LLM в роли классификатора текста

LLM как классификаторы текста

Классификация тональности

import anthropic, json

client = anthropic.Anthropic(api_key='YOUR_API_KEY')

def classify_sentiment(text):
    prompt = f'''
Classify the sentiment of the text below.
Return ONLY JSON: {{"sentiment": "positive|negative|neutral", "confidence": "high|medium|low"}}

Definitions:
- positive: Overall favorable opinion or emotion
- negative: Overall unfavorable opinion or dissatisfaction
- neutral: Factual, balanced, or no clear sentiment

Text: {text}
'''
    r = client.messages.create(
        model='claude-opus-4-5', max_tokens=50,
        messages=[{'role': 'user', 'content': prompt}]
    )
    return json.loads(r.content[0].text)

print(classify_sentiment('The product works but setup was painful.'))
print(classify_sentiment('Delivery was incredibly fast and packaging was perfect!'))

Классификация намерений

INTENT_CATEGORIES = [
    'purchase_intent: User wants to buy or is ready to purchase',
    'complaint: User is dissatisfied and reporting a problem',
    'question: User is asking for information or help',
    'cancellation: User wants to cancel a service or subscription',
    'compliment: User is expressing satisfaction or praise',
    'other: Does not fit any category above'
]

def classify_intent(message):
    categories_str = '\n'.join(f'- {c}' for c in INTENT_CATEGORIES)
    prompt = f'''
Classify the intent of this customer message.
Return JSON: {{"intent": str, "confidence": "high|medium|low"}}

Categories:
{categories_str}

Message: {message}
'''
    r = client.messages.create(model='claude-opus-4-5', max_tokens=80, messages=[{'role': 'user', 'content': prompt}])
    return json.loads(r.content[0].text)

print(classify_intent('I love the app but I need to cancel my plan.'))
print(classify_intent('How do I export my data to CSV?'))

Классификация по темам

def classify_topic(article_text, topics):
    topic_list = ', '.join(topics)
    prompt = f'''
Classify the topic of this article. Choose EXACTLY ONE from the list.
Return JSON: {{"topic": str, "secondary_topic": str or null}}

Available topics: {topic_list}

Article (first 300 chars): {article_text[:300]}
'''
    r = client.messages.create(model='claude-opus-4-5', max_tokens=80, messages=[{'role': 'user', 'content': prompt}])
    return json.loads(r.content[0].text)

topics = ['technology', 'sports', 'politics', 'business', 'science', 'health', 'entertainment']
text = 'The FDA approved a new mRNA vaccine for seasonal influenza, marking a breakthrough in vaccine technology.'
result = classify_topic(text, topics)
print(result)

Классификация срочности

URGENCY_PROMPT = '''
Classify the urgency of this support ticket.
Return JSON: {{"urgency": str, "reason": str}}

Urgency levels:
- critical: Service is completely down, data loss occurring, or security breach
- high: Major functionality broken, many users affected, no workaround
- medium: Non-critical feature broken, workaround exists, single user affected
- low: Cosmetic issue, enhancement request, general question

Be conservative: only use critical if the ticket explicitly describes a system-wide outage or data loss.

Ticket: {ticket}
'''

def classify_urgency(ticket_text):
    prompt = URGENCY_PROMPT.replace('{ticket}', ticket_text)
    r = client.messages.create(model='claude-opus-4-5', max_tokens=100, messages=[{'role': 'user', 'content': prompt}])
    return json.loads(r.content[0].text)

print(classify_urgency('The entire production database is down. All customers affected.'))
print(classify_urgency('Dark mode button is slightly off-center.'))

Многометочная классификация

def multi_label_classify(text, labels):
    label_list = ', '.join(labels)
    prompt = f'''
Classify this text. It may belong to multiple categories.
Return JSON: {{"labels": [str], "primary_label": str}}

Available labels: {label_list}

Rules:
- Include all labels that clearly apply
- Do NOT include labels that only marginally apply
- primary_label is the single most relevant label

Text: {text}
'''
    r = client.messages.create(model='claude-opus-4-5', max_tokens=100, messages=[{'role': 'user', 'content': prompt}])
    return json.loads(r.content[0].text)

labels = ['technical', 'billing', 'account', 'bug_report', 'feature_request', 'security']
text = 'I found a bug that exposes other users billing information on my account page.'
print(multi_label_classify(text, labels))

