日本語NLP趋势报告

Research Japanese NLP trends for topic: "$ARGUMENTS" by combining the bundled dataset with the latest web information.

Instructions

Preamble — Establish the current date

Before doing anything else, run this once and remember the values — every subsequent step that mentions a year, month, or report date refers to them:

echo "YEAR_NOW=$(date +%Y)"
echo "YEAR_PREV=$(($(date +%Y) - 1))"
echo "REPORT_DATE_EN=$(LC_TIME=C date '+%B %Y')"
echo "REPORT_DATE_JP=$(date '+%Y年%-m月')"

Substitute these values everywhere this skill writes ${YEAR_NOW}, ${YEAR_PREV}, ${REPORT_DATE_EN}, or ${REPORT_DATE_JP} below. Do not hardcode dates — the skill must always reflect the current month.

Step 0 — Handle empty input

If $ARGUMENTS is empty or blank, treat it as a request for a general overview of current Japanese NLP trends. Use the following defaults for the rest of the steps:

• Topic label for output headings: "Japanese NLP Overall Trends" (use "日本語NLP 全体トレンド" only when the user's query was written in Japanese)
• Keywords for Step 1 (local dataset survey): japanese nlp, llm, bert, embed, speech, morpholog, translat — These broad keywords give a cross-category snapshot of the most popular resources
• WebSearch queries for Step 5: cover multiple active sub-fields rather than one topic:
  • japanese NLP trends ${YEAR_NOW} overview
  • 日本語 NLP 最新動向 ${YEAR_NOW}
  • japanese LLM embedding benchmark ${YEAR_NOW} github
  • 日本語 自然言語処理 注目 モデル ${YEAR_NOW}
  • huggingface japanese models trending ${YEAR_NOW}
• Report title: ## 📊 Japanese NLP Trend Report (as of ${REPORT_DATE_EN}) instead of ## 📊 Trend Report for "$ARGUMENTS" (use ## 📊 日本語NLP 全体トレンドレポート (${REPORT_DATE_JP}時点) only when output language is Japanese)
• Section 1 (Overview): write a broad 3–4 sentence overview covering the major active sub-fields (LLMs, embeddings/RAG, speech, morphological analysis, benchmarks)

Then continue normally from Step 1 using the above defaults.

Step 1 — Interpret the topic

The user's topic is: "$ARGUMENTS"

Generate two keyword sets:

1. English stem keywords (4–6) for searching the local dataset (descriptions are mostly English). Use stems like morpholog, embed, classif, translat, generat, recogni, etc. Add well-known Japanese-specific tool/model names where applicable: mecab, sudachi, ginza, bert, gpt, llama, swallow, elyza, rinna, calm, ruri, whisper, voicevox, manga-ocr, jglue, llm-jp-eval, etc.

2. Web search phrases (3–5) mixing English and Japanese for the latest information.

Step 2 — Locate the data file

The data file ships with the plugin. Resolve its path via ${CLAUDE_PLUGIN_ROOT} (Claude Code substitutes this inline in skill content), falling back to a scoped search only if the install is unusual:

RESOURCES_PATH="${CLAUDE_PLUGIN_ROOT}/data/resources.json"
[ -f "$RESOURCES_PATH" ] || RESOURCES_PATH="$(find "${HOME}/.claude/plugins" -type f -name resources.json 2>/dev/null | grep "awesome-japanese-nlp-resources/" | head -1)"
echo "RESOURCES_PATH=$RESOURCES_PATH"

Save the resulting absolute path as RESOURCES_PATH.

