A newer version of this model is available: ZombitX64/Thai-sentiment-e5

MultiSent-E5

A Thai sentiment analysis model fine-tuned from multilingual-e5-large for classifying sentiment in Thai text into positive, negative, neutral, and question categories.

Model Details

Model Description

This model is a fine-tuned version of intfloat/multilingual-e5-large specifically trained for Thai sentiment analysis. It can classify Thai text into four sentiment categories: positive, negative, neutral, and question. The model demonstrates strong performance on Thai language sentiment classification tasks with high accuracy and good understanding of Thai linguistic nuances including sarcasm and implicit sentiment.

The model is particularly effective at:

Sarcasm Detection: Understanding when positive words are used in a negative context
Cultural Context: Recognizing Thai-specific expressions and cultural references
Implicit Sentiment: Detecting sentiment even when not explicitly stated
Colloquial Language: Processing informal Thai text from social media and conversations

Developed by: ZombitX64, Krittanut Janutsaha, Chanyut Saengwichain
Model type: Sequence Classification (Sentiment Analysis)
Language(s) (NLP): Thai (th) - Primary, with limited multilingual capability
License: Creative Commons Attribution-NonCommercial-NoDerivatives 4.0
Finetuned from model: intfloat/multilingual-e5-large

Model Sources

Repository: https://huggingface.co/ZombitX64/Thai-sentiment-e5
Base Model: https://huggingface.co/intfloat/multilingual-e5-large

Uses

Direct Use

This model can be directly used for sentiment analysis of Thai text. It's particularly useful for:

Social Media Analysis: Monitoring sentiment on Thai social platforms like Twitter, Facebook, and Pantip
Customer Feedback Analysis: Processing reviews and feedback in Thai for e-commerce and services
Product Review Classification: Automatically categorizing product reviews by sentiment
Opinion Mining: Extracting sentiment from Thai news articles, blogs, and forums
Customer Service: Categorizing customer inquiries and complaints by sentiment and intent

Downstream Use

The model can be integrated into larger applications such as:

Customer Service Chatbots: Automatically routing messages based on sentiment
Social Media Analytics Platforms: Real-time sentiment monitoring dashboards
E-commerce Review Systems: Automated review scoring and categorization
Content Moderation Systems: Identifying potentially problematic content
Market Research Tools: Analyzing consumer sentiment towards brands or products
News Analysis Systems: Tracking public opinion on political or social issues

Out-of-Scope Use

This model should not be used for:

Question Classification: The model has poor performance on question detection due to insufficient training data. Questions are often misclassified with moderate confidence (50-60%). Use a dedicated question classification model instead.
Mixed Sentiment Analysis: Complex texts with both positive and negative elements may be misclassified or produce low confidence scores. Consider using aspect-based sentiment analysis for such cases.
Non-Thai Languages: While it has some multilingual capability, accuracy is significantly lower for languages other than Thai
Fine-grained Emotion Detection: The model only classifies into 4 broad categories, not specific emotions like anger, joy, fear, etc.
Clinical Applications: Should not be used for mental health diagnosis or psychological assessment without proper validation
High-stakes Decision Making: Avoid using for critical decisions affecting individuals without human oversight, especially for predictions with confidence < 60%
Legal or Financial Decisions: The model's predictions should not be the sole basis for legal or financial determinations

🌐 Multilingual Sentiment Capability

The MultiSent-E5 model has been developed as an extension of the intfloat/multilingual-e5-large base model, which is a multilingual embedding model supporting over 50 languages. This gives the model some capability for sentiment prediction in multiple languages beyond Thai.

Language Support Details

Primary Language: Thai - The model has been fine-tuned specifically for Thai and performs best with Thai text
Secondary Languages: The model can provide basic sentiment analysis for other languages such as English, Chinese, Japanese, Indonesian, and other languages supported by the base multilingual model
Performance Considerations: Accuracy for non-Thai languages may be significantly lower and results may be less reliable, depending on the similarity of linguistic structures and vocabulary to Thai

Multilingual Performance Expectations

Language Family	Expected Performance	Use Case Recommendation
Thai	Excellent (99%+ accuracy)	Primary use case
Southeast Asian (Indonesian, Malay, Vietnamese)	Good (70-85% accuracy)	Limited use with validation
East Asian (Chinese, Japanese, Korean)	Moderate (60-75% accuracy)	Experimental use only
European Languages	Moderate (55-70% accuracy)	Not recommended
Other Languages	Poor (40-60% accuracy)	Not recommended

Recommendations for Multilingual Use

Primary Recommendation: Use this model primarily for Thai sentiment analysis where it excels
Secondary Use: For other languages, consider using language-specific models for maximum accuracy
Validation Required: Always validate results when using with non-Thai languages
Experimental Use: Multilingual capability can be useful for initial exploration or when Thai-specific models are unavailable

This multilingual capability makes the model suitable for basic multilingual sentiment classification tasks while maintaining excellent performance for Thai text analysis.

