algo-nlp-summarization
用抽取式與生成式方法做文字摘要
中文翻譯
---
name: "algo-nlp-summarization"
description: "用抽取式與生成式兩種方法實作文字摘要。當使用者需要濃縮長文件、建立自動摘要流程，或比較摘要策略時使用，即使他們說的是『summarize this document』『TLDR』或『key points extraction』。"
metadata:
category: "WP-45 NLP 演算法"
tags: ["nlp", "summarization", "extractive", "abstractive"]
---
# 文字摘要
## 概述
文字摘要在保留關鍵資訊的前提下濃縮文件。抽取式：從原文挑出重要的句子串接起來。生成式：生成改寫內容的新文字。抽取式比較簡單、也更忠於原文；生成式比較流暢，但可能產生幻覺。
## 何時使用
**觸發條件：**
- 濃縮長文件、報告或文章集合
- 為內容策展建立自動摘要流程
- 針對某個使用情境比較抽取式與生成式
**不該使用的情況：**
- 需要完整理解整份文件時（摘要會損失細節）
- 結構化資料抽取（用 NER 或資訊抽取）
## 演算法
```
IRON LAW: Abstractive Summarization Can HALLUCINATE
Abstractive models may generate fluent text containing facts NOT in
the source. Always verify key claims in abstractive summaries against
the original document. For high-stakes use cases (legal, medical),
prefer extractive or use abstractive with factual consistency checking.
```
### 階段 1：輸入驗證
確定：輸入長度、目標摘要長度（比例或字數）、單文件 vs 多文件、領域。
**關卡：** 有輸入文字，已定義目標長度。
### 階段 2：核心演算法
**抽取式（TextRank/LexRank）：**
1. 把文件切成句子
2. 建立相似度圖（句子為節點，cosine 相似度為邊）
3. 在句子圖上跑 PageRank
4. 依排名選出前 k 個句子，再依原始位置重新排序
**生成式（transformer 類）：**
1. 使用預訓練模型（BART、T5、Pegasus）
2. 編碼輸入文件（有長度限制時，視需要做分塊處理）
3. 用 beam search 生成摘要
4. 後處理：檢查有沒有重複、事實一致性
### 階段 3：驗證
評估：對參考摘要算 ROUGE 分數（ROUGE-1、ROUGE-2、ROUGE-L）。人工檢查事實準確度和連貫性。
**關卡：** ROUGE 分數對這個領域來說合理，抽樣檢查沒有幻覺。
### 階段 4：輸出
回傳摘要與 metadata。
## 輸出格式
```json
{
"summary": "The company reported Q4 revenue of...",
"method": "extractive_textrank",
"metadata": {"input_words": 2000, "summary_words": 200, "compression_ratio": 0.10, "sentences_selected": 5}
}
```
## 範例
### 輸入輸出範例
**輸入：** 一篇 2000 字、關於季度財報的新聞文章
**預期：** 200 字的摘要，涵蓋：營收、獲利、財測、重點亮點。抽取式：挑 5-6 個句子。生成式：一段連貫的文字。
### 邊界情況
| 輸入 | 預期 | 原因 |
|-------|----------|-----|
| 非常短的輸入（< 100 字） | 原樣回傳或稍微精簡 | 本來就夠精簡了 |
| 多個互相矛盾的段落 | 摘要可能漏掉細微差異 | 摘要偏向主導的主題 |
| 專業術語 | 抽取式會保留，生成式可能簡化 | 領域專業會影響品質 |
## 注意事項
- **ROUGE ≠ 品質**：ROUGE 衡量的是和參考摘要的 n-gram 重疊度。ROUGE 高的摘要可能不連貫，ROUGE 低的摘要用詞不同，也可能非常好。
- **輸入長度限制**：Transformer 模型有最大 token 限制（512-4096）。長文件需要分塊策略（先分塊再摘要，或階層式摘要）。
- **重複**：生成式模型有時會重複片語。生成時要用重複懲罰（no_repeat_ngram_size）。
- **位置偏誤**：新聞文字的重要資訊都在前面（倒金字塔）。單純「取前 N 句」是很強的抽取式基準。
- **多文件摘要**：摘要多份相關文件時，需要處理跨來源的冗餘和矛盾。
## 參考資料
- TextRank/LexRank 的實作細節，見 `references/graph-based-extraction.md`
- 事實一致性檢查，見 `references/factual-consistency.md`英文原文
---
name: "algo-nlp-summarization"
description: "Implement text summarization using extractive and abstractive approaches. Use this skill when the user needs to condense long documents, build an automatic summarization pipeline, or compare summarization strategies — even if they say 'summarize this document', 'TLDR', or 'key points extraction'."