Шаблоны промптов для классификации

def build_classifier(category_definitions, additional_rules=''):
    cats = '\n'.join(f'- {k}: {v}' for k, v in category_definitions.items())
    return f'''
Classify the input text into exactly one category below.
Return JSON: {{"category": str, "confidence": "high|medium|low"}}

Categories:
{cats}
{('\nAdditional rules:\n' + additional_rules) if additional_rules else ''}

Text: {{text}}
'''

language_classifier = build_classifier({
    'formal': 'Business or academic writing, professional context',
    'informal': 'Casual, conversational, slang or colloquial',
    'technical': 'Domain-specific jargon, code, or specialized terminology',
    'emotional': 'High emotional content, personal, expressive'
})

print(language_classifier[:200])

Обработка неоднозначных результатов классификации

AMBIGUITY_PROMPT = '''
Classify this customer message. If the message is ambiguous or could fit multiple categories,
choose the category that would be most useful for routing it to the correct team.

Return JSON:
{{
  "category": str,
  "is_ambiguous": true | false,
  "alternative": str or null,
  "reasoning": str
}}

Categories: billing, technical_support, sales, account_management

Message: {message}
'''

message = 'I upgraded my plan but I am still seeing the free tier features.'
r = client.messages.create(
    model='claude-opus-4-5', max_tokens=150,
    messages=[{'role': 'user', 'content': AMBIGUITY_PROMPT.format(message=message)}]
)
print(json.loads(r.content[0].text))

Пакетная классификация для повышения эффективности

def batch_classify(items, categories):
    items_str = '\n'.join(f'{i+1}. {item}' for i, item in enumerate(items))
    cat_str = ', '.join(categories)

    prompt = f'''
Classify each item below. Categories: {cat_str}
Return JSON: {{"results": [{{"id": int, "category": str, "confidence": str}}]}}

Items:
{items_str}
'''
    r = client.messages.create(model='claude-opus-4-5', max_tokens=300, messages=[{'role': 'user', 'content': prompt}])
    return json.loads(r.content[0].text)['results']

texts = [
    'Absolutely love this product!',
    'It crashed twice today.',
    'What is your refund policy?',
    'The color options are limited.'
]
results = batch_classify(texts, ['positive_feedback', 'bug_report', 'inquiry', 'feature_request'])
for r in results:
    print(f'{texts[r["id"]-1][:30]}... -> {r["category"]} ({r["confidence"]})')

Оценка точности классификатора

labeled_test_set = [
    {'text': 'Great product, no issues!', 'expected': 'positive'},
    {'text': 'The app keeps crashing on startup.', 'expected': 'negative'},
    {'text': 'The screen resolution is 1080p.', 'expected': 'neutral'},
    {'text': 'Worst experience of my life.', 'expected': 'negative'},
    {'text': 'Works as advertised, decent value.', 'expected': 'positive'},
]

def evaluate_classifier(test_set, classify_fn):
    correct = 0
    for example in test_set:
        result = classify_fn(example['text'])
        if result['sentiment'] == example['expected']:
            correct += 1
        else:
            print(f'WRONG: "{example["text"][:40]}" -> got {result["sentiment"]}, expected {example["expected"]}')
    accuracy = correct / len(test_set)
    print(f'Accuracy: {correct}/{len(test_set)} = {accuracy:.0%}')
    return accuracy

evaluate_classifier(labeled_test_set, classify_sentiment)

Примеры на нескольких образцах в промптах классификации

few_shot_classifier = '''
Classify each customer message as: billing, technical, or general.

Examples:
  Input: "I was charged twice for this month." -> billing
  Input: "The app crashes when I open the dashboard." -> technical
  Input: "Do you have a mobile app?" -> general
  Input: "My invoice shows the wrong plan." -> billing
  Input: "I cannot log in, I get error 403." -> technical

Now classify:
Input: {message}
Return JSON: {"category": str, "confidence": "high|medium|low"}
'''

message = "My subscription was renewed but I cancelled last week."
prompt = few_shot_classifier.replace("{message}", message)
print(prompt)

Быстрая проверка

LLM в роли классификатора — основные выводы