How to Get Started with the Model

Basic Usage

from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch

# Load the model and tokenizer
model_name = "ZombitX64/MultiSent-E5"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForSequenceClassification.from_pretrained(model_name)

# Example Thai text
text = "ผลิตภัณฑ์นี้ดีมาก ใช้งานง่าย"  # "This product is very good, easy to use"

# Tokenize and predict
inputs = tokenizer(text, return_tensors="pt", truncation=True, padding=True)

with torch.no_grad():
    outputs = model(**inputs)
    predictions = torch.nn.functional.softmax(outputs.logits, dim=-1)
    predicted_class = torch.argmax(predictions, dim=-1)

# Label mapping: 0=Question, 1=Negative, 2=Neutral, 3=Positive
labels = ["Question", "Negative", "Neutral", "Positive"]
predicted_label = labels[predicted_class.item()]
confidence = predictions[0][predicted_class.item()].item()

print(f"Text: {text}")
print(f"Predicted sentiment: {predicted_label} ({confidence:.2%})")

Batch Processing

# List of texts to analyze (multilingual examples)
texts = [
    "ผลิตภัณฑ์นี้ดีมาก ใช้งานง่าย",               # Thai: "This product is very good, easy to use"
    "The service was terrible and disappointing",  # English
    "商品质量还可以",                               # Chinese: "Product quality is okay"
    "บริการแย่มาก ไม่ประทับใจเลย",                 # Thai: "Service is terrible, not impressed at all"
    "Ce produit est excellent",                     # French: "This product is excellent"
]

print("Predicting sentiment for multiple texts:")
for text in texts:
    inputs = tokenizer(text, return_tensors="pt", truncation=True, padding=True)

    with torch.no_grad():
        outputs = model(**inputs)
        predictions = torch.nn.functional.softmax(outputs.logits, dim=-1)
        predicted_class = torch.argmax(predictions, dim=-1)

    predicted_label = labels[predicted_class.item()]
    confidence = predictions[0][predicted_class.item()].item()

    print(f"\nText: \"{text}\"")
    print(f"Predicted sentiment: {predicted_label} ({confidence:.2%})")

Pipeline Usage

from transformers import pipeline

# Create a sentiment analysis pipeline
classifier = pipeline("text-classification", 
                     model="ZombitX64/MultiSent-E5",
                     tokenizer="ZombitX64/MultiSent-E5")

# Analyze sentiment
texts = [
    "วันนี้อากาศดีจังเลย",  # "The weather is so nice today"
    "แย่ที่สุดเท่าที่เคยเจอมา"  # "The worst I've ever encountered"
]

results = classifier(texts)
for text, result in zip(texts, results):
    print(f"Text: {text}")
    print(f"Sentiment: {result['label']} (Score: {result['score']:.4f})")

Training Details

Training Data

The model was trained on a carefully curated Thai sentiment dataset with the following characteristics:

Total samples: 2,730 (2,729 after data cleaning and filtering)
Data Distribution:
- Question samples: Minimal representation (specific count not provided)
- Negative samples: 102 (3.7% of dataset)
- Neutral samples: 317 (11.6% of dataset)
- Positive samples: 2,310 (84.7% of dataset)

Data Split Strategy:

Training set: 2,456 samples (90% of total data)
Validation set: 273 samples (10% of total data)

Data Quality and Preprocessing:

Data was manually reviewed and cleaned to ensure quality
Duplicate entries were removed
Text was normalized for consistent formatting
Class imbalance was noted but maintained to reflect real-world distribution

Training Procedure

The model was fine-tuned using state-of-the-art techniques with careful hyperparameter optimization:

Training Hyperparameters

Base Model: intfloat/multilingual-e5-large (1.02B parameters)
Model Architecture: XLMRobertaForSequenceClassification
Training Epochs: 5 (with early stopping monitoring)
Total Training Steps: 770
Batch Size: 8 (effective batch size with gradient accumulation)
Learning Rate: 2e-5 with linear warmup and decay
Weight Decay: 0.01
Warmup Steps: 77 (10% of total steps)
Max Sequence Length: 512 tokens
Optimization: AdamW optimizer
Training Runtime: 1,633.3 seconds (~27 minutes)
Training Samples per Second: 7.519
Training Steps per Second: 0.471

