Research Issues

Research current challenges in Japanese NLP 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 challenges. Use the following defaults for the rest of the steps:

• Topic label for output headings: "Japanese NLP Current Challenges" (use "日本語NLP 現状の課題" only when the user's query was written in Japanese)
• Keywords for Step 1 (local dataset survey): japanese nlp, llm, evaluat, benchmark, embed, speech, morpholog

— These broad keywords give a cross-category snapshot for inferring coverage gaps

• WebSearch queries for Step 5: cover challenge-language across multiple sub-fields:
• japanese NLP challenges ${YEAR_NOW} overview
• 日本語 NLP 課題 ${YEAR_NOW}
• japanese LLM limitations evaluation ${YEAR_NOW}
• 日本語 自然言語処理 問題点 未解決 ${YEAR_NOW}
• japanese NLP benchmark error analysis ${YEAR_NOW}
• Report title: ## 🔍 Japanese NLP Issue Report (as of ${REPORT_DATE_EN}) instead of ## 🔍 Issue Report for "$ARGUMENTS" (use ## 🔍 日本語NLP 現状の課題レポート (${REPORT_DATE_JP}時点) only when output language is Japanese)
• Section 1 (Overview): write a broad 3–4 sentence overview covering the most pressing challenges across active sub-fields (LLM evaluation, low-resource domains, embedding quality, speech, 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.

1. Web search phrases (3–5) mixing English and Japanese, explicitly biased to challenge language:

• English-side: challenges, limitations, problems, bottleneck, unsolved, failure modes, error analysis
• Japanese-side: 課題, 問題点, 限界, 未解決, 誤り分析

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  challenges ${YEAR_NOW}`
- `Japanese NLP  limitations`
- `日本語  課題 ${YEAR_NOW}`
- `日本語  問題点 未解決`
- `arxiv japanese  ${YEAR_PREV} ${YEAR_NOW} challenges`
- Optional: ` japanese benchmark error analysis`, ` japanese low-resource`

When a specific high-value URL surfaces (e.g. arXiv abstract describing a failure mode, an evaluation paper, a position paper, a benchmark leaderboard with error analysis), use **WebFetch** to extract details:

WebFetch url="https://..." prompt="Extract: publication date, problem statement (what challenge or limitation does this describe), proposed solution if any, dataset/model used, and a 1-sentence summary of the result. Note if it cites Japanese-specific issues."

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

### Step 6 — Cross-reference and synthesize

Sort the findings from Step 5 into three buckets:

1. **Known challenges** — 3–6 concrete, dated items with URLs. Each should be a *specific* problem ("evaluation suites still over-rely on machine-translated GLUE-style tasks", not "evaluation is hard").
2. **Current efforts / proposed solutions** — 2–4 ongoing projects, papers, or releases attempting to address the challenges in bucket 1. Each needs a URL. If no public efforts surfaced, say so explicitly.
3. **Open gaps** — items in bucket 1 that bucket 2 does NOT cover. These become Section 5's "Still unsolved" line.

Also note any **mismatch with the dataset survey**: web items not in the existing list are candidates for `/awesome-japanese-nlp-resources:find-new-resources`.

### Step 7 — Format the issue 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):**

🔍 Issue Report for "$ARGUMENTS" (as of ${REPORTDATEEN})

1. Overview

2–3 sentences. "The main challenges in X are A, B, and C. Recent work tackled D; E remains open."

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

Top 5 resources covering this topic:

| # | Resource | Category | Popularity | Coverage note | |---|---|---|---|---| | 1 | [name](url) | category | ⭐N or 📥N | what aspect it handles / its known limitation | | 2 | ... | ... | ... | ... |

What's covered well: What's weak / missing:

3. Known Challenges (from the web)

• **** — .
• YYYY-MM — Evaluation benchmarks still over-rely on translated GLUE tasks, missing Japanese-specific phenomena (honorifics, particles). https://arxiv.org/...
• ... (3–6 items)

4. Current Efforts & Proposed Solutions

• **** — .
• ... (2–4 items)

If no public efforts surfaced, write: "No widely-cited public efforts directly addressing the above challenges surfaced in this scan."

5. Key Takeaways

• Most pressing issue:
• Promising direction:
• Still unsolved:
• Next step: run /awesome-japanese-nlp-resources:find-new-resources "$ARGUMENTS" to discover repos that may address these issues

6. References

(See the Sources section below)

Sources:

• Title 1
• Title 2

**Japanese output template (when query is in Japanese):**

🔍 "$ARGUMENTS" の課題レポート (${REPORTDATEJP}時点)

1. 概要

2–3 文の要約。「主要な課題は X, Y, Z。最近 D は進展、E は未解決」のように端的に。

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

このトピックを扱う代表的なリソース top 5:

| # | リソース | カテゴリ | 人気度 | カバー範囲メモ | |---|---|---|---|---| | 1 | [name](url) | category | ⭐N or 📥N | 何を扱っているか / 既知の限界 | | 2 | ... | ... | ... | ... |

カバーできている領域: 未対応 / 弱い領域:

3. 既知の課題 (Web より)

• **** — 。
• YYYY-MM — 評価ベンチマークが翻訳版 GLUE 系に偏り、敬語・助詞など日本語固有現象を捉えきれていない。 https://arxiv.org/...
• ... (3–6 項目)

4. 現状の取り組み / 提案されている解決策

• **** — 。
• ... (2–4 項目)

公開された取り組みが見つからない場合は「上記課題に直接取り組んでいる広く引用された公開プロジェクトは今回の調査では発見できませんでした。」と明記。

5. 注目ポイント

• 最重要課題:
• 有望な方向性:
• 未解決の点:
• 次の一手: /awesome-japanese-nlp-resources:find-new-resources "$ARGUMENTS" で課題に対応するリポジトリ候補を探索

6. 参考リンク

(下の Sources セクションを参照)

Sources:

• Title 1
• Title 2

**Rules:**
- Keep total length under ~700 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 speculation.
- Section 4 must cite at least 2 efforts with URLs, OR explicitly state none were found.
- Section 5 must include all four sub-bullets in order.
- The `Sources:` block at the very end is **mandatory** — WebSearch results require it.

### Step 8 — Edge cases

- **No existing results** (Step 3 returns empty): 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。" (Japanese) or "No direct resources found in the existing list for this topic." (English). Then make Sections 3–5 the focus.
- **No challenges surface from the web**: Note in Section 3 that little challenge discussion was found — the field may be mature or under-discussed. Suggest a broader / adjacent query.
- **Topic is too broad** (e.g. just "NLP"): Add a line in Section 5 suggesting a narrower sub-topic to re-query.

## Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

- **Author:** [taishi-i](https://github.com/taishi-i)
- **Source:** [taishi-i/awesome-japanese-nlp-resources](https://github.com/taishi-i/awesome-japanese-nlp-resources)
- **License:** CC0-1.0
- **Homepage:** https://taishi-i.github.io/awesome-japanese-nlp-resources/

Install and usage instructions live in the source repository linked above.