citation-impact

分类: 工具与效率 | 上传者: yuanmu97 | 下载: 0 | 版本: v1.0（最新）
Analyze how academic papers are cited. Given a config.yaml (generated by the web tool), run the analysis pipeline to download PDFs, extract citation contexts, classify sentiment, and produce a Markdown report. Use when the user asks to analyze citations, citation impact, or provides a config.yaml for citation analysis.
更新日志: Source: GitHub https://github.com/yuanmu97/citation-impact
目录结构

当前层级: agent/
📁 data/
- 📄 ccf_rankings.json 200.5 KB
📁 scripts/
- 📄 __init__.py 0 B
- 📄 ccf_lookup.py 2.3 KB
- 📄 citation_extractor.py 13.0 KB
- 📄 config_parser.py 5.2 KB
- 📄 main.py 18.4 KB
- 📄 openalex_client.py 5.8 KB
- 📄 pdf_processor.py 5.1 KB
- 📄 report_generator.py 5.5 KB
- 📄 semantic_scholar_client.py 4.6 KB
📄 requirements.txt 39 B
📄 SKILL.md 10.0 KB
SKILL.md

---
name: citation-impact
description: Analyze how academic papers are cited. Given a config.yaml (generated by the web tool), run the analysis pipeline to download PDFs, extract citation contexts, classify sentiment, and produce a Markdown report. Use when the user asks to analyze citations, citation impact, or provides a config.yaml for citation analysis.
---

# Citation Impact Analyzer

Analyze how academic papers are cited — extract citation contexts from PDFs and classify each citation as high praise, neutral reference, or critical mention.

## Using in Cursor

- **Trigger**: When the user asks to "analyze citations", "analyze citation impact", or provides/mentions a `config.yaml` for citation analysis.
- **Paths**: Skill scripts are in `~/.cursor/skills/citation-impact/agent/scripts/`. The user's `config.yaml`, `citation_pdfs/` folder, and output are all in the **user's current workspace**.
- **Python**: The skill requires Python 3.9+. Before first run, install dependencies:
  ```bash
  pip install -r ~/.cursor/skills/citation-impact/agent/requirements.txt
  ```

## Workflow

When triggered, follow these steps **in order**:

### Step 1: Locate config.yaml

Find the `config.yaml` in the user's workspace. If the user specifies a path, use that. Otherwise look in the current working directory.

Read the config to understand:
- Which target papers to analyze
- The pre-selected citing papers for each target (v2.0 configs have these pre-built)
- Where PDFs are stored: **always under workspace root** `citation_pdfs/` (script uses `options.pdf_dir`, default `"."`; never use a subfolder name like the workspace name)
- Where to write output (`options.output_dir`, defaults to `./citation_output`)

### Step 2: Install dependencies (if needed)

Check if required packages are installed. If not:
```bash
pip install -r ~/.cursor/skills/citation-impact/agent/requirements.txt
```

### Step 3: Run the pipeline

```bash
python ~/.cursor/skills/citation-impact/agent/scripts/main.py --config <path-to-config.yaml>
```

The script will:
- Create the output directory structure
- For v2.0 configs: use the pre-selected citing papers (no API calls needed for filtering)
- Create per-paper folders under **workspace root** `citation_pdfs/` (pdf_dir is `"."` in generated configs)
- Download Open Access PDFs automatically
- Find user-placed PDFs in `citation_pdfs/<paper_folder>/` under the workspace root
- Extract text from each PDF using PyMuPDF
- Locate citation contexts (sentences referencing the target paper)
- Generate `summary.json` files and an initial `citation_report.md`

### Step 4: Analyze and complete the report

After the script completes, read each paper's `summary.json` under `<output_dir>/<paper_name>/summary.json`.

For each citing paper, fill in **four areas** in `<output_dir>/citation_report.md`:

**引用原文 (citation context)**:
- Check `citation_contexts` in `summary.json`. If non-empty, the pipeline already extracted them.
- If empty (`[]`) but a PDF exists in `citation_pdfs/<pdf_folder>/`, **read the PDF directly** to find where the target paper is cited. Search for the target paper's title keywords, author names, or reference numbers like `[N]`. Extract the surrounding 2-3 sentences as context.
- Only mark as "(未提取到)" when the PDF is genuinely unavailable (not downloaded).

**机构 (institutions)**:
- Read `first_pages_text` from `summary.json` — it contains the first 2 pages where author affiliations typically appear.
- If unclear, **read the PDF directly** from `citation_pdfs/<pdf_folder>/`.

**评价类型 (sentiment)** — classify each citation as one of:

| Sentiment | Chinese Label | Description |
|-----------|--------------|-------------|
| High praise | 高度评价 | Explicitly praises, builds upon, or highlights the contribution as significant/influential. Also applies when the paper is cited as a "representative", "pioneering", or "foundational" work, or cited multiple times across different sections (Introduction + Related Work + Experiments, etc.), indicating the authors consider it an important reference in the field. |
| Neutral | 一般引用 | Single standard reference for background, comparison, or related work listing — typically a one-time mention in a "[X], [Y], [Z]" list within Related Work |
| Critical | 批评性引用 | Critiques, identifies limitations, proposes improvements, or disagrees |

Consider the full context: look for words like "seminal", "groundbreaking", "builds upon", "representative" (praise), "proposed", "introduced", "similar to" (neutral), or "fails to", "limited", "outperforms" with the target as the weaker method (critical). Also consider **citation frequency**: being cited 2+ times across different sections is a strong signal of high praise — it means the authors view the work as important enough to reference repeatedly.