Step 3 — Survey existing resources (inline scoring)

Do NOT use the Read tool on resources.json — it exceeds the read limit. Run this Python block in Bash, substituting RESOURCES_PATH and your English stem keywords:

python3 << 'EOF'
import json
from collections import Counter

with open("RESOURCES_PATH") as f:
    data = json.load(f)

keywords = ["keyword1", "keyword2", "keyword3"]  # English stems from Step 1

results = []
for item in data:
    n = item.get("n", "").lower()
    d = item.get("d", "").lower()
    s = item.get("s", "").lower()
    c = item.get("c", "").lower()

    text_score = 0
    for kw in keywords:
        kw = kw.lower()
        if n == kw:       text_score += 20
        elif kw in n:     text_score += 10
        if kw in d:       text_score += 5
        if kw in s:       text_score += 3
        if kw in c:       text_score += 2

    if text_score < 8:
        continue

    ns = item.get("ns") or 0
    nd = item.get("nd") or 0
    sc = item.get("sc") or 0
    pop = (ns if ns else nd) * 2.5
    qual = min(5, sc * 5 / 21)
    combined = text_score + pop + qual

    results.append((combined, text_score, item))

results.sort(key=lambda x: -x[0])

# Print top 10 for the trend snapshot
print("=== TOP MATCHES ===")
for combined, ts, item in results[:10]:
    st = item.get("st", 0) or 0
    dl = item.get("dl", 0) or 0
    print(f"score={combined:.1f} | {item.get('c','')} | ⭐{st} 📥{dl}")
    print(f"  n={item['n']}")
    print(f"  u={item['u']}")
    print(f"  d={item.get('d','')[:120]}")
    print()

# Print category distribution of full result set
print("=== CATEGORY DISTRIBUTION ===")
cats = Counter(item.get("c", "") for _, _, item in results)
for cat, count in cats.most_common(8):
    print(f"  {count:3d}  {cat}")
EOF

This gives you:

• Top 10 items currently in the dataset for the topic
• Category distribution showing which resource types dominate
• Popularity stats (stars / downloads) to gauge maturity

Step 4 — Identify trend angles to investigate

Before searching the web, decide what to look for. Based on the survey in Step 3, ask:

• What's the dominant architecture in the top matches (BERT vs. GPT vs. T5 vs. LLaMA)?
• What's the dominant resource type (libraries vs. models vs. corpora)?
• Are the top items recent (within the last 2 years) or older (>3 years ago)?
• Is there an apparent gap (e.g., no recent multimodal models, no public benchmarks)?

Use these angles to shape Step 5's queries.

Step 5 — Web research

Use WebSearch + WebFetch only — do not use the gh CLI in this project.

Run 4–6 WebSearch queries. Always include ${YEAR_NOW} (and optionally ${YEAR_PREV}) to bias toward recency. Mix English and Japanese:

• Japanese NLP <topic-en> ${YEAR_NOW}
• 日本語 <topic> 最新 モデル ${YEAR_NOW}
• arxiv japanese <topic-en> ${YEAR_PREV} ${YEAR_NOW}
• huggingface japanese <topic-en> new release
• <topic-en> github trending japanese ${YEAR_NOW}
• Optional domain-specific: <topic-en> jp benchmark ${YEAR_NOW}, 日本語 <topic> 評価

When a specific high-value URL surfaces (e.g. an arXiv abstract, a HuggingFace model card, a blog post announcing a release), use WebFetch to extract details:

WebFetch url="https://..." prompt="Extract: release date, model/library name, key contribution, GitHub/HuggingFace URL if any, parameter count or dataset size if applicable."

Limit WebFetch to at most 3 calls to keep latency in check.

Step 6 — Cross-reference and synthesize

Combine signals:

1. Web items already in the dataset — confirm the survey's top items remain relevant; note if anything new dethrones them.
2. Web items NOT in the dataset — these are candidates the user could also surface via /awesome-japanese-nlp-resources:find-new-resources "$ARGUMENTS"; mention this in section 4 of the output.
3. Directional signals — write 2–4 specific observations about where the field is heading. Examples:
   • "Parameter-count growth: 1B → 7B → 70B for Japanese LLMs since 2024"
   • "Shift from encoder-only (BERT) to decoder-only (LLaMA-derived) base models"
   • "Embedding models specialized for RAG dominate 2025–2026 releases (Ruri-v3, GLuCoSE)"
   • "Multimodal Japanese models (image+text) are emerging but still rare"
4. Gaps — what's missing from the existing list that the web shows exists?

