Pythonzero_shot.py
Copy
1
import os
2
import openai
3
 
4
client = openai.OpenAI(api_key=os.environ["OPENAI_API_KEY"])
5
 
6
# ── Zero-shot prompt anatomy ─────────────────────────────────────────
7
# 1. Role (who the model is)
8
# 2. Task (what to do)
9
# 3. Format (how to respond)
10
# 4. Constraints (what not to do)
11
 
12
def prompt_v1_vague(text: str) -> str:
13
    """Bad: vague, no structure."""
14
    return f"Classify this: {text}"
15
 
16
def prompt_v2_structured(text: str) -> str:
17
    """Better: role + task + format."""
18
    return f"""You are an expert sentiment analyst.
19
Task: Classify the sentiment of the customer review below.
20
Output format: A single word — exactly one of: positive, negative, neutral
21
Review: {text}
22
Sentiment:"""
23
 
24
def prompt_v3_detailed(text: str) -> str:
25
    """Best: full structure with constraints."""
26
    return f"""You are a senior NLP specialist with expertise in customer sentiment analysis.
27
 
28
Task: Analyze the sentiment of the customer review below, considering:
29
- Overall emotional tone
30
- Presence of complaints vs praise
31
- Implicit vs explicit sentiment
32
 
33
Output requirements:
34
- Return ONLY a JSON object
35
- Use exactly this schema: {{"sentiment": "positive|negative|neutral", "confidence": 0.0-1.0, "key_phrase": "<most revealing phrase>"}}
36
 
37
Customer Review: {text}"""
38
 
39
# ── Test on sample reviews ───────────────────────────────────────────
40
REVIEWS = [
41
    "The product quality is excellent but shipping took 3 weeks.",
42
    "Complete waste of money. Broke after one day.",
43
    "Delivered on time and does exactly what it says.",
44
]
45
 
46
def call(prompt: str) -> str:
47
    response = client.chat.completions.create(
48
        model="gpt-4o-mini",
49
        messages=[{"role": "user", "content": prompt}],
50
        max_tokens=80, temperature=0,
51
    )
52
    return response.choices[0].message.content.strip()
53
 
54
import json
55
 
56
print("=== Zero-Shot Prompt Comparison ===")
57
for review in REVIEWS[:2]:
58
    print(f"\nReview: {review[:60]}...")
59
    
60
    r1 = call(prompt_v1_vague(review))
61
    print(f"  V1 (vague):      {r1[:50]}")
62
    
63
    r2 = call(prompt_v2_structured(review))
64
    print(f"  V2 (structured): {r2[:50]}")
65
    
66
    r3 = call(prompt_v3_detailed(review))
67
    try:
68
        parsed = json.loads(r3)
69
        print(f"  V3 (detailed):   {parsed['sentiment']} ({parsed['confidence']:.0%}) — '{parsed['key_phrase']}'")
70
    except json.JSONDecodeError:
71
        print(f"  V3 (detailed):   {r3[:50]}")
72
 
73
# ── Role impact experiment ───────────────────────────────────────────
74
def with_role(role: str, text: str) -> str:
75
    return f"{role}\n\nSummarize in one sentence: {text}"
76
 
77
TEXT = "Transformer models use self-attention mechanisms to process sequences in parallel, enabling massive parallelization during training and achieving state-of-the-art results across NLP tasks."
78
 
79
print("\n=== Role Impact ===")
80
roles = [
81
    "Answer the following:",
82
    "You are a helpful assistant.",
83
    "You are an expert ML researcher who specializes in explaining complex concepts clearly.",
84
]
85
for role in roles:
86
    result = call(with_role(role, TEXT))
87
    print(f"  [{role[:40]}...]\n  → {result[:80]}\n")
88