Research Survey Skill

If ambiguous, ask: "特定の論文を登録したいですか？それとも分野全体のサーベイを行いますか？"

Mode 1: Deep Search Workflow

Step 1 — Clarify Survey Scope

Ask the user for:

• サーベイ名 (directory name, kebab-case, e.g. dialogue-user-profiling)
• リサーチクエスチョン in one sentence
• 対象年範囲 (default: last 10 years)
• アンカー論文 — any known key papers or authors
• 言語設定 — survey document language (default: Japanese)

Create surveys/{survey-name}/config.json:

{
  "name": "...",
  "description": "...",
  "created": "YYYY-MM-DD",
  "query_field": "...",
  "year_range": "20XX-20XX",
  "anchor_papers": [],
  "searches_performed": []
}

Step 2 — Plan Search Queries

Decompose the field into 5–10 sub-queries covering:

• Core method / approach
• Application domain variants
• Adjacent techniques
• Historical foundational terms
• Japanese-language equivalents (重要: NLP/HCI 分野では日本語論文も多い)

Show the query plan to the user. Wait for approval before proceeding. Record approved queries in config.json under searches_performed.

Step 3 — Multi-Source Search

For each approved query, search in this order. Stop accumulating when 30+ high-relevance results are collected.

1. search_semantic_scholar — best for citation counts + structured abstracts
2. search_arxiv — preprints, essential for NLP/ML/HCI
3. search_crossref — broad DOI coverage for published papers
4. search_google_scholar — fallback for hard-to-find papers

For specialized fields also try: search_pubmed (biomedical), search_springer, search_scopus.

After all searches: deduplicate by DOI, then by title similarity (>85% match = same paper). Report the deduplicated list count to the user before proceeding.

Step 4 — Per-Paper Registration Loop

For each unique paper in the deduplicated list:

1. Compute doi-safe-id. Check if papers/{doi-safe-id}/ already exists.

   • If exists: skip creation; add this survey to its surveys list in info.json and continue.
   • If not: create directory.

2. Write info.json with all available bibliographic fields. Leave tags.method, tags.dataset, tags.evaluation_metrics empty for now. Set "surveys": ["{survey-name}"].

3. Attempt PDF download (try in order, stop at first success):

   • download_paper with platform: "arxiv" if arXiv ID is known
   • download_paper with platform: "scihub" using DOI
   • download_paper with platform: "semantic" if Semantic Scholar paper ID available Set savePath to the paper directory (papers/{doi-safe-id}/). On failure: set pdf_path: null, record pdf_url from search result if available.

4. Generate notes.md using the template in references/templates.md.

   • Fill all sections from abstract and any full-text content available.
   • For sections requiring full text that is unavailable, write: PDF未取得のため詳細不明
   • Write in Japanese unless the survey's config.json specifies otherwise.

⚠️ STOP after all papers are registered. Do NOT assign tags yet — tags are assigned globally in Step 5.

Step 5 — Global Tag Assignment

With the full set of papers now visible, do ONE pass across all papers:

1. Read all info.json and abstracts / notes in memory.
2. Apply tags from references/tag-taxonomy.md to each paper:
   • tags.method — primary methods/models used
   • tags.dataset — datasets used for training or evaluation
   • tags.evaluation_metrics — evaluation metrics reported
   • tags.has_user_study — true if the paper recruited real users for evaluation
   • tags.language — ["ja"] for Japanese venue, ["en"] for English, ["ja","en"] for bilingual
3. Write updated info.json for every paper.

Apply tags consistently: if three papers use BERT, all three must use the same tag string "BERT". If a method isn't in the taxonomy, add it to references/tag-taxonomy.md first, then use it.

Step 6 — Duplicate and Research Group Detection

Scan all papers in this survey:

1. Same-paper duplicates: papers with identical DOI, title similarity >90%, or same first author + same year + substantially overlapping abstract. → Set duplicate_papers in each paper's info.json pointing to the other DOI(s). Key case: Japanese domestic workshop paper + extended English journal version by same author.

2. Research group clustering: group papers sharing first/last authors or institutional keywords. → Set research_group (e.g. "Kyoto-NLP-Lab") and same_research_group list.

3. Cross-references: for each paper, identify other papers in this survey that it would likely cite or be cited by. → Set related_papers.

Before writing, report detected clusters to the user for confirmation: "以下の論文が同一研究グループと判定されました: ..."

Step 7 — Subfield Identification (Bottom-Up)

Do NOT pre-define subfields. Instead:

1. Read all notes and tags.
2. Cluster papers by shared method/topic themes emerging from the content.
3. Propose 3–8 subfield names to the user with paper counts and representative papers.
4. Adjust based on user feedback.
5. Assign each paper to 1–2 subfields. Store assignment in info.json under a new subfields key.

Step 8 — Write Survey Documents

surveys/{survey-name}/subfields/{subfield-name}.md (one per subfield):

# {Subfield Name}

## 概要
[What this subfield addresses, 2-3 sentences]

## 歴史的流れ
[Oldest → newest papers; note when key ideas emerged]

## 論文比較表
| 論文 | 年 | 手法 | データセット | 主要結果 | 備考 |
|------|----|----- |------------|---------|------|

## 他サブフィールドとの関係
## 未解決問題・今後の方向性

surveys/{survey-name}/overview.md:

# {Survey Name} サーベイ概要

## エグゼクティブサマリー
## サブフィールドマップ
[各サブフィールドと相互関係]
## 全体的なトレンド（時系列）
## 初学者向け推奨読書順序

surveys/{survey-name}/paper-list.md: Full comparison table sorted by year:

| 論文 | 年 | 手法 | データセット | User Study | 評価指標 | サブフィールド |
|------|----|----- |------------|-----------|---------|-------------|

Mode 2: Paper Registration Workflow

Step 1 — Identify the Paper

Extract DOI, title, authors from user input (BibTeX / plain text / PDF path / URL).

If DOI available: call get_paper_by_doi to fetch full metadata. If only title/authors: call search_semantic_scholar with title; show top match to user for confirmation.

Step 2 — Duplicate Check

Compute doi-safe-id. Check if papers/{doi-safe-id}/ exists.

• Exists: Report existing entry. Ask: "notes.md を更新しますか？既存のサーベイに追加しますか？"
• Not exists: proceed to Step 3.

Step 3 — Create Paper Entry

1. Create papers/{doi-safe-id}/.
2. Write info.json from fetched metadata.
3. PDF handling:
   • If user provided a PDF path: record it in pdf_path (absolute path).
   • Otherwise attempt download_paper (arxiv → scihub → semantic).
4. Generate notes.md from abstract and any available full text.
5. Apply tags from references/tag-taxonomy.md to info.json.

Step 4 — Survey Linkage

Read all surveys/*/config.json. For each survey, check if this paper's tags and topic match the survey's query_field. Suggest:

"この論文は以下のサーベイに関連しそうです: [X, Y]. 追加しますか？"

If added: update info.json surveys list and append paper row to paper-list.md.

File Templates

Full templates are in references/templates.md. Tag vocabulary is in references/tag-taxonomy.md.

Error Handling

Constraints

• Never delete existing papers/ entries. Updates are always additive.
• Do not overwrite notes.md unless user explicitly requests re-analysis.
• Tag vocabulary must come from references/tag-taxonomy.md; add new terms there before using them.
• Survey documents are written in the language from config.json (default: Japanese).
• overview.md is always the last document written — it synthesizes all subfield docs.