
├── outputs/ # Results, plots, logs
├── app.py # Gradio demo application
├── requirements.txt # Python dependencies
└── docs/ # Architecture & results documentation
| Parameter | Value |
|---|---|
| Backbone | answerdotai/ModernBERT-base |
| Max sequence length | 96 |
| Batch size (effective) | 16 (4 × 4 accumulation) |
| BERT learning rate | 1e-5 |
| Head learning rate | 3e-4 |
| Weight decay | 0.05 |
| Warmup | 10% of total steps |
| Label smoothing | 0.1 |
| Head dropout | 0.4 |
| Unfrozen BERT layers | 8 / 22 |
| Mixed precision | FP16 |
| Early stopping patience | 5 epochs |
| # | Notebook | Description |
|---|---|---|
| 01 | EDA & Data Profiling | Comprehensive analysis of 3 SemEval datasets with NLP statistics |
| 02 | Data Pipeline | Tokenization, BIO labeling, adjacency tensors, DataLoader creation |
| 03 | Model Architecture | ModernBERT-RGAT design, parameter analysis, forward pass validation |
| 04 | Training | Full training with mixed-precision, early stopping, checkpointing |
| 05 | Evaluation | Metrics computation, confusion matrices, cross-dataset comparison |
| 06 | Inference Demo | 3-model comparison, visual highlighting, agreement analysis |
-
Multi-aspect BIO labeling — Each training sample marks all aspects in the sentence (not just one), preventing conflicting targets for the ATE head when the same sentence appears multiple times.
-
Differential learning rates — BERT backbone fine-tuned at 1e-5, task heads at 3e-4 — prevents catastrophic forgetting while allowing heads to learn quickly.
-
RGAT for syntax — Dependency parse relations capture long-range aspect-opinion connections that positional encodings alone miss (e.g., "The pasta, which arrived late, was cold" — RGAT links pasta to cold via dependency path).
-
Joint loss balancing — Equal weighting (α=0.5) between ATE and ASC losses, both using standard CrossEntropyLoss to prevent magnitude imbalance.
- ModernBERT — Warner et al., "Smarter, Better, Faster, Longer: A Modern Bidirectional Encoder for Fast, Memory Efficient, and Long Context Finetuning and Inference", 2024
- RGAT — Busbridge et al., "Relational Graph Attention Networks", 2019
- SemEval ABSA — Pontiki et al., "SemEval-2014/2015/2016 Task 4/12: Aspect Based Sentiment Analysis", 2014-2016
- Joint ATE+ASC — Li et al., "A Unified Model for Opinion Target Extraction and Target Sentiment Prediction", 2019
This project is licensed under the MIT License — see LICENSE for details.
Built for advancing aspect-level sentiment analysis in food-tech