Sentiment Analysis

1from langchain_openai import ChatOpenAI
2from langchain_core.prompts import ChatPromptTemplate
3from pydantic import BaseModel, Field
4from typing import Literal
5
6llm = ChatOpenAI(model="gpt-4o-mini", temperature=0)
7
8class SentimentResult(BaseModel):
9    sentiment: Literal["positive", "negative", "neutral"]
10    confidence: float = Field(ge=0, le=1, description="0-1 confidence score")
11    reasoning: str = Field(description="Short explanation")
12
13sentiment_chain = (
14    ChatPromptTemplate.from_messages([
15        ("system", """Phan tich sentiment cua text.
16        Tra ve: sentiment (positive/negative/neutral), confidence (0-1), reasoning."""),
17        ("human", "{text}")
18    ])
19    | llm.with_structured_output(SentimentResult)
20)
21
22result = sentiment_chain.invoke({
23    "text": "San pham tot nhung gia hoi cao. Dich vu cham soc khach hang tuyet voi!"
24})
25print(f"Sentiment: {result.sentiment} ({result.confidence:.0%})")
26print(f"Reason: {result.reasoning}")

1from typing import List
2
3class AspectSentiment(BaseModel):
4    aspect: str = Field(description="Khia canh duoc danh gia")
5    sentiment: Literal["positive", "negative", "neutral", "mixed"]
6    mentions: List[str] = Field(description="Cau lien quan")
7
8class DetailedAnalysis(BaseModel):
9    overall: Literal["positive", "negative", "neutral", "mixed"]
10    aspects: List[AspectSentiment]
11    key_emotions: List[str]
12
13aspect_chain = (
14    ChatPromptTemplate.from_messages([
15        ("system", """Phan tich sentiment theo tung khia canh (aspect).
16        Identify: product quality, price, service, delivery, etc.
17        Cho moi aspect: sentiment + cau mention cu the."""),
18        ("human", "{review}")
19    ])
20    | llm.with_structured_output(DetailedAnalysis)
21)
22
23review = """
24Da dat hang tren Shopee. Giao hang nhanh, dong goi can than. 
25San pham dung nhu mo ta, chat luong tot. Tuy nhien gia dat hon 
26cac shop khac khoang 50k. Ho tro khach hang tra loi nhanh. 
27Se mua lai lan sau.
28"""
29
30analysis = aspect_chain.invoke({"review": review})
31for aspect in analysis.aspects:
32    print(f"  {aspect.aspect}: {aspect.sentiment}")

1import pandas as pd
2
3reviews = [
4    "San pham tuyet voi, se mua lai!",
5    "Chat luong kem, khong dung mo ta",
6    "Binh thuong, khong co gi dac biet",
7    "Giao hang cham nhung san pham tot",
8    "Gia qua dat so voi chat luong",
9]
10
11# Batch process
12results = sentiment_chain.batch(
13    [{"text": r} for r in reviews],
14    config={"max_concurrency": 5}
15)
16
17# To DataFrame
18df = pd.DataFrame([
19    {
20        "review": reviews[i],
21        "sentiment": r.sentiment,
22        "confidence": r.confidence
23    }
24    for i, r in enumerate(results)
25])
26
27print(df.to_string())
28print(f"\nDistribution:\n{df['sentiment'].value_counts()}")

1from langchain_core.prompts import FewShotChatMessagePromptTemplate
2
3examples = [
4    {"text": "Rat hai long voi dich vu", "label": "satisfied"},
5    {"text": "Se khong bao gio mua lai", "label": "churning"},
6    {"text": "Can cai thien them", "label": "constructive"},
7    {"text": "Toi se gioi thieu cho ban be", "label": "advocate"},
8]
9
10example_prompt = ChatPromptTemplate.from_messages([
11    ("human", "{text}"),
12    ("ai", "{label}")
13])
14
15few_shot = FewShotChatMessagePromptTemplate(
16    example_prompt=example_prompt,
17    examples=examples
18)
19
20custom_chain = (
21    ChatPromptTemplate.from_messages([
22        ("system", "Classify customer feedback: satisfied/churning/constructive/advocate"),
23        few_shot,
24        ("human", "{text}")
25    ])
26    | llm
27)

1class EmotionAnalysis(BaseModel):
2    primary_emotion: str
3    secondary_emotions: List[str]
4    intensity: Literal["low", "medium", "high"]
5    triggers: List[str]
6
7emotion_chain = (
8    ChatPromptTemplate.from_messages([
9        ("system", """Detect emotions in text.
10        Emotions: joy, anger, sadness, fear, surprise, disgust, trust, anticipation.
11        Identify primary emotion, secondary, intensity, and triggers."""),
12        ("human", "{text}")
13    ])
14    | llm.with_structured_output(EmotionAnalysis)
15)

1from collections import Counter
2from datetime import datetime
3
4class SentimentMonitor:
5    def __init__(self, chain):
6        self.chain = chain
7        self.history = []
8    
9    def analyze(self, text, source="unknown"):
10        result = self.chain.invoke({"text": text})
11        self.history.append({
12            "timestamp": datetime.now().isoformat(),
13            "text": text,
14            "sentiment": result.sentiment,
15            "confidence": result.confidence,
16            "source": source
17        })
18        return result
19    
20    def get_stats(self):
21        sentiments = [h["sentiment"] for h in self.history]
22        return {
23            "total": len(self.history),
24            "distribution": dict(Counter(sentiments)),
25            "avg_confidence": sum(h["confidence"] for h in self.history) / len(self.history)
26        }
27
28monitor = SentimentMonitor(sentiment_chain)