owen4512/minilm-finance-term-aligner

Multilingual Finance Term Matching Model (MiniLM)

A multilingual sentence embedding model fine-tuned for cross-lingual financial terminology matching.

---

🧠 Model Description

This model is fine-tuned from sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2 to improve semantic similarity between financial terms across languages.

It is designed to map bilingual terminology (e.g., Chinese–English term pairs) into a shared embedding space, enabling efficient cross-lingual retrieval and alignment.

---

🎯 Intended Use

This model is suitable for:

• Cross-lingual terminology matching
• Financial term alignment (ZH ↔ EN)
• Translation memory enhancement
• Semantic search / retrieval
• Terminology normalization

---

🚫 Out-of-Scope Use

This model is not intended for:

• Long document semantic comparison
• General-purpose multilingual understanding
• High-risk decision systems without human validation

---

📊 Training Data

• Source: WMT 2025 terminology dataset
• Format: bilingual term pairs (Chinese ↔ English)
• Structure: CSV file with two columns
  • Column 1: Chinese term
  • Column 2: English term

No explicit labels or negative samples are required.

---

🏗️ Training Pipeline

The model is fine-tuned using the Sentence-Transformers v3 training API.

Data Processing

• Dataset loaded via Hugging Face datasets
• Only first two columns used (anchor, positive)
• Dataset shuffled with fixed seed
• Optional train/dev split (dev_ratio=0.1)

---

⚙️ Training Details

• Base model: paraphrase-multilingual-MiniLM-L12-v2
• Framework: Sentence-Transformers
• Loss: MultipleNegativesRankingLoss

Key Mechanism

• Uses in-batch negatives
• Each batch:
  • positive pair = (zh_term, en_term)
  • all other pairs = implicit negatives

---

🧪 Training Configuration

• num_epochs: 5–10
• batch_size: 32–64
• learning_rate: 2e-5
• warmup_ratio: 0.1

Training Strategy

• evaluation: per epoch (if dev set provided)
• checkpoint saving: per epoch
• best model selection: enabled when validation exists
• no-duplicate batch sampling

---

⚡ Hardware & HPC Setup

This training pipeline is designed for offline HPC environments:

• Custom Hugging Face cache (HF_HOME)
• Offline mode enabled:
  • HF_HUB_OFFLINE=1
  • TRANSFORMERS_OFFLINE=1
• Optional GPU acceleration (fp16)

---

🚀 Usage

from sentence_transformers import SentenceTransformer, util

model = SentenceTransformer("yourname/minilm-finance-term-matcher")

emb1 = model.encode("可转换债券")
emb2 = model.encode("convertible bond")

score = util.cos_sim(emb1, emb2)
print(score)
🧾 Example

Input:

"可转换债券"
"convertible bond"

Output:

cosine similarity ≈ high (close match)
🔗 Integration with Term Extraction

This model is designed to work together with:

BERT-based term extractors (EN & ZH)

Pipeline:

Extract terms from text
Encode terms
Perform cross-lingual matching
⚠️ Limitations
Domain-specific (finance)
Performance depends on term coverage
Not optimized for long sentences
Sensitive to domain shift
📜 License

This model is trained on data licensed under CC BY-NC 4.0:

✅ Non-commercial use allowed
❌ Commercial use restricted
✅ Attribution required

The base model is Apache 2.0, but fine-tuned weights inherit dataset restrictions.

🙏 Acknowledgements
Base model: sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2
Dataset: WMT 2025 terminology resources
Framework: Sentence-Transformers
Training API: SentenceTransformerTrainer