Training Infrastructure

Hardware: GPU-accelerated training (specific GPU not specified)
Framework: Hugging Face Transformers 4.x
Distributed Training: Single GPU setup
Memory Optimization: Gradient checkpointing enabled

Training Results

The model showed excellent convergence with minimal overfitting:

Epoch	Training Loss	Validation Loss	Accuracy	Notes
1	0.0812	0.0699	98.53%	Strong initial performance
2	0.0053	0.0527	99.27%	Rapid improvement
3	0.0041	0.0350	99.63%	Near-optimal performance
4	0.0002	0.0384	99.63%	Slight validation loss increase
5	0.0002	0.0410	99.63%	Stable performance

Training Observations:

Very low training loss achieved by epoch 3
Validation loss remained stable, indicating minimal overfitting
Accuracy plateaued at 99.63% from epoch 3 onwards
Early convergence suggests effective transfer learning from the base model

Evaluation

Model Comparison Metrics

Model Comparison Metrics (Scatter)

🥇อันดับ	ชื่อโมเดล	Accuracy (%)	หมายเหตุ
1	MultiSent-E5	84.88	★ โมเดลที่แม่นยำที่สุด
2	ZombitX64/sentiment-103	68.60	รองชนะเลิศ
3	Thai-sentiment-e5	67.44	ดีเด่นด้านความเข้าใจภาษาไทย
4	xlm-roberta	34.30	multilingual baseline
5	mMiniLM	28.49	ขนาดเล็ก ใช้ทรัพยากรน้อย
6	wangchan-sentiment-thai-text-model	25.58	ภาษาไทยโดยเฉพาะ
7	WangchanBERTa-finetuned-sentiment	25.00	fine-tuned Thai BERT
8	wangchanberta	22.67	Thai BERT base
9	Thaweewat/hyperopt-sentiment	21.51	ปรับจูนด้วย hyperopt
10	sentiment-thai-text-model	17.44	baseline keyword model
11	e5-base	16.86	multilingual encoder
12	twitter-xlm-roberta-base-sentiment	7.56	fine-tuned บน Twitter (อังกฤษ)

============================================================ Evaluating Model: MultiSent-E5

Accuracy: 0.849 F1-Macro: 0.839 F1-Weighted: 0.850

=== ERROR ANALYSIS FOR MultiSent-E5 === Total Errors: 26 / 172 (15.1%)

Error Types: error_type negative -> positive 7 question -> neutral 7 neutral -> positive 3 negative -> neutral 3 question -> positive 2 positive -> neutral 2 neutral -> negative 1 neutral -> question 1 Name: count, dtype: int64

Low Confidence Errors (< 60%): 4 High Confidence Errors (> 80%): 19

=== ERROR EXAMPLES ===

negative -> positive: Text: 'สุดยอดไปเลย! เธอใช้เวลาทั้งวันทำงานชิ้นนี้ออกมาได้แค่นี้เองเหรอเนี่ย!' Confidence: 0.517 Text: 'ไอเดียสร้างสรรค์มาก! ไม่มีใครคิดจะเสนออะไรที่ไม่มีทางเป็นไปได้แบบนี้หรอก' Confidence: 1.000 Text: 'ไอเดียสร้างสรรค์มาก! ไม่มีใครคิดจะเสนออะไรที่ไม่มีทางเป็นไปได้แบบนี้หรอก' Confidence: 1.000

question -> neutral: Text: 'คุณคิดว่าอย่างไรกับเรื่องนี้' Confidence: 0.999 Text: 'How was your day today?' Confidence: 1.000 Text: '你觉得怎么样？' Confidence: 0.999

neutral -> positive: Text: 'ก็แข็งแรงอยู่นะ' Confidence: 0.727 Text: 'ก็แข็งแรงอยู่นะ' Confidence: 0.727 Text: 'บรรยากาศดีมาก เหมาะกับการนั่งเงียบๆ คนเดียว' Confidence: 0.723

negative -> neutral: Text: 'Good day. Unfortunately, I had to walk 10 kilometers from home to school, and now I'm feeling quite ...' Confidence: 0.970 Text: 'Good day. Unfortunately, I had to walk 10 kilometers from home to school, and now I'm feeling quite ...' Confidence: 0.970 Text: 'ส่งของไวมาก...ถ้านับวันเป็นเดือน' Confidence: 0.999

question -> positive: Text: 'ลำไยอร่อยดีสดมากและลูกใหญ่ด้วยแต่เน่าไปครึ่งนึงมั้ย' Confidence: 0.550 Text: 'ลำไยอร่อยดีสดมากและลูกใหญ่ด้วยแต่เน่าไปครึ่งนึงมั้ย' Confidence: 0.550

