algo-social-sentiment
用 VADER 替社群媒體文字做情緒評分
中文翻譯
---
name: "algo-social-sentiment"
description: "實作 VADER 情緒分析，替社群媒體文字評分。當使用者需要分析推文、評論或社群貼文的情緒、計算綜合情緒分數、或分類文字的正負面時使用，即使他們說的是「這是正面還是負面」、「這些留言的情緒」或「社群媒體的氛圍分析」也一樣。"
metadata:
category: "WP-38 社群演算法"
tags: ["social-media", "sentiment-analysis", "vader", "nlp"]
---
# VADER 情緒分析（VADER Sentiment Analysis）
## 概述
VADER（Valence Aware Dictionary and sEntiment Reasoner）是針對社群媒體最佳化、以詞典和規則為基礎的情緒工具。回傳綜合分數 compound，範圍 [-1, +1]，結合了正面、負面和中性的比例。每則文字的執行複雜度是 O(n)，n = 字數。不需要訓練。
## 何時使用
**觸發條件：**
- 要分析社群媒體貼文、推文或評論的情緒
- 要在不訓練 ML 模型的情況下，快速做情緒評分
- 要處理帶有俚語、表情符號和非正式用語的文字
**何時不要用：**
- 正式／學術文字（VADER 是針對社群媒體調校的）
- 領域專屬的情緒很重要時（例如金融情緒，用 FinBERT）
- 偵測反諷很關鍵時（VADER 偵測不到反諷）
## 演算法
```
鐵律：VADER 是為「社群媒體」文字設計的
它把俚語、表情符號、大小寫和標點，都當成情緒的修飾因子處理。
把它套用在正式文件（法律、學術、醫療）上，
產生的分數不可靠。領域專屬的文字，
要改用該領域訓練的模型。
```
### 階段 1：輸入驗證
把文字斷詞。要保留：大小寫（全大寫 = 強調）、標點（! 會放大）、表情符號／emoji。
**關卡：** 文字非空、編碼處理正確。
### 階段 2：核心演算法
1. 在 VADER 詞典中查每個 token（7,500 個以上帶情緒評分的詞彙）
2. 套用文法規則：否定（"not good" = 負面）、程度修飾詞（"very good" > "good"）、大小寫加成、標點放大
3. 計算正面、負面、中性比例的原始價數分數
4. 計算 compound 分數：用公式把所有價數分數的總和正規化：compound = sum / √(sum² + α)，其中 α = 15
### 階段 3：驗證
分類：compound ≥ 0.05 → 正面、≤ -0.05 → 負面，其餘為中性。抽查樣本結果。
**關卡：** 分類通過 10-20 個範例的人工抽查。
### 階段 4：輸出
回傳每則文字的 compound 分數和正負面分類。
## 輸出格式
```json
{
"results": [{"text": "...", "compound": 0.76, "pos": 0.45, "neu": 0.55, "neg": 0.0, "label": "positive"}],
"metadata": {"texts_analyzed": 500, "distribution": {"positive": 0.45, "neutral": 0.35, "negative": 0.20}}
}
```
## 範例
### 輸入輸出範例
**輸入：** "This product is AMAZING!!! 😍"
**預期：** compound ≈ 0.87（正面）。被這些加成：全大寫、!!!、😍 emoji。
### 邊界情況
| 輸入 | 預期 | 原因 |
|-------|----------|-----|
| "Not bad at all" | 略微正面（約 0.2） | 雙重否定有被部分處理 |
| "😂😂😂" | 正面 | 詞典中有對應 emoji |
| 空字串 | Compound = 0，中性 | 沒有 token 可評分 |
## 常見陷阱
- **反諷看不出來**："Oh great, another meeting" 會被判成正面。VADER 沒有反諷偵測。
- **否定的視窗**：VADER 在 3 個字的視窗內套用否定。"I do not think this is bad" 可能會把否定鏈解析錯。
- **Emoji 涵蓋範圍**：VADER 的 emoji 詞典可能沒涵蓋較新的 emoji。視需要更新或補充。
- **語言限制**：VADER 只支援英文。中文／日文要用該語言專用的工具（例如中文用 SnowNLP）。
- **compound 門檻的敏感度**：0.05 這個界線是隨意定的。要依你的使用情境，以及對偽陽性的容忍度來調整門檻。
## 參考資料
- VADER 詞典與規則的文件，見 `references/vader-rules.md`
- 與 transformer 情緒模型的比較，見 `references/model-comparison.md`英文原文
---
name: "algo-social-sentiment"
description: "Implement VADER sentiment analysis for social media text scoring. Use this skill when the user needs to analyze sentiment in tweets, reviews, or social posts, compute compound sentiment scores, or classify text polarity — even if they say 'is this positive or negative', 'sentiment of these comments', or 'social media mood analysis'."
