curate-topics

name	curate-topics
description	Autonomously curate candidate keywords into the AI-Trend taxonomy. Given a conference papers CSV, extract candidate keywords, decide for each whether it is noise, belongs to an existing topic, seeds a new emerging topic, or should be parked in the "other" bucket, then apply the decision and assign topics. Use when adding a new conference-year or refreshing topic assignments.

Curate Topics

You are the reasoning core that replaces the manual keyword→topic curation step in the AI-Trend pipeline. The deterministic parts (spaCy candidate extraction, substring assignment) are CLI tools; your job is the judgement in the middle.

Environment

All commands run from the repo root with the project env, isolated from the user site-packages:

RUN="PYTHONNOUSERSITE=1 ./env/bin/ai-trend"

The taxonomy lives in config/taxonomy.json (topic → keyword list) and config/useless_keywords.json (noise blocklist). config/other_keywords.json is the ledger of real-but-unmapped keywords. Treat these as the single source of truth.

Inputs

• CSV — path to the papers file, e.g. data/2025/5_iclr.csv
• CONFERENCE — short name, e.g. iclr
• YEAR — e.g. 2025

Procedure

1. Extract candidates

PYTHONNOUSERSITE=1 ./env/bin/ai-trend candidates "$CSV" \
  --conference "$CONFERENCE" --year "$YEAR" -o /tmp/candidates.json

Read /tmp/candidates.json. It contains:

• existing_topics — the current taxonomy topic labels (your anchor set)
• candidates — [{keyword, count, examples}], most frequent first

2. Decide each candidate

For every candidate, choose exactly one action. Use count and examples to judge what the keyword actually means in context.

Action	When	Requires
existing	The keyword is a clear synonym/variant/instance of an existing topic	topic ∈ existing_topics
new	A genuinely distinct, substantive research theme not covered by any existing topic, with enough volume to matter (rule of thumb: count ≥ 8 and clearly coherent across examples)	topic (a concise lowercase label)
noise	Generic ML vocabulary, methodology words, or anything not a research topic (e.g. "approach", "framework", "evaluation", "method")	—
other	A real, specific concept that is not yet worth its own topic (too niche/rare, or you are unsure). Parks it for later review instead of forcing a bad fit	—

Decision principles (these encode the project's intent — follow them):

• Anchor on the existing taxonomy. Prefer existing over new. Only create a new topic when nothing existing fits; this keeps trends comparable year-over-year.
• Never force a fit. If a keyword does not clearly belong to an existing topic and is not clearly its own theme, use other, not a stretched existing.
• Be conservative with new. A handful of new topics per conference-year is normal; dozens is a sign you are over-splitting — reconsider toward existing/other.
• Keywords are matched case-sensitively against lowercased text. Always emit keywords in lowercase, or they will never match any paper.
• Reuse topic labels exactly as they appear in existing_topics (mind the existing case-variant labels; pick the one that already carries the most keywords).

3. Write the decision

Write /tmp/decision.json:

{
  "decisions": [
    {"keyword": "gaussian splatting", "action": "existing", "topic": "splatting"},
    {"keyword": "state space model",  "action": "existing", "topic": "state space model"},
    {"keyword": "test-time training", "action": "new", "topic": "test-time training"},
    {"keyword": "framework",          "action": "noise"},
    {"keyword": "spectral graph",     "action": "other"}
  ]
}

Include an entry for every candidate — do not silently drop any.

4. Apply and verify

# Preview first
PYTHONNOUSERSITE=1 ./env/bin/ai-trend curate --dry-run /tmp/decision.json
# Then apply (writes config/taxonomy.json, useless_keywords.json, other_keywords.json)
PYTHONNOUSERSITE=1 ./env/bin/ai-trend curate /tmp/decision.json

If curate reports a DecisionError (unknown topic, duplicate keyword, bad action), fix /tmp/decision.json and re-run — do not edit the config by hand.

5. Assign topics

PYTHONNOUSERSITE=1 ./env/bin/ai-trend assign "$CSV"   # writes <CSV>_topics.csv

Output

Report a short summary: counts per action, any new topics created, and how many keywords were parked in other. Surface the new topics explicitly so a human can later confirm they belong in the taxonomy.

Guardrails

• Do not hand-edit config/*.json; go through ai-trend curate so validation runs.
• Do not delete existing topics or keywords; this skill only adds.
• Keep the run reproducible: the decision file is the record of what you chose.