=== LOW CONFIDENCE PREDICTIONS === Total Low Confidence: 7 (4.1%)

Low Confidence Examples: 'ลำไยอร่อยดีสดมากและลูกใหญ่ด้วยแต่เน่าไปครึ่งนึงมั้ย' Predicted: positive, Confidence: 0.550 True: question, Correct: False

'ลำไยอร่อยดีสดมากและลูกใหญ่ด้วยแต่เน่าไปครึ่งนึงรึเปล่า' Predicted: question, Confidence: 0.521 True: question, Correct: True

'สุดยอดไปเลย! เธอใช้เวลาทั้งวันทำงานชิ้นนี้ออกมาได้แค่นี้เองเหรอเนี่ย!' Predicted: positive, Confidence: 0.517 True: negative, Correct: False

'เกือบดีแล้วล่ะ เหลือแค่ดีจริงๆ นิดเดียว' Predicted: neutral, Confidence: 0.546 True: neutral, Correct: True

'ลำไยอร่อยดีสดมากและลูกใหญ่ด้วยแต่เน่าไปครึ่งนึงมั้ย' Predicted: positive, Confidence: 0.550 True: question, Correct: False

📊 สรุปผลการประเมินโมเดล: MultiSent-E5

Metric	ค่า (Value)
Accuracy	84.9%
F1 Macro	83.9%
F1 Weighted	85.0%
จำนวนตัวอย่างทั้งหมด	172
จำนวนข้อผิดพลาด (Error)	26
เปอร์เซ็นต์ความผิดพลาด	15.1%
Low Confidence Errors (<60%)	4
High Confidence Errors (>80%)	19

🧩 ประเภทความผิดพลาด (Error Types)

ผิดจาก (True Label)	เป็น (Predicted Label)	จำนวนครั้ง (Count)
negative	positive	7
question	neutral	7
neutral	positive	3
negative	neutral	3
question	positive	2
positive	neutral	2
neutral	negative	1
neutral	question	1

🔍 ตัวอย่าง Error ที่น่าสนใจ

1. ประชด/เสียดสี ที่ผิดเป็น Positive

ข้อความ	ทำนาย	จริง	Confidence
สุดยอดไปเลย! เธอใช้เวลาทั้งวันทำงานชิ้นนี้ออกมาได้แค่นี้เองเหรอเนี่ย!	positive	negative	0.517
ไอเดียสร้างสรรค์มาก! ไม่มีใครคิดจะเสนออะไรที่ไม่มีทางเป็นไปได้แบบนี้หรอก	positive	negative	1.000

2. คำถาม ผิดเป็น Neutral

ข้อความ	ทำนาย	จริง	Confidence
คุณคิดว่าอย่างไรกับเรื่องนี้	neutral	question	0.999
How was your day today?	neutral	question	1.000
你觉得怎么样？	neutral	question	0.999

3. ประโยคคลุมเครือที่ Low Confidence

ข้อความ	ทำนาย	จริง	Confidence
ลำไยอร่อยดีสดมากและลูกใหญ่ด้วยแต่เน่าไปครึ่งนึงมั้ย	positive	question	0.550
เกือบดีแล้วล่ะ เหลือแค่ดีจริงๆ นิดเดียว	neutral	neutral	0.546

Testing Data, Factors & Metrics

Testing Data

The model was evaluated on a carefully selected validation set with the following characteristics:

Total Validation Samples: 273
Selection Method: Stratified random sampling to maintain class distribution
Data Quality: Manually verified and cleaned validation samples
Evaluation Period: Final model checkpoint from epoch 5

Evaluation Metrics

The model was comprehensively evaluated using multiple metrics:

Primary Metrics:
- Accuracy: Overall classification accuracy across all classes
- F1-Score: Both macro and weighted averages
Secondary Metrics:
- Precision: Per-class and overall precision scores
- Recall: Per-class and overall recall scores
- Support: Number of samples per class in validation set

Results

Final Test Results

Per-Class Performance:

