brycewang-stanford SIGIR Related Work
At SIGIR, related work is a credibility exam graded by reviewers who may have written the papers you cite. IR is an old field with a living canon: a related-work section that starts retrieval history at BERT signals a visitor from an adjacent community, which primes every other judgment. This skill builds positioning that demonstrates lineage fluency and states the delta mechanically.
The lineage obligation
An IR paper is expected to know which research line it extends and to cite the
line's origin plus its current SIGIR-era frontier. The canonical lines and their
anchors (all venue-verified — see ../../resources/exemplars/library.md):
| Line | Origin anchor | What your delta sentence must contrast |
|---|---|---|
| Probabilistic/lexical ranking | Robertson-era BM25 work; Ponte & Croft's language-modeling approach (SIGIR 1998) | Which term-weighting/estimation assumption you change |
| LM smoothing & estimation | Zhai & Lafferty (SIGIR 2001) | Estimation method, prior, or evidence source |
| Graded evaluation | Järvelin & Kekäläinen (SIGIR 2000; nDCG lineage) | Which validity gap your metric/protocol closes |
| Neural re-ranking / late interaction | ColBERT (SIGIR 2020) | Interaction granularity, storage, or supervision |
| Graph collaborative filtering | NGCF (SIGIR 2019), LightGCN (SIGIR 2020) | Propagation structure or training signal |
You do not cite all lines — you identify yours and prove you know both ends of it.
Venue misattribution: the IR-specific trap
Retrieval's most-cited recent papers are scattered across NLP and ML venues, and misattributing them is a reviewer's one-glance credibility test:
- DPR (dense passage retrieval): EMNLP 2020, not SIGIR.
- ANCE: ICLR 2021, not SIGIR.
- Neural Collaborative Filtering: WWW 2017, not SIGIR (its graph successors NGCF/LightGCN are SIGIR).
- RankNet: ICML 2005. BERT itself: NAACL 2019.
- BM25's standard citations are TREC/SIGIR-era Robertson et al. papers, not "the Anserini paper"; cite the toolkit separately when you mean the implementation.
- Järvelin & Kekäläinen's "Cumulated gain-based evaluation" is TOIS 2002 — the SIGIR 2000 paper is the earlier "highly relevant documents" one; the pair is the classic conference-then-journal split, and conflating them is common and noticed.
# Bibliography audit: flag entries whose venue field is suspiciously vague
grep -nE "booktitle.*(arXiv|CoRR|preprint)" refs.bib # published since?
grep -nE "@misc|howpublished" refs.bib # citing repos as papers?
grep -in "sigir" refs.bib | wc -l # lineage presence check
The delta sentence, mechanized
For each of the 3-5 nearest neighbors (not the 30 nearest keywords), write one sentence with three fixed slots: what it does (mechanism, not adjective), what we do differently (mechanism again), what that buys (measured, or honestly "which we test in §5"). Neighbors are papers a reviewer would confuse with yours — the correct list usually includes at least one paper from the last two SIGIR volumes, one from a sibling venue (ECIR/CIKM/WSDM/WWW/RecSys), and one journal or TREC-track antecedent that pre-dates the neural era. Missing that third slot is how "novel" claims collide with 2009.
Concurrent and arXiv-era work
IR moves at preprint speed while SIGIR reviews at conference speed:
- Sweep arXiv (cs.IR) and the latest SIGIR/ECIR/CIKM accepted-paper lists in the final week before submission; add a "concurrent work" note for anything published after your experiments froze. Reviewers respect the note and punish the silence.
- Cite preprints as preprints; when a preprint you rely on has since been published, cite the published version — mixed-provenance bibliographies age badly in the version of record.
- Anonymity interaction: discussing your own preprint must stay third-person and
must not make the identification trivial (see
sigir-submission).
Positioning failures that cost accepts
- The adjacent-community rewrite: framing a retrieval problem entirely in LLM/ agent vocabulary with no IR literature — reviewers read it as "wrong venue."
- The strawman baseline lineage: citing only pre-neural or untuned versions of the line you claim to beat.
- The enumeration: three paragraphs of "X did A. Y did B." with no contrast work — spends budget without answering "so what is new?"
- The self-citation cluster: heavy self-lineage both risks anonymity and reads as insularity; balance against external anchors.
- TREC blindness: proposing an evaluation methodology or task without checking whether a TREC/NTCIR/CLEF track already operationalized it.
The reviewer-eye test
Before freezing the section, simulate the three reads it will get:
- The 30-second AC skim: do the paragraph topic sentences alone convey "line identified, frontier known, delta stated"? If the deltas hide mid-paragraph, surface them.
- The competitor's read: the author of your nearest neighbor may be R2. Is their paper described accurately enough that they nod before you pivot to the difference? Misdescribing a rival's mechanism converts a fixable delta dispute into hostile review.
- The veteran's read: is there any sentence a 20-year IR researcher would mark "we knew this in the 2000s"? Pre-neural antecedents cited by you defuse what they discover for you.
Placement under the 9-page budget
Related work earns ~0.75 page in a full paper. Two structures work: after the intro (when positioning is the motivation) or before the conclusion (when the method needs no prior-work vocabulary). In short papers, related work compresses into the introduction as positioning sentences with dense citations — a standalone section is usually a page-budget error.
Output format
[Line identified] <which research line, origin + frontier cited y/n>
[Neighbor set] <3-5 papers with mechanism-contrast sentences present y/n>
[Misattribution scan] venues checked against exemplar library: violations <list>
[Era coverage] pre-neural antecedent / recent SIGIR volume / sibling venue: present each
[Concurrent-work note] needed y/n, drafted y/n
[Budget] <fraction of page used> and placement choice