metadata:
category: "WP-38 社群演算法"
tags: ["social-media", "sentiment-analysis", "vader", "nlp"]
---
# VADER Sentiment Analysis
## Overview
VADER (Valence Aware Dictionary and sEntiment Reasoner) is a lexicon and rule-based sentiment tool optimized for social media. Returns compound score [-1, +1] combining positive, negative, and neutral proportions. Runs in O(n) per text where n = word count. No training required.
## When to Use
**Trigger conditions:**
- Analyzing sentiment in social media posts, tweets, or reviews
- Quick sentiment scoring without ML model training
- Processing text with slang, emoticons, and informal language
**When NOT to use:**
- For formal/academic text (VADER is tuned for social media)
- When domain-specific sentiment matters (e.g., financial sentiment — use FinBERT)
- When sarcasm detection is critical (VADER doesn't detect sarcasm)
## Algorithm
```
IRON LAW: VADER Is Designed for SOCIAL MEDIA Text
It handles slang, emoticons, capitalization, and punctuation as
sentiment modifiers. Applying it to formal documents (legal, academic,
medical) produces unreliable scores. For domain-specific text, use
domain-trained models instead.
```
### Phase 1: Input Validation
Tokenize text. Preserve: capitalization (ALL CAPS = emphasis), punctuation (! amplifies), emoticons/emoji.
**Gate:** Text is non-empty, encoding handled correctly.
### Phase 2: Core Algorithm
1. Look up each token in VADER lexicon (7,500+ sentiment-rated terms)
2. Apply grammatical rules: negation ("not good" = negative), degree modifiers ("very good" > "good"), capitalization boost, punctuation amplification
3. Compute raw valence scores for positive, negative, neutral proportions
4. Compute compound score: normalized sum of all valence scores using formula: compound = sum / √(sum² + α) where α = 15
### Phase 3: Verification
Classify: compound ≥ 0.05 → positive, ≤ -0.05 → negative, else neutral. Spot-check sample results.
**Gate:** Classifications pass manual spot-check on 10-20 examples.
### Phase 4: Output
Return compound score and polarity classification per text.
## Output Format
```json
{
"results": [{"text": "...", "compound": 0.76, "pos": 0.45, "neu": 0.55, "neg": 0.0, "label": "positive"}],
"metadata": {"texts_analyzed": 500, "distribution": {"positive": 0.45, "neutral": 0.35, "negative": 0.20}}
}
```
## Examples
### Sample I/O
**Input:** "This product is AMAZING!!! 😍"
**Expected:** compound ≈ 0.87 (positive). Boosted by: CAPS, !!!, 😍 emoji.
### Edge Cases
| Input | Expected | Why |
|-------|----------|-----|
| "Not bad at all" | Slightly positive (~0.2) | Double negation partially handled |
| "😂😂😂" | Positive | Emoji mapped in lexicon |
| Empty string | Compound = 0, neutral | No tokens to score |
## Gotchas
- **Sarcasm is invisible**: "Oh great, another meeting" reads as positive. VADER has no sarcasm detection.
- **Negation window**: VADER applies negation within a 3-word window. "I do not think this is bad" may misparse the negation chain.
- **Emoji coverage**: VADER's emoji lexicon may not cover newer emoji. Update or supplement as needed.
- **Language limitation**: VADER is English-only. For Chinese/Japanese, use language-specific tools (e.g., SnowNLP for Chinese).
- **Compound threshold sensitivity**: The 0.05 boundary is arbitrary. Adjust thresholds based on your specific use case and tolerance for false positives.
## References
- For VADER lexicon and rules documentation, see `references/vader-rules.md`
- For comparison with transformer-based sentiment models, see `references/model-comparison.md`