Class	Precision	Recall	F1-Score	Support	Performance Notes
Question	N/A	N/A	N/A	0	No question samples in validation set
Negative	1.00	1.00	1.00	231	Perfect classification
Neutral	1.00	0.90	0.95	10	1 misclassification due to small sample size
Positive	1.00	1.00	1.00	32	Perfect classification

Overall Performance Summary:

Metric	Value	Interpretation
Overall Accuracy	100% (273/273)	Exceptional performance
Macro Average F1	0.98	Excellent across all represented classes
Weighted Average F1	1.00	Perfect when weighted by class frequency
Total Correct Predictions	272/273	Only 1 misclassification

Detailed Confusion Matrix Results

Classification Breakdown:

Negative Class: 231/231 correctly classified (100% accuracy)
Neutral Class: 9/10 correctly classified (90% accuracy)
- 1 neutral sample misclassified (likely as positive due to ambiguous language)
Positive Class: 32/32 correctly classified (100% accuracy)
Question Class: Not present in validation set

Model Capabilities

Demonstrated Strengths

The model shows exceptional capability in understanding various aspects of Thai sentiment:

1. Straightforward Sentiment Classification:

Clear positive expressions: "วันนี้อากาศดีจังเลย" (The weather is so nice today) → Positive (99.96%)
Clear negative expressions: "แย่ที่สุดเท่าที่เคยเจอมา" (The worst I've ever encountered) → Negative (99.99%)
Neutral expressions: "ก็งั้นๆ แหละ ไม่มีอะไรพิเศษ" (It's just okay, nothing special) → Neutral (99.70%)

2. Advanced Linguistic Understanding:

Sarcasm Detection:

"เก่งจังเลยนะ ทำผิดซ้ำได้เหมือนเดิมเป๊ะเลย" (So talented! You can make the same mistake repeatedly) → Negative (99.99%)
The model correctly identifies that "เก่งจัง" (so talented) is used sarcastically

Implicit Criticism:

"ไอเดียสร้างสรรค์มาก! ไม่มีใครคิดจะเสนออะไรที่ไม่มีทางเป็นไปได้แบบนี้หรอก" (Very creative idea! No one would think to propose something this impossible) → Negative (99.43%)
Successfully detects negative sentiment despite seemingly positive words

3. Cultural Context Understanding:

Thai-specific expressions and idioms
Formal vs. informal language registers
Regional variations in expression

Performance Analysis by Text Type

Text Type	Accuracy	Confidence Range	Notes
Direct statements	99-100%	90-100%	Excellent performance
Sarcastic content	95-99%	85-99%	Very good sarcasm detection (e.g., "เก่งจังเลยนะ ทำผิดซ้ำได้เหมือนเดิมเป๊ะเลย" → 99.98% negative)
Implicit sentiment	90-95%	80-95%	Good at reading between the lines
Mixed sentiment	60-75%	50-60%	Struggles with texts containing both positive and negative aspects
Question-like text	40-60%	50-60%	Poor question detection, often classified as other categories
Star ratings	95-100%	99%+	Excellent (e.g., "ให้5ดาวเลย" → 99.98% positive, "ให้1ดาวเลย" → 99.49% negative)
Formal language	98-100%	85-100%	Strong performance on formal text
Colloquial language	95-99%	80-95%	Handles informal text well

Real-World Performance Issues

Low Confidence Predictions (< 60%): Based on empirical testing, these text types frequently produce low confidence:

Mixed Sentiment Examples:
- "ลำไยอร่อยดีสดมากและลูกใหญ่ด้วยแต่เน่าไปครึ่งนึ..." → Positive (55.0%) or Question (52.1%)
- "เกือบดีแล้วล่ะ เหลือแค่ดีจริงๆ นิดเดียว" → Neutral (54.6%)
Ambiguous Praise with Criticism:
- "สุดยอดไปเลย! เธอใช้เวลาทั้งวันทำงานชิ้นนี้ออกม..." → Positive (51.7%)

High Confidence Predictions (> 99%): The model excels at:

Clear sarcasm: "เก่งจังเลยนะ ทำผิดซ้ำได้เหมือนเดิมเป๊ะเลย" → Negative (99.98%)
Obvious negative sentiment: "ไม่ให้ดาวเลย" → Negative (99.94%)
Simple positive expressions: "ให้5ดาวเลย" → Positive (99.98%)

