Classification

Overview

This report covers two classification pipelines in Underthesea: Text Classification and Sentiment Analysis. Both use underthesea_core.TextClassifier and support multiple domains.

---

Text Classification

The text classification module categorizes Vietnamese text into predefined categories. It supports a general news domain (10 categories) and a bank domain (14 categories), with an optional OpenAI prompt-based model.

Architecture

Text Classification Pipeline
├── Text Input
│   └── Raw Vietnamese text
├── Model Selection
│   ├── General Domain (default)
│   │   └── underthesea_core.TextClassifier
│   ├── Bank Domain (domain='bank')
│   │   └── underthesea_core.TextClassifier
│   └── Prompt Model (model='prompt')
│       └── OpenAI API
└── Output
    └── List of predicted categories

Models

Model	File	Description
General	sen-classifier-general-1.0.0-20260207.bin	Vietnamese news classification
Bank	sen-bank-1.0.0-20260207.bin	Banking feedback classification
Prompt	OpenAI API	LLM-based classification

Categories

General Domain (10 categories)

Category	Description
The thao	Sports
Kinh doanh	Business
Chinh tri Xa hoi	Politics & Society
Van hoa	Culture
Khoa hoc	Science
Phap luat	Law
Suc khoe	Health
Doi song	Lifestyle
The gioi	World
Vi tinh	Technology

Bank Domain (14 categories)

Category	Description
ACCOUNT	Account management
CARD	Card services
CUSTOMER_SUPPORT	Customer support
DISCOUNT	Discounts
INTEREST_RATE	Interest rates
INTERNET_BANKING	Internet banking
LOAN	Loan services
MONEY_TRANSFER	Money transfers
OTHER	Other topics
PAYMENT	Payments
PROMOTION	Promotions
SAVING	Savings
SECURITY	Security
TRADEMARK	Brand-related

Usage

from underthesea import classify

text = "HLV đầu tiên ở Premier League bị sa thải sau 4 vòng đấu"
category = classify(text)
# ['The thao']

# Bank domain
classify("Lãi suất tiết kiệm quá thấp", domain='bank')
# ['INTEREST_RATE']

# Access labels
classify.labels        # General domain labels
classify.bank.labels   # Bank domain labels

Parameters

Parameter	Type	Default	Description
X	str	required	Text to classify
domain	str	None	Domain — None for general, 'bank' for banking
model	str	None	Model type — None for default, 'prompt' for OpenAI

---

Sentiment Analysis

The sentiment analysis module analyzes the sentiment of Vietnamese text. The general domain returns positive/negative/neutral classification. The bank domain provides aspect-based sentiment analysis.

Architecture

Sentiment Analysis Pipeline
├── Text Input
│   └── Raw Vietnamese text
├── Model Selection
│   ├── General Domain (default)
│   │   └── underthesea_core.TextClassifier
│   │       └── 3-class: positive / negative / neutral
│   └── Bank Domain (domain='bank')
│       └── underthesea_core.TextClassifier
│           └── Aspect-based sentiment
└── Output
    ├── General: sentiment string
    └── Bank: list of aspect#sentiment pairs

Models

Model	File	Description
General	sen-sentiment-general-1.0.0-20260207.bin	3-class sentiment
Bank	sen-sentiment-bank-1.0.0-20260207.bin	Aspect-based sentiment

Sentiment Labels

General Domain

Label	Description
positive	Positive sentiment
negative	Negative sentiment
neutral	Neutral sentiment

Bank Domain — Aspects

Aspect	Description
INTEREST_RATE	Interest rate related
CUSTOMER_SUPPORT	Customer service quality
PRODUCT	Product/service quality
TRADEMARK	Brand perception

Usage

from underthesea import sentiment

text = "Sản phẩm hơi nhỏ so với tưởng tượng nhưng chất lượng tốt"
result = sentiment(text)
# 'positive'

# Bank domain
sentiment("Lãi suất quá cao, nhân viên hỗ trợ tốt", domain='bank')
# ['INTEREST_RATE#negative', 'CUSTOMER_SUPPORT#positive']

# Access labels
sentiment.labels        # General domain labels
sentiment.bank.labels   # Bank domain labels

Parameters

Parameter	Type	Default	Description
X	str	required	Text to analyze
domain	str	'general'	Domain — 'general' or 'bank'

---

References

1. Underthesea GitHub Repository