🎯 MultiSent-E5-Pro: Advanced Thai Sentiment Classifier

🇹🇭 State-of-the-art Thai sentiment analysis with multilingual capabilities

📋 Quick Overview

MultiSent-E5-Pro is a fine-tuned sentiment analysis model based on intfloat/multilingual-e5-large, specially optimized for Thai with support for multilingual contexts. The model classifies text into four categories: Positive, Negative, Neutral, and Question.

🎯 Key Features

• Handles Thai-specific expressions, colloquialisms, and sarcasm effectively
• Performs well on real-world social media & review data
• Multilingual support for Southeast and East Asian languages

---

🏆 Benchmark Summary

Rank	Model	Accuracy	F1-Macro	Notes
🥇 1st	MultiSent-E5-Pro	84.61%	84.61%	Best overall
2nd	MultiSent-E5	80.62%	80.62%	Baseline model
3rd	sentiment-103	57.40%	49.87%	Moderate baseline

---

📊 Detailed Metrics (2,183 samples)

Metric	Score
Accuracy	84.61%
F1-Macro	84.61%
F1-Weighted	84.75%
Avg Confidence	98.53%
Low Confidence Rate (<60%)	0.96%

Per-Class Performance

Class	Precision	Recall	F1	Notes
Negative	91.0%	84.6%	87.7%	Excellent
Positive	83.0%	94.3%	88.3%	Excellent
Neutral	71.9%	81.6%	76.4%	Moderate
Question	94.4%	79.0%	86.0%	Good

---

⚡ Quick Start

from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch

model = "ZombitX64/MultiSent-E5-Pro"
tokenizer = AutoTokenizer.from_pretrained(model)
model = AutoModelForSequenceClassification.from_pretrained(model)

text = "ผลิตภัณฑ์นี้ดีมาก ใช้งานง่าย"
inputs = tokenizer(text, return_tensors="pt", truncation=True)
with torch.no_grad():
    outputs = model(**inputs)
    probs = torch.nn.functional.softmax(outputs.logits, dim=-1)
    predicted = torch.argmax(probs, dim=-1)

labels = ["Question", "Negative", "Neutral", "Positive"]
print(f"Sentiment: {labels[predicted.item()]} (Confidence: {probs[0][predicted].item():.2%})")

---

🌟 Use Cases

Application	Suitability
Product Reviews	🟢 Excellent
Social Media	🟢 Excellent
Customer Support	🟢 Excellent
Content Moderation	🟡 Good
Research Analysis	🟡 Good

---

⚠ Known Limitations

• Sarcasm Misclassification (especially in Chinese)
• Mixed Sentiments lean toward Neutral
• Low recall for Question class due to limited data
• Bias toward Positive due to class imbalance
• Overconfidence in some multilingual predictions

---

🛠 Technical Info

Config	Value
Base Model	multilingual-e5-large
Params	~1.02B
Classes	4
Max Length	512
Training Time	~27 min

Data Summary:

• Training: 2,456 samples
• Validation: 273 samples
• Evaluation: 2,183 samples

---

📄 Citation

@misc{MultiSent-E5-Pro-2024,
  title={MultiSent-E5-Pro: Advanced Thai Sentiment Analysis},
  author={ZombitX64, Janutsaha K., Saengwichain C.},
  year={2024},
  url={https://huggingface.co/ZombitX64/MultiSent-E5-Pro},
  note={Hugging Face Model Card}
}

@article{wang2024multilingual,
  title={Multilingual E5 Text Embeddings: A Technical Report},
  author={Wang, Liang and Yang, Nan and Huang, Xiaolong and Yang, Linjun and Majumder, Rangan and Wei, Furu},
  journal={arXiv preprint arXiv:2402.05672},
  year={2024}
}

---

👨‍💼 Authors

Role	Name
Lead Dev	ZombitX64
Data Scientist	Krittanut Janutsaha
Engineer	Chanyut Saengwichain

---

😊 Feedback & Contributions

• 💬 Open Discussion
• 🐛 Report Issue
• 🌟 Star the repo if useful!

---

Last Updated: Dec 2024 | Version: 1.1 | Docs: v2.0