Known Limitations

1. Question Class Performance Issues:

Insufficient Training Data: The question class has minimal representation in the training dataset
Low Confidence Predictions: Question classification often results in confidence scores below 60%
Misclassification: Questions are frequently classified as positive, negative, or neutral instead
Example Issue: "ลำไยอร่อยดีสดมากและลูกใหญ่ด้วยแต่เน่าไปครึ่งนึ..." (Longans are delicious and fresh, big fruits too, but half are rotten...) → Classified as neutral (97.7% confidence) instead of recognizing mixed sentiment

2. Mixed Sentiment Challenges:

Complex Sentiment: Texts with both positive and negative aspects may be misclassified
Moderate Confidence: Mixed sentiment often results in lower confidence scores (50-60%)
Example: Product reviews mentioning both good and bad aspects tend toward neutral classification

3. Class Imbalance Effects:

Model may be biased toward positive classifications due to training data imbalance (84.7% positive samples)
Neutral class performance slightly lower due to limited training examples (11.6% of data)
Negative class well-represented but still only 3.7% of training data

4. Low Confidence Predictions:

Predictions with confidence < 60% should be treated with caution
Common in mixed sentiment, ambiguous language, or question-like texts
Recommend implementing confidence thresholding for production use

Environmental Impact

Carbon Footprint Considerations

Training Emissions: Specific carbon emission data not available
Efficiency Benefits: Model was fine-tuned from a pre-trained multilingual model, significantly reducing computational cost compared to training from scratch
Resource Usage: Relatively efficient training with only 27 minutes of GPU time required
Deployment Efficiency: Model can be deployed efficiently for inference with standard hardware

Sustainable AI Practices

Transfer Learning: Leveraged existing multilingual model to reduce training requirements
Efficient Architecture: Uses proven transformer architecture optimized for efficiency
Reusability: Single model can handle multiple languages, reducing need for separate models

Technical Specifications

Model Architecture and Objective

Architecture: XLMRobertaForSequenceClassification
Base Model: intfloat/multilingual-e5-large
Model Parameters: ~1.02 billion parameters
Classification Head: Linear layer with 4 output classes
Task: Multi-class text classification (4 classes: Question, Negative, Neutral, Positive)
Objective Function: Cross-entropy loss minimization
Activation Function: Softmax for final predictions
Input Processing: Tokenization with XLM-RoBERTa tokenizer
Maximum Input Length: 512 tokens

Performance Characteristics

Inference Speed: Fast inference suitable for real-time applications
Memory Requirements: Standard transformer model memory usage
Scalability: Can handle batch processing efficiently
Hardware Requirements: Compatible with CPU and GPU inference

Integration Specifications

Framework Compatibility:
- Hugging Face Transformers
- PyTorch
- ONNX (convertible)
- TensorFlow (via conversion)
API Support: Compatible with Hugging Face Inference API
Deployment Options:
- Cloud deployment (AWS, GCP, Azure)
- Edge deployment (with optimization)
- Local deployment

Compute Infrastructure

Hardware Requirements

Training Infrastructure

GPU: Modern NVIDIA GPU with sufficient VRAM (16GB+ recommended)
Memory: 32GB+ RAM recommended for training
Storage: SSD storage for fast data loading

Inference Infrastructure

Minimum Requirements:
- CPU: Modern multi-core processor
- RAM: 8GB+ for batch processing
- Storage: 2GB for model files
Recommended for Production:
- GPU: NVIDIA T4 or better
- RAM: 16GB+
- Multiple instances for load balancing

Software Dependencies

Core Requirements

Python: 3.8+
PyTorch: 1.9+
Transformers: 4.15+
NumPy: 1.21+
Tokenizers: 0.11+

Optional Dependencies

ONNX: For model conversion and optimization
TensorRT: For NVIDIA GPU optimization
Gradio/Streamlit: For web interface development

Usage Examples and Best Practices

Best Practices for Implementation

Text Preprocessing

def preprocess_thai_text(text):
    """
    Recommended preprocessing for Thai text
    """
    # Remove excessive whitespace
    text = ' '.join(text.split())
    
    # Handle common Thai punctuation
    text = text.replace('...', ' ')
    text = text.replace('!!', '!')
    
    # Normalize quotation marks
    text = text.replace('"', '"').replace('"', '"')
    
    return text.strip()

Confidence Thresholding

def classify_with_confidence(text, threshold=0.6):
    """
    Classification with confidence thresholding
    Recommended threshold: 0.6 based on empirical testing
    """
    inputs = tokenizer(text, return_tensors="pt", truncation=True, padding=True)
    
    with torch.no_grad():
        outputs = model(**inputs)
        predictions = torch.nn.functional.softmax(outputs.logits, dim=-1)
        confidence = torch.max(predictions).item()
        predicted_class = torch.argmax(predictions, dim=-1).item()
    
    if confidence >= threshold:
        return labels[predicted_class], confidence
    else:
        return "Low Confidence - Manual Review Needed", confidence