metadata:
category: "WP-45 NLP 演算法"
tags: ["nlp", "summarization", "extractive", "abstractive"]
---
# Text Summarization
## Overview
Text summarization condenses documents while preserving key information. Extractive: selects and concatenates important sentences from the original. Abstractive: generates new text that paraphrases the content. Extractive is simpler and more faithful; abstractive is more fluent but may hallucinate.
## When to Use
**Trigger conditions:**
- Condensing long documents, reports, or article collections
- Building automated summary pipelines for content curation
- Comparing extractive vs abstractive approaches for a use case
**When NOT to use:**
- When full document understanding is needed (summarization loses detail)
- For structured data extraction (use NER or information extraction)
## Algorithm
```
IRON LAW: Abstractive Summarization Can HALLUCINATE
Abstractive models may generate fluent text containing facts NOT in
the source. Always verify key claims in abstractive summaries against
the original document. For high-stakes use cases (legal, medical),
prefer extractive or use abstractive with factual consistency checking.
```
### Phase 1: Input Validation
Determine: input length, target summary length (ratio or word count), single-doc vs multi-doc, domain.
**Gate:** Input text available, target length defined.
### Phase 2: Core Algorithm
**Extractive (TextRank/LexRank):**
1. Split document into sentences
2. Build similarity graph (sentence nodes, cosine similarity edges)
3. Run PageRank on sentence graph
4. Select top-k sentences by rank, reorder by original position
**Abstractive (transformer-based):**
1. Use pre-trained model (BART, T5, Pegasus)
2. Encode input document (handle length limits with chunking if needed)
3. Generate summary with beam search
4. Post-process: check for repetition, factual consistency
### Phase 3: Verification
Evaluate: ROUGE scores (ROUGE-1, ROUGE-2, ROUGE-L) against reference summaries. Manual check for factual accuracy and coherence.
**Gate:** ROUGE scores reasonable for domain, no hallucinations in spot-check.
### Phase 4: Output
Return summary with metadata.
## Output Format
```json
{
"summary": "The company reported Q4 revenue of...",
"method": "extractive_textrank",
"metadata": {"input_words": 2000, "summary_words": 200, "compression_ratio": 0.10, "sentences_selected": 5}
}
```
## Examples
### Sample I/O
**Input:** 2000-word news article about quarterly earnings
**Expected:** 200-word summary covering: revenue, profit, guidance, key highlights. Extractive: 5-6 selected sentences. Abstractive: coherent paragraph.
### Edge Cases
| Input | Expected | Why |
|-------|----------|-----|
| Very short input (< 100 words) | Return as-is or minimal trimming | Already concise |
| Multiple contradicting sections | Summary may miss nuance | Summarization favors dominant theme |
| Technical jargon | Extractive preserves, abstractive may simplify | Domain expertise affects quality |
## Gotchas
- **ROUGE ≠ quality**: ROUGE measures n-gram overlap with references. A high-ROUGE summary can be incoherent, and a low-ROUGE summary can be excellent with different word choices.
- **Input length limits**: Transformer models have max token limits (512-4096). Long documents need chunking strategies (chunk-then-summarize or hierarchical summarization).
- **Repetition**: Abstractive models sometimes repeat phrases. Use repetition penalty during generation (no_repeat_ngram_size).
- **Position bias**: In news text, important information is front-loaded (inverted pyramid). Simple "take first N sentences" is a strong extractive baseline.
- **Multi-document summarization**: Summarizing multiple related documents requires handling redundancy and contradiction across sources.
## References
- For TextRank/LexRank implementation details, see `references/graph-based-extraction.md`
- For factual consistency checking, see `references/factual-consistency.md`