sigir-topic-selection
GitHub辅助判断论文是否属于SIGIR及选择具体Track。核心通过测试检索是否为贡献主体来筛选，并根据内容类型（方法、短发现、资源等）路由至Full/Short/Resources等轨道，或推荐ECIR/CHIIR等替代会议。
Trigger Scenarios
Install
npx skills add brycewang-stanford/Awesome-Journal-Skills --skill sigir-topic-selection -g -y
SKILL.md
Frontmatter
{
"name": "sigir-topic-selection",
"description": "Use when deciding whether a project belongs at SIGIR and in which of its tracks — testing whether retrieval is the contribution or the plumbing, forking across full\/short\/resources\/reproducibility\/perspectives\/industry, and routing misfits to ECIR, CIKM, WSDM, CHIIR, ICTIR, RecSys, TheWebConf, or NLP\/ML venues."
}
SIGIR Topic Selection
Two decisions, in order: (1) is this an information-retrieval contribution at all, and (2) which SIGIR track owns it. Both are cheaper to get right in October than to learn from reviews in March. SIGIR 2026's verified lineup — Full, Short, Perspectives, Resources, Reproducibility, Low Resource Environments, and Industry tracks — makes the second decision genuinely multi-way.
Test 1: is retrieval the contribution or the plumbing?
The LLM era floods SIGIR with papers where retrieval is present but incidental. The discriminating question: if you swapped the retrieval component for an oracle (or a black box), would the contribution survive?
- Survives → the contribution is the downstream system (agent, QA, generation quality); route to an NLP/ML venue and cite IR work properly.
- Dies → the contribution is retrieval: ranking quality, efficiency, evaluation validity, user search behavior, or the retrieval-specific part of RAG (candidate quality, index design, retrieval evaluation). SIGIR-shaped.
Corollary signals of SIGIR fit: the natural headline table has metric@cutoff
columns on named collections; the natural reviewers are IR-evaluation people; the
natural related work includes SIGIR/TOIS/TREC lineage (see sigir-related-work).
Test 2: the track fork
| Your strongest honest sentence starts with... | Track | 2026 anchor |
|---|---|---|
| "We propose/analyze a method and show on N collections..." | Full papers | ≤9 pp excl. refs, double-blind |
| "We report one focused finding..." | Short papers | ≤4 pp excl. refs, own OpenReview group |
| "We built and share a corpus/toolkit/judgments..." | Resources | ≤6 pp + refs, single-anonymous |
| "We reproduced (or failed to reproduce) published results..." | Reproducibility | Own track in 2026; details 待核实 |
| "We argue the field should..." | Perspectives | Details 待核实 |
| "We deployed retrieval in production and learned..." | Industry (SIRIP lineage) | Details 待核实 |
| "Search under compute/connectivity/language constraints..." | Low Resource Environments | Feb 5 abstract, Apr 2 notify (2026) |
Fork rules of thumb:
- A full paper needs a rounded evidence program; a strong single finding with one collection and one deep analysis is a better short paper than a thin full paper.
- The dataset built for your method paper is a separate routing decision — the cross-track double-submission ban means the dataset goes to Resources only if it is not the method paper's claimed contribution.
- Reproducibility is for studies about prior results, not for "our paper has good
reproducibility" (that's
sigir-reproducibilityhygiene).
Neighbor routing when SIGIR isn't the answer
Contribution center of gravity -> Better first target
---------------------------------------------------------------
Interactive/user-study search behavior -> CHIIR (then SIGIR full)
Formal/theoretical IR models -> ICTIR (SIGIR-affiliated)
Recommender systems end-to-end -> RecSys (SIGIR takes rec too,
esp. retrieval-flavored: NGCF/LightGCN lineage)
Web platform/graph/user data emphasis -> TheWebConf (WWW) or WSDM
IR + knowledge management/databases -> CIKM
European timing or repro emphasis -> ECIR
Language-technology contribution -> ACL/EMNLP (retrieval as application)
Efficiency/systems below the ranking -> systems venues; keep an IR-eval slice for SIGIR
Evaluation campaign participation -> TREC/NTCIR/CLEF tracks (not archival
substitutes: they *feed* SIGIR papers)
TREC deserves the explicit note: it is an evaluation campaign, not a proceedings competitor. Building on (or contributing to) a TREC track is often the strongest possible SIGIR pipeline — campaign participation year 1, analysis paper year 2.
Borderline cases, adjudicated
- RAG for a downstream task, with a retrieval ablation: if the ablation is the deepest result, extract it into a retrieval-evaluation paper for SIGIR and let the system paper go to NLP — two focused papers beat one blurry one.
- An LLM that rewrites queries: SIGIR-shaped if evaluated as a retrieval intervention (per-topic deltas on standard collections), NLP-shaped if evaluated on rewrite quality.
- A new efficiency trick inside an index structure: SIGIR wants the quality-cost frontier; a pure throughput result without ranking-quality context fits systems venues better.
- A user study of searchers: CHIIR first unless there is a measurable system-side consequence (a re-ranker, an interface signal) evaluated IR-style.
- An evaluation critique with no fix: Perspectives material, or pair it with a demonstrated protocol improvement and aim for full papers.
Timing and prestige honesty
- SIGIR's single annual cycle (January-pattern deadline) means a misrouted
submission costs a year unless the fallback ring is pre-planned
(
sigir-workflow). Route by contribution type, not venue rank. - Short papers at SIGIR are archival, cited, and respected — routing a focused finding there is a strategy, not a demotion.
- If the honest evidence today is one collection, no significance testing, and an unfinished ablation, the choice is not "which track" but "which cycle" — a rushed January submission burns the year anyway.
One more timing lever: the staggered secondary-track calendar (Resources and LRE ran weeks behind full papers in 2026) sometimes lets a January full-paper miss become a same-year February resource or reproducibility submission — but only if the contribution genuinely re-frames, not as a costume change reviewers see through.
Scope vocabulary check
SIGIR's own topic language (2026 full-papers page): search and ranking, queries and query analysis, web search, retrieval models and ranking, efficiency and scalability (indexing, crawling, compression), recommender systems — plus the community's standing breadth: evaluation, conversational search, and retrieval-augmented systems. If the abstract cannot be written naturally in this vocabulary, revisit Test 1.
Output format
[IR-contribution test] survives-oracle-swap: yes (route out) / no (SIGIR-shaped)
[Track] full / short / resources / reproducibility / perspectives / industry / LRE
[Fork rationale] <one sentence: why this track owns the strongest honest claim>
[Cross-track conflicts] dataset/method ownership resolved y/n
[Fallback ring] <next venue if rejected, with deadline>
[Cycle honesty] submit this cycle / build one more quarter
Version History
- 9f86f09 Current 2026-07-19 17:36