# Enhanced classification with question detection fallback
def enhanced_classify(text, confidence_threshold=0.6):
    """
    Enhanced classification with special handling for low confidence
    and potential question detection
    """
    sentiment, confidence = classify_with_confidence(text, confidence_threshold)
    
    # Special handling for low confidence predictions
    if confidence < confidence_threshold:
        # Simple question detection fallback
        question_indicators = ['?', 'ไหม', 'หรือ', 'ครับ', 'คะ', 'มั้ย']
        if any(indicator in text for indicator in question_indicators):
            return "Question (Detected by Rules)", confidence
        else:
            return f"Uncertain ({sentiment})", confidence
    
    return sentiment, confidence

# Example usage with test cases
test_texts = [
    "ลำไยอร่อยดีสดมากและลูกใหญ่ด้วยแต่เน่าไปครึ่งนึ...",  # Mixed sentiment
    "สุดยอดไปเลย! เธอใช้เวลาทั้งวันทำงานชิ้นนี้ออกม...",     # Low confidence positive
    "เก่งจังเลยนะ ทำผิดซ้ำได้เหมือนเดิมเป๊ะเลย",            # High confidence sarcasm
]

for text in test_texts:
    result, conf = enhanced_classify(text)
    print(f"Text: {text[:50]}...")
    print(f"Result: {result} (Confidence: {conf:.1%})")
    print()

Production Deployment Example

from fastapi import FastAPI
from pydantic import BaseModel
import logging

app = FastAPI()

class SentimentRequest(BaseModel):
    text: str
    
class SentimentResponse(BaseModel):
    sentiment: str
    confidence: float
    warning: str = None

def classify_with_warnings(text, confidence_threshold=0.6):
    """
    Production-ready classification with warnings for low confidence
    """
    inputs = tokenizer(text, return_tensors="pt", truncation=True, padding=True)
    
    with torch.no_grad():
        outputs = model(**inputs)
        predictions = torch.nn.functional.softmax(outputs.logits, dim=-1)
        confidence = torch.max(predictions).item()
        predicted_class = torch.argmax(predictions, dim=-1).item()
    
    sentiment = labels[predicted_class]
    warning = None
    
    # Add warnings based on empirical testing
    if confidence < confidence_threshold:
        warning = "Low confidence prediction - manual review recommended"
    
    if predicted_class == 0:  # Question class
        warning = "Question classification has known accuracy issues - consider manual review"
    
    # Detect potential mixed sentiment
    if confidence < 0.7 and any(pos_word in text for pos_word in ['ดี', 'อร่อย', 'สวย']) and any(neg_word in text for neg_word in ['แย่', 'เน่า', 'แต่']):
        warning = "Possible mixed sentiment detected - consider aspect-based analysis"
    
    return sentiment, confidence, warning

@app.post("/analyze-sentiment", response_model=SentimentResponse)
async def analyze_sentiment(request: SentimentRequest):
    try:
        # Preprocess text
        text = preprocess_thai_text(request.text)
        
        # Get prediction with warnings
        sentiment, confidence, warning = classify_with_warnings(text)
        
        # Log low confidence predictions for monitoring
        if confidence < 0.6:
            logging.warning(f"Low confidence prediction: {text[:50]}... -> {sentiment} ({confidence:.3f})")
        
        return SentimentResponse(
            sentiment=sentiment,
            confidence=confidence,
            warning=warning
        )
    
    except Exception as e:
        logging.error(f"Error processing text: {str(e)}")
        return SentimentResponse(
            sentiment="Error",
            confidence=0.0,
            warning="Processing error occurred"
        )

# Batch processing endpoint for efficiency
@app.post("/analyze-batch")
async def analyze_batch(texts: list[str]):
    """
    Batch processing for multiple texts
    """
    results = []
    for text in texts:
        sentiment, confidence, warning = classify_with_warnings(text)
        results.append({
            "text": text[:100] + "..." if len(text) > 100 else text,
            "sentiment": sentiment,
            "confidence": confidence,
            "warning": warning
        })
    return {"results": results}

Citation

Academic Citation

BibTeX:

@misc{MultiSent-E5,
  title={Thai-sentiment-e5: A Fine-tuned Multilingual Sentiment Analysis Model for Thai Text Classification},
  author={ZombitX64 and Janutsaha, Krittanut and Saengwichain, Chanyut},
  year={2024},
  url={https://huggingface.co/ZombitX64/MultiSent-E5},
  note={Hugging Face Model Repository}
}

