2nd/10_Wiki/Topics/AI_and_ML/Figurative-Language.md

---
id: wiki-2026-0508-figurative-language
title: Figurative Language
category: 10_Wiki/Topics
status: verified
canonical_id: self
aliases: [figurative language, metaphor, simile, idiom, sarcasm, NLP figurative]
duplicate_of: none
source_trust_level: A
confidence_score: 0.88
verification_status: applied
tags: [linguistics, nlp, figurative, metaphor, simile, idiom, sarcasm]
raw_sources: []
last_reinforced: 2026-05-10
github_commit: pending
tech_stack:
  language: NLP / Python
  framework: Transformers / spaCy
---

# Figurative Language

## 매 한 줄
> **"매 literal meaning 의 X — 매 implied meaning"**. 매 metaphor, simile, idiom, sarcasm, irony, hyperbole, personification. 매 NLP 의 challenge — 매 LLM 의 의 의 better. 매 Lakoff conceptual metaphor → 매 modern transformer.

## 매 핵심

### 매 type
- **Metaphor**: "Time is money".
- **Simile**: "Brave as a lion".
- **Idiom**: "Kick the bucket".
- **Sarcasm / Irony**: 매 opposite literal.
- **Hyperbole**: "I died laughing".
- **Personification**: "The wind whispered".
- **Metonymy**: "The crown" (= king).

### 매 NLP challenge
- 매 literal interpreter 의 fail.
- 매 cross-cultural varies.
- 매 context-dependent.
- 매 sarcasm 의 매 hardest (no surface marker).

### 매 modern method
- **LLM** (GPT, Claude): 매 zero-shot decent.
- **Sentence embedding** + classifier.
- **Conceptual metaphor identification**.
- **Multi-task** (sentiment + sarcasm).

### 매 응용
1. **Sentiment analysis**: 매 sarcasm 의 detect.
2. **Translation**: 매 idiom localize.
3. **Content moderation**.
4. **Search / IR**.
5. **Education**: 매 figurative interpretation teaching.

## 💻 패턴

### Detect sarcasm (transformer)
```python
from transformers import pipeline
sarcasm_clf = pipeline('text-classification', model='helinivan/english-sarcasm-detector')
result = sarcasm_clf("Oh great, another Monday!")
# 매 [{'label': 'sarcastic', 'score': 0.92}]
```

### Detect metaphor (with LLM)
```python
def detect_metaphor(sentence, llm):
    prompt = f"""Is the following sentence metaphorical? If yes, identify the metaphor.

Sentence: "{sentence}"

Output JSON:
- is_metaphor: bool
- source_domain: ...
- target_domain: ...
- explanation: ..."""
    return json.loads(llm.generate(prompt))

detect_metaphor("Her career took off after that promotion.", llm)
# 매 source: flight, target: career
```

### Idiom translation
```python
IDIOM_DB = {
    'kick the bucket': {'en': 'die', 'ko': '죽다', 'fr': 'mourir'},
    'break a leg': {'en': 'good luck', 'ko': '행운을 빌다'},
    'piece of cake': {'en': 'easy', 'ko': '식은 죽 먹기'},
}

def translate_idiom(text, target_lang):
    for idiom, trans in IDIOM_DB.items():
        if idiom in text.lower():
            return text.replace(idiom, trans[target_lang])
    return text
```

### LLM idiom-aware translation
```python
def smart_translate(text, target_lang, llm):
    prompt = f"""Translate to {target_lang}, preserving figurative meaning.
If an idiom exists, use the equivalent target idiom rather than literal translation.

Source: "{text}"
Output: translation only."""
    return llm.generate(prompt)
```