**分析说明 (reasoning)** — briefly explain the classification (e.g. which section, how the target paper is referenced, any praise or criticism).

For citations where the PDF exists but the pipeline extracted no context, you **must** read the PDF yourself and extract the citation context — do not leave it as "(未提取到)". Only mark as "(未提取到)" and default to "一般引用" when the PDF is truly unavailable (download failed, not provided).

### Step 5: Add summary statistics

At the end of each paper's table, add:

```markdown
### 总结
- 高度评价: X citations (XX.X%)
- 一般引用: X citations (XX.X%)
- 批评性引用: X citations (XX.X%)
```

Add a final overall summary section at the end of the report.

### Step 6: Deep annotation for high-praise citations

For each citation classified as **高度评价**, add a blockquote section **after the citation table and before the 总结**, with author background and institution assessment:

**Author background research** — For the top 2-3 authors and corresponding author, **search** for notable titles/honors:
- IEEE Fellow, ACM Fellow, or other society fellowships
- National-level talent titles (杰青、长江学者、优青、院士, etc.)
- Leadership positions (department chair, dean, lab director, etc.)
- Only report what you find with confidence. Do not guess or fabricate.

**Institution assessment** — Based on the already-extracted institution info, assess whether it qualifies as a top-tier research institution using your own knowledge:
- Top CS departments (e.g. CSRankings top-ranked universities)
- Elite enterprise labs (e.g. MSRA, Google DeepMind, Meta FAIR, etc.)
- State key laboratories, national research centers

**Format** — Insert one blockquote per high-praise citation, between the table and the 总结:

```markdown
> **#N 高度评价背景**:
> - **知名作者**: Author Name (IEEE Fellow), Another Author (杰青)
> - **头部机构**: University Name (CSRankings top-N in XX); Lab Name
```

If no notable authors or institutions are found, simply omit that bullet — do not write "无".

**Overall summary** — At the end of the final 整体总结 paragraph, append a sentence summarizing how many notable scholars and top institutions appear among the high-praise citations, as evidence of citation quality. For example:
```
高度评价引用中包含 X 位知名学者（IEEE/ACM Fellow 等）和 Y 所头部机构，进一步表明该工作在领域内的认可度。
```

## Config Format (v2.0)

The web tool generates v2.0 configs where citing papers are pre-selected:

```yaml
version: '2.0'
researcher:
  name: "Researcher Name"
  google_scholar_id: "XXXXX"
  openalex_id: "AXXXXXXXXXX"
target_papers:
  - openalex_id: W1234567890
    title: "Paper Title"
    year: 2023
    doi: 10.xxxx/xxxxx
    citing_papers:
      - openalex_id: W9876543210
        title: "Citing Paper Title"
        year: 2025
        doi: 10.yyyy/yyyyy
        venue: "ICML"
        ccf_rank: A
        authors: "Alice, Bob"
        pdf_source: oa          # "oa" = auto-download, "local" = user placed, "unknown" = try both
        pdf_folder: "Citing Paper Title"   # subfolder name under citation_pdfs/
options:
  pdf_dir: "."                      # 固定为 "."：citation_pdfs 位于工作区根目录，不可改为工作区名或子目录名
  output_dir: "./citation_output"
```

Key points:
- **pdf_dir**: 必须为 `"."`。流水线在工作区根目录执行，PDF 目录为 **工作区根目录下的 `citation_pdfs/`**。Web 导出的 config 已固定为 `"."`；若你生成或修改 config，也必须写 `options.pdf_dir: "."`，切勿使用工作区文件夹名（如 `test-run`）或其它子路径。提示用户放置 PDF 时，只说 **`citation_pdfs/<pdf_folder>/`**（相对工作区根目录），不要写为 `<某子目录>/citation_pdfs/...`。
- `pdf_source: oa` — the script will auto-download from the Open Access URL
- `pdf_source: local` — the user places a PDF in **workspace root** `citation_pdfs/<pdf_folder>/`
- `pdf_source: unknown` — the script tries online download first, then checks the local folder
- The script accepts any `.pdf` file found in the subfolder (filename doesn't matter)

## Output Structure

```
citation_output/
├── citation_report.md          # Final report (you fill in sentiment)
├── Paper Title 1/
│   ├── summary.json            # Structured data with citation contexts
│   └── pdfs/                   # Downloaded PDFs (for OA papers)
└── Paper Title 2/
    ├── summary.json
    └── pdfs/
```

## Report Format

The report is in Chinese (学术论文被引情况报告) with this table structure:

| # | 他引论文 | 作者 | 机构 | 发表出处 | 年份 | 评级 | 引用原文 | 评价类型 | 分析说明 |
|---|---------|------|------|---------|------|------|---------|---------|---------|
| 1 | Title   | Authors | Institutions | Venue | Year | CCF Rank | Context | Sentiment | Reasoning |

Your job is to fill:
- "机构" — extract from `first_pages_text` in summary.json, or read the PDF directly
- "评价类型" — 高度评价 / 一般引用 / 批评性引用
- "分析说明" — 简要说明判定理由（如：在哪个章节、以什么方式引用、是否有褒贬倾向）
登录后下载/点赞/收藏 ❤ 2 | ★ 0
citation-impact

目录结构

SKILL.md

举报内容

提示