APA Style: ZombitX64, Janutsaha, K., & Saengwichain, C. (2024). MultiSent-E5: A Fine-tuned Multilingual Sentiment Analysis Model for Thai Text Classification. Hugging Face. https://huggingface.co/ZombitX64/MultiSent-E5

IEEE Style: ZombitX64, K. Janutsaha, and C. Saengwichain, "MultiSent-E5: A Fine-tuned Multilingual Sentiment Analysis Model for Thai Text Classification," Hugging Face, 2024. [Online]. Available: https://huggingface.co/ZombitX64/MultiSent-E5

Usage in Publications

If you use this model in your research or applications, please cite both this model and the base model:

@misc{wang2022text,
  title={Text Embeddings by Weakly-Supervised Contrastive Pre-training}, 
  author={Liang Wang and Nan Yang and Xiaolong Huang and Binxing Jiao and Linjun Yang and Daxin Jiang and Rangan Majumder and Furu Wei},
  year={2022},
  eprint={2212.03533},
  archivePrefix={arXiv},
  primaryClass={cs.CL}
}

@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}
}

Model Card Authors

Primary Contributors:

ZombitX64 - Lead developer and model architect
Krittanut Janutsaha - Data curation and evaluation
Chanyut Saengwichain - Model optimization and documentation

Model Card Contact

Support and Issues

For questions, issues, or contributions regarding this model, please use the following channels:

Primary Contact: Hugging Face model repository issues and discussions
Repository: https://huggingface.co/ZombitX64/MultiSent-E5
Community: Hugging Face community forums for general questions

Collaboration Opportunities

We welcome collaboration on:

Improving the model's performance
Expanding to other Southeast Asian languages
Creating domain-specific variants
Integration into larger NLP systems

Feedback and Improvements

Your feedback helps improve this model. Please report:

Performance issues on specific text types
Suggestions for additional evaluation metrics
Use cases where the model performs unexpectedly
Ideas for model enhancements

Last updated: 2024 Model version: 1.0 Documentation version: 2.0

Downloads last month: 44

Safetensors

Model size

560M params

Tensor type

F32

Inference Providers NEW

Text Classification

This model isn't deployed by any Inference Provider. 🙋 1 Ask for provider support

Model tree for ZombitX64/MultiSent-E5

Base model

intfloat/multilingual-e5-large

Finetuned

(110)

this model

MultiSent-E5

Model Details

Model Description

Model Sources

Uses

Direct Use

Downstream Use

Out-of-Scope Use

🌐 Multilingual Sentiment Capability

Language Support Details

Multilingual Performance Expectations

Recommendations for Multilingual Use

How to Get Started with the Model

Basic Usage

Batch Processing

Pipeline Usage

Training Details

Training Data

Training Procedure

Training Hyperparameters

Training Infrastructure

Training Results

Evaluation

Model Comparison Metrics

Model Comparison Metrics (Scatter)

============================================================ Evaluating Model: MultiSent-E5

📊 สรุปผลการประเมินโมเดล: MultiSent-E5

🧩 ประเภทความผิดพลาด (Error Types)

🔍 ตัวอย่าง Error ที่น่าสนใจ

1. ประชด/เสียดสี ที่ผิดเป็น Positive

2. คำถาม ผิดเป็น Neutral

3. ประโยคคลุมเครือที่ Low Confidence

Testing Data, Factors & Metrics

Testing Data

Evaluation Metrics

Results

Final Test Results

Detailed Confusion Matrix Results

Model Capabilities

Demonstrated Strengths

Performance Analysis by Text Type

Real-World Performance Issues

Known Limitations

Environmental Impact

Carbon Footprint Considerations

Sustainable AI Practices

Technical Specifications

Model Architecture and Objective

Performance Characteristics

Integration Specifications

Compute Infrastructure

Hardware Requirements

Training Infrastructure

Inference Infrastructure

Software Dependencies

Core Requirements

Optional Dependencies

Usage Examples and Best Practices

Best Practices for Implementation

Text Preprocessing

Confidence Thresholding

Production Deployment Example

Citation

Academic Citation

Usage in Publications

Model Card Authors

Model Card Contact

Support and Issues

Collaboration Opportunities

Feedback and Improvements

Model tree for ZombitX64/MultiSent-E5

Space using ZombitX64/MultiSent-E5 1