Research

Research Japanese NLP trends and challenges for the user's topic by combining the bundled dataset with the latest web information.

Claude Code and Codex

This skill is shared by the Claude Code and Codex versions of the plugin. The steps are the same in both tools; only these details differ:

• Topic — Claude Code: the arguments of /awesome-japanese-nlp-resources:research, appended at the end of this skill as ARGUMENTS: …. Codex: the user's message that invoked $awesome-japanese-nlp-resources:research, minus that $… mention. If the skill was picked automatically rather than invoked by name, use the user's request as the topic.
• Plugin root — Claude Code: ${CLAUDE_PLUGIN_ROOT}. Codex: the directory two levels above this SKILL.md (use its absolute path).
• Shell — run the commands below with Claude Code's Bash tool or Codex's shell tool. Copy each Python script in full and run it as written, changing only its placeholders (RESOURCES_PATH, the keyword list) — don't shorten it, drop passes, or alter its scores and thresholds.
• Web — Claude Code: WebSearch to search and WebFetch to read a page. Codex: the built-in web search tool (search, then open the page). Do not use the gh CLI, curl, or other network commands from the shell.
• Commands — write any command you show the user in the current tool's form: /awesome-japanese-nlp-resources:<skill> in Claude Code, $awesome-japanese-nlp-resources:<skill> in Codex.

Section 1 of the report ("What's already in awesome-japanese-nlp-resources") must come from the bundled data. If the data file can't be read (for example, shell commands are blocked or fail to start), say so in that section instead of filling it from memory or web search.

Instructions

Preamble — Establish the current date

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

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

Substitute these wherever this skill writes ${YEAR_NOW}, ${YEAR_PREV}, ${REPORT_DATE_EN} or ${REPORT_DATE_JP}. Do not hardcode years.

Step 0 — Validate input

If the topic is empty or blank, treat it as a request for a general overview of the current Japanese NLP landscape (both trends and challenges). Use the following defaults for the rest of the steps:

• Topic label for output headings: "Japanese NLP Overall Landscape" (use "日本語NLP 全体動向" only when the user's query was written in Japanese)
• Keywords for Step 1 (local dataset survey): llm, bert, embed, speech, morpholog, translat, evaluat, benchmark — short stems, since Step 3 matches by literal substring and a multi-word phrase like japanese nlp rarely occurs verbatim in a description — This broad set gives a cross-category snapshot of the most popular resources and of coverage gaps
• Web searches for Step 5: cover both trend and challenge language across multiple sub-fields:
  • japanese NLP trends ${YEAR_NOW} overview
  • 日本語 NLP 最新動向 ${YEAR_NOW}
  • japanese LLM embedding benchmark ${YEAR_NOW} github
  • japanese NLP challenges ${YEAR_NOW} overview
  • 日本語 NLP 課題 ${YEAR_NOW}
  • japanese LLM limitations evaluation ${YEAR_NOW}
• Report title: ## 🔭 Japanese NLP Research Report (as of ${REPORT_DATE_EN}) instead of ## 🔭 Research Report for "<topic>" (use ## 🔭 日本語NLP リサーチレポート (${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) and the most pressing shared challenges

Step 1 — Interpret the topic

Translate the topic intent to English keywords for the local dataset survey. Aim for 4–6 keywords, using the same stem + tool-name conventions as the search skill (morpholog, embed, classif, translat, recogni, plus well-known tool names for the domain).

Step 2 — Locate the data file

The data file ships with the plugin at data/resources.json under the plugin root (see "Claude Code and Codex" above). Resolve its absolute path, falling back to a scoped search only if the install is unusual:

PLUGIN_ROOT="${CLAUDE_PLUGIN_ROOT}"  # Codex: replace with the plugin root, two levels above this SKILL.md
RESOURCES_PATH="$PLUGIN_ROOT/data/resources.json"
[ -f "$RESOURCES_PATH" ] || RESOURCES_PATH="$(find "${CODEX_HOME:-$HOME/.codex}/plugins" "${HOME}/.claude/plugins" -type f -name resources.json 2>/dev/null | grep "awesome-japanese-nlp-resources/" | head -1)"
echo "RESOURCES_PATH=$RESOURCES_PATH"

Use the resulting absolute RESOURCES_PATH wherever Step 3 opens the data file — write the path itself into the script, since shell variables may not persist between commands.

The plugin also ships data/multilingual_resources.json (same item format) listing multilingual libraries, models, and datasets (GitHub repositories) that also support Japanese, from docs/multilingual.md. The scripts below load it automatically when it exists; its items have categories like Multilingual (Speech recognition).

Step 3 — Survey the existing dataset (inline Python)

Do not read the data file directly (no Read tool, cat, or head) — it is about 660 KB. Run the scoring in a single shell command using Python.

python3 << 'EOF'
import json, os

with open("RESOURCES_PATH") as f:    # absolute path from Step 2
    data = json.load(f)
multilingual_path = os.path.join(os.path.dirname("RESOURCES_PATH"), "multilingual_resources.json")
if os.path.exists(multilingual_path):
    with open(multilingual_path) as f:
        data += json.load(f)

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

results = []
for item in data:
    if item.get("status") == "not_found":
        continue

    n = item.get("n", "").lower()
    d = item.get("d", "").lower()
    s = " ".join(item.get("s") or []).lower()
    c = item.get("c", "").lower()
    al = " ".join(item.get("al") or []).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 kw in al:      text_score += 10

    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, item))

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

# Category distribution across ALL matches (not just the top slice) — used to
# spot which resource types dominate and, by inference, which are thin.
from collections import Counter
cat_counts = Counter(item["c"] for _, item in results)

print(f"=== {len(results)} local matches; top 10 shown ===")
for combined, item in results[:10]:
    st = item.get("st", 0) or 0
    dl = item.get("dl", 0) or 0
    print(f"score={combined:.1f} st={st} dl={dl}")
    print(f"  n={item['n']}")
    print(f"  u={item['u']}")
    print(f"  c={item['c']}")
    print(f"  s={item.get('s','')}")
    print(f"  d={item.get('d','')[:120]}")
    print()

print("=== category distribution (all matches) ===")
for cat, count in cat_counts.most_common(10):
    print(f"  {count:4d}  {cat}")
EOF

Substitute RESOURCES_PATH with the absolute path from Step 2 and keywords with your keywords list from Step 1.