Step 7 — Format the trend report

Language detection rule (apply before writing any output):

• $ARGUMENTS is empty → English
• $ARGUMENTS contains Japanese characters (hiragana / katakana / kanji) → Japanese
• Otherwise → English

Apply the detected language to all headings and prose.

English output template (default):

## 📊 Trend Report for "$ARGUMENTS" (as of ${REPORT_DATE_EN})

### 1. Overview
2–3 sentence summary. "The current focus is X, the latest trend is Y, and the highlight is Z."

### 2. Current Resources (awesome-japanese-nlp-resources)

Top 5 resources:

| # | Resource | Category | Popularity | Summary |
|---|---|---|---|---|
| 1 | [name](url) | category | ⭐N or 📥N | 10–15 word summary |
| 2 | ... | ... | ... | ... |

Category distribution: <Python library: 45, HuggingFace Model: 30, ...>

Maturity comment: 1–2 sentences. Assessment of "mature / growing / sparse."

### 3. Latest Trends (from the web)

- **<date or year-month>** — <finding>. <URL>
- **YYYY-MM** — `XYZ-LLM-7B` released, achieves SOTA on Japanese JGLUE, N downloads on HuggingFace. https://huggingface.co/...
- ... (3–6 items)

### 4. Key Takeaways

- **Direction**: <directional signal 1 and 2>
- **Gaps**: important resources not yet in the list
  - <name> (https://...) — short reason
  - ... (if any)
- **Next step**: run `/awesome-japanese-nlp-resources:find-new-resources "$ARGUMENTS"` to get the full candidate list

### 5. References
(See the Sources section below)

Sources:
- [Title 1](https://...)
- [Title 2](https://...)

Japanese output template (when query is in Japanese):

## 📊 "$ARGUMENTS" トレンドレポート (${REPORT_DATE_JP}時点)

### 1. 概要
2–3 文の要約。「現状の中心は X、最新の動向は Y、注目は Z」のように端的に。

### 2. 既存リソースの現状 (awesome-japanese-nlp-resources)

代表的なリソース top 5:

| # | リソース | カテゴリ | 人気度 | 一言 |
|---|---|---|---|---|
| 1 | [name](url) | category | ⭐N or 📥N | 10–15 字の要約 |
| 2 | ... | ... | ... | ... |

カテゴリ分布: <Python library: 45, HuggingFace Model: 30, ...>

成熟度コメント: 1–2 文。「成熟・拡大期・空白期」の判定。

### 3. 最新トレンド (Web より)

- **<日付 or 年月>** — <発見事項>。<関連 URL>
- **YYYY-MM** — `XYZ-LLM-7B` 公開、日本語 JGLUE で SOTA、HuggingFace で N DL。 https://huggingface.co/...
- ... (3–6 項目)

### 4. 注目ポイント

- **方向性**: <directional signal 1 と 2>
- **ギャップ**: 既存リストに未収録の重要リソース
  - <name> (https://...) — 短い理由
  - ... (もしあれば)
- **次の一手**: `/awesome-japanese-nlp-resources:find-new-resources "$ARGUMENTS"` を実行すると候補一覧を取得できる

### 5. 参考リンク
(下の Sources セクションを参照)

Sources:
- [Title 1](https://...)
- [Title 2](https://...)

Rules:

• Keep total length under ~600 words. The report should be scannable, not exhaustive.
• Each line in section 3 must include a date (or year-month) and a URL. No undated rumors.
• Section 4 must include at least one directional signal and at least one gap (or explicitly note "no obvious gaps").
• The Sources: block at the very end is mandatory — WebSearch results require it.

Step 8 — Edge cases

• No existing results (Step 3 returns empty): Skip section 2's table; in section 2 write "既存リストにこのトピックの直接的なリソースは見つかりませんでした。" Then make section 3 + 4 the focus.
• No recent web results: Note in section 3 that the field is quiet — that's itself a signal.
• Topic is too broad (e.g. just "NLP"): Suggest in section 4 a narrower sub-topic to re-query.