### Conceptual metaphor (Lakoff)
```python
CONCEPTUAL_METAPHORS = {
    'TIME_IS_MONEY': ['save time', 'spend time', 'waste time', 'invest time'],
    'ARGUMENT_IS_WAR': ['attack', 'defend', 'win', 'lose'],
    'IDEAS_ARE_OBJECTS': ['grasp', 'hold', 'pass on'],
    'LOVE_IS_JOURNEY': ['go separate ways', 'crossroads'],
    'UP_IS_GOOD': ['high spirits', 'rise', 'top'],
}

def detect_conceptual(text):
    found = []
    for cm, markers in CONCEPTUAL_METAPHORS.items():
        if any(m in text.lower() for m in markers):
            found.append(cm)
    return found
```

### Multi-task (sentiment + sarcasm)
```python
class MultiTaskModel(torch.nn.Module):
    def __init__(self, base):
        super().__init__()
        self.base = base
        self.sentiment_head = torch.nn.Linear(768, 3)
        self.sarcasm_head = torch.nn.Linear(768, 2)
    
    def forward(self, x):
        feat = self.base(x).pooler_output
        return {
            'sentiment': self.sentiment_head(feat),
            'sarcasm': self.sarcasm_head(feat),
        }
```

### Eval (with figurative)
```python
def adjusted_sentiment(text, sarcasm_score, sentiment_score):
    """매 sarcastic → 매 flip sentiment."""
    if sarcasm_score > 0.7:
        return -sentiment_score
    return sentiment_score
```

### Hyperbole detection
```python
HYPERBOLE_MARKERS = ['always', 'never', 'died', 'a million', 'the worst', 'the best ever']

def has_hyperbole(text):
    return any(m in text.lower() for m in HYPERBOLE_MARKERS)
```

### Cross-cultural figurative test
```python
def cultural_metaphor_test(metaphor, languages, llm):
    results = {}
    for lang in languages:
        prompt = f"In {lang}, how is the metaphor '{metaphor}' typically expressed? If different, give the cultural equivalent."
        results[lang] = llm.generate(prompt)
    return results
```

### Personification detector
```python
def detect_personification(text, llm):
    prompt = f"""Identify personification (giving human traits to non-human).

Text: "{text}"
Output: list of personifications + the human trait + the entity."""
    return llm.generate(prompt)
```

### Figurative-aware embedding
```python
from sentence_transformers import SentenceTransformer
m = SentenceTransformer('all-mpnet-base-v2')

# 매 idiom and literal 의 should be different
emb1 = m.encode("It's raining cats and dogs.")  # 매 idiom
emb2 = m.encode("It is raining heavily.")  # 매 literal
emb3 = m.encode("Cats and dogs are falling from the sky.")  # 매 absurd literal
# 매 1-2 close, 1-3 distant (good model)
```

## 매 결정 기준
| 상황 | Approach |
|---|---|
| Sentiment / sarcasm | Multi-task transformer |
| Translation | LLM idiom-aware |
| Linguistics research | Conceptual metaphor |
| Search | Idiom DB + paraphrase |
| Education | LLM explainer |

**기본값**: 매 modern LLM 의 default + 매 sarcasm 의 specialized + 매 idiom DB + 매 cultural awareness.

## 🔗 Graph
- 부모: [[NLP]]
- 변형: [[Metaphor]] · [[Idiom]]
- 응용: [[Sentiment-Analysis]]
- Adjacent: [[Embodied Cognition]] · [[Pragmatics]]

## 🤖 LLM 활용
**언제**: 매 sentiment. 매 translation. 매 educational.
**언제 X**: 매 strict literal task.

## ❌ 안티패턴
- **Literal-only NLP**: 매 sarcasm miss.
- **Word-by-word translate**: 매 idiom break.
- **No cultural check**: 매 offense / confusion.
- **Single language assumption**: 매 i18n fail.

## 🧪 검증 / 중복
- Verified (Lakoff & Johnson, NLP figurative literature).
- 신뢰도 A.

## 🕓 Changelog
| 날짜 | 변경 |
|---|---|
| 2026-04-20 | Auto-reinforced |
| 2026-05-08 | Phase 1 |
| 2026-05-10 | Manual cleanup — types + 매 sarcasm / metaphor / idiom / multi-task code |