| Presentation link: https://aaudk-my.sharepoint.com/:v:/g/personal/ev52yu_student_aau_dk/IQBhbanv3sOmR4qpqKzNWJ8LASJjBYmnSK-GsC4yRPtShL8?nav=eyJyZWZlcnJhbEluZm8iOnsicmVmZXJyYWxBcHAiOiJPbmVEcml2ZUZvckJ1c2luZXNzIiwicmVmZXJyYWxBcHBQbGF0Zm9ybSI6IldlYiIsInJlZmVycmFsTW9kZSI6InZpZXciLCJyZWZlcnJhbFZpZXciOiJNeUZpbGVzTGlua0NvcHkifX0&e=spnj1M |
| |
| # 🌱 Green Patent Classification using PatentSBERTa |
| ### Silver Supervision + Active Learning + Human-in-the-Loop Fine-Tuning |
| |
| This repository contains a fine-tuned version of **PatentSBERTa** for binary classification of patent claims into: |
| |
| - **0 — Non-Green** |
| - **1 — Green Technology** |
| |
| |
| We created a balanced dataset of 50,000 claims: |
| |
| - 25,000 Green (silver) |
| - 25,000 Non-Green (silver) |
| |
| Stratified split: |
| |
| - 70% Train |
| - 15% Pool (for active learning) |
| - 15% Evaluation |
| |
| --- |
| |
| # 🧠 Baseline Model (Frozen Embeddings) |
| |
| We computed claim embeddings using: |
| |
| |
| A Logistic Regression classifier was trained on frozen embeddings. |
| |
| ### Silver Evaluation Performance |
| |
| - Accuracy ≈ 77.5% |
| - Balanced precision and recall across classes |
| |
| This confirms that domain-specific embeddings capture sustainability semantics effectively even without fine-tuning. |
| |
| --- |
| |
| # 🔍 Active Learning Strategy |
| |
| To improve labeling efficiency, we selected high-risk examples from the unlabeled pool using uncertainty sampling: |
| |
| |
| Where: |
| - `p_green` is the model’s predicted probability for class 1. |
| - Higher `u` indicates higher classification uncertainty. |
|
|
| The top 100 most uncertain examples were selected for expert review. |
|
|
| --- |
|
|
| # 🤖 LLM → Human HITL Workflow |
|
|
| ## LLM Used |
|
|
| qwen2.5:7b-instruct (via Ollama) |
|
|
| The LLM provided: |
| - Suggested label (0/1) |
| - Confidence level (low / medium / high) |
| - Short rationale quoting the claim |
|
|
| Human reviewers assigned the final gold label. |
|
|
| --- |
|
|
| # 🔄 Human Overrides (Required Reporting) |
|
|
| Override defined as: |
|
|
|
|
| This measures disagreement between the LLM and human expert. |
|
|
| ## Example Overrides |
|
|
| 1) **LLM 0 → Human 1** |
| Claim: Conversion of biological material into energy. |
| Reason: Biomass-based energy conversion qualifies as renewable energy under green technology criteria. |
|
|
| 2) **LLM 0 → Human 1** |
| Claim: Light concentrating optic for photovoltaic device. |
| Reason: The invention directly improves solar energy generation, which is renewable and climate-mitigating. |
|
|
| ### Observed Pattern |
|
|
| The LLM tends to under-classify cases where renewable energy impact is implied but not explicitly framed as “climate mitigation.” |
|
|
| This highlights the importance of human-in-the-loop review for borderline sustainability cases. |
|
|
| --- |
|
|
| # 🚀 Final Fine-Tuning |
|
|
| We fine-tuned: |
|
|
|
|
| Training data: |
| - Silver training split |
| - + 100 human-labeled gold examples |
|
|
| Training configuration: |
|
|
| - Optimizer: AdamW |
| - Learning rate: 2e-5 |
| - Epochs: 1 |
| - Max sequence length: 256 |
| - Batch size: 4 |
|
|
| --- |
|
|
| # 📈 Evaluation |
|
|
| The fine-tuned model was evaluated on: |
|
|
| - Silver evaluation set |
| - Gold HITL set |
|
|
| Fine-tuning improves performance especially on previously uncertain claims. |
|
|
| --- |
|
|
| # 💻 Usage Example |
|
|
| ```python |
| from transformers import AutoTokenizer, AutoModelForSequenceClassification |
| import torch |
| |
| model_path = "your-model-path" |
| |
| tokenizer = AutoTokenizer.from_pretrained(model_path) |
| model = AutoModelForSequenceClassification.from_pretrained(model_path) |
| |
| text = "A system for converting biomass into renewable energy." |
| |
| inputs = tokenizer(text, return_tensors="pt", truncation=True, padding=True, max_length=256) |
| |
| with torch.no_grad(): |
| outputs = model(**inputs) |
| |
| prediction = torch.argmax(outputs.logits, dim=-1).item() |
| |
| print("Green" if prediction == 1 else "Non-Green") |
| |
