ingest-proceedings

Ingesting Anthology proceedings

Templatized workflow for ingesting one or more proceedings (ACLPUB or aclpub2) into the Anthology with bin/ingest.py, one volume at a time, committing each. Works for main/standalone events and colocated workshops — the pre-flight and venue rules are identical; only the run flags and event handling differ (§3).

Fill in the template variables:

• {{REPO}} — the acl-anthology checkout (e.g. /Users/you/code/acl-anthology)
• {{WS}} — directory holding the proceedings tarballs/clones (e.g. .../ws)
• {{PARENT_EVENT}} — workshops only: parent event id to colocate under (e.g. acl-2026). Omit for a main/standalone event.
• {{PY}} — the project interpreter, usually {{REPO}}/.venv/bin/python (run bin/ingest.py with this; fixedcase is a local module under bin/)

0. Golden rules

• Only use the public acl_anthology API; never parse the XML by hand.
• Never git add/commit unless the human asked. (This task explicitly does.) When committing, stage only data/ — a failed ingest can leave a stray data/yaml/venues/*.yaml that git add -A/git add data will otherwise sweep into the next commit. Verify git status after failures.
• Ingest one volume per invocation so each can be committed/reverted alone. bin/ingest.py calls save_all() only at the very end, so a failed run saves nothing — but PDFs may already be copied to ~/anthology-files (harmless).

1. Find each ingestion root

The root is the directory passed to bin/ingest.py. Detection (detect_ingestion_format):

• aclpub = meta file + cdrom/ dir
• aclpub2 = (conference_details.yml or inputs/conference_details.yml) and (papers.yml or inputs/papers.yml)

The most reliable signal for an aclpub2 root is the watermarked_pdfs/ directory. Locate it and confirm inputs/ sits beside it:

for d in {{WS}}/ws_*/; do
  find "$d" -type d -name watermarked_pdfs | while read wm; do
    root="${wm%/watermarked_pdfs}"
    echo "$d -> ROOT=$root  conf=$([ -e "$root/inputs/conference_details.yml" ] && echo y) papers=$([ -e "$root/inputs/papers.yml" ] && echo y)"
  done
done

Ignore build/ matches (build artifacts). If watermarked_pdfs is missing or misnamed (e.g. watermarked_pdf, or under outputs/), fix with a symlink in the tarball (non-destructive), so the script finds ROOT/watermarked_pdfs/ and ROOT/watermarked_pdfs/0.pdf (frontmatter):

ln -sfn watermarked_pdf            {{WS}}/ws_x/watermarked_pdfs      # misnamed
ln -sfn outputs/watermarked_pdfs   {{WS}}/ws_y/watermarked_pdfs      # wrong dir
ln -sfn outputs/proceedings.pdf    {{WS}}/ws_y/proceedings.pdf       # book pdf

If the dir exists but individual PDFs are missing (often organizer-added "Overview"/shared-task papers the watermarking step skipped), the watermarked file simply was never produced. The un-watermarked source usually still exists elsewhere in the tarball (e.g. <repo>/.../papers/<id>.pdf); copy it into ROOT/watermarked_pdfs/<id>.pdf so the paper can be ingested:

# which paper ids in papers.yml have no watermarked PDF?
comm -23 <(grep -oE '^  id: [0-9]+' inputs/papers.yml | grep -oE '[0-9]+' | sort -nu) \
         <(ls watermarked_pdfs/*.pdf | xargs -n1 basename | sed 's/.pdf//' | sort -nu)
cp <repo>/.../papers/77.pdf ROOT/watermarked_pdfs/77.pdf   # supply the missing one

2. PRE-FLIGHT validation (do this BEFORE ingesting — organizers get these wrong)

Read each conference_details.yml and check anthology_venue_id, volume_name, and book_title. Cross-reference data/yaml/venues/*.yaml.

2a. Venue id must be the workshop's OWN venue — not the parent conference

A common error is anthology_venue_id: ACL (the parent). That would ingest the workshop into the YYYY.acl collection and clobber sibling workshops sharing a volume name. Replace it with the workshop's real venue slug.

2b. Don't invent a new venue for an existing or joint venue

The venue slug = slugify(anthology_venue_id.replace("-","")) and must match ^[a-z][a-z0-9]+$ — letters and digits ONLY, no hyphens/+/spaces (the human-readable acronym may contain more). If a same-named venue already exists under a different slug, the id is wrong — use the existing slug. Examples:

• NLP+CSS → slugifies to nlp-css (has a hyphen → invalid), and the venue is nlpcss → set id to nlpcss.
• LT-EDI → ltedi (hyphen stripped) — fine.
• A joint workshop (e.g. CODI-CRAC) must NOT create a merged codicrac venue. Ingest under the primary venue (codi) and add the secondary as an extra <venue> tag (crac). A volume may carry multiple <venue> tags (e.g. codi + crac + ws). Delete any merged venue file that leaked in.

2b-bis. Findings

The Findings volume (usually already ingested with the main conference) must use findings as the venue and one of acl/eacl/naacl as the volume name (e.g. 2026.findings-acl, 2026.findings-eacl). Verify this if present.

2c. New-venue names: strip ordinals, years, and conference prefixes

Venue name: in data/yaml/venues/<slug>.yaml must be the timeless name. Remove all of these (organizers routinely add them):

• ordinals: "The 9th ...", "First/Second ..."
• years (2- or 4-digit): "... 2026", "... '26"
• conference prefixes: "ACL 2026 Workshop on ..."
• a trailing acronym in parens: "... (CDL)" / "... (EvalEval)"
• the literal venue name acl (or any parent-conference name) is never a valid workshop venue name

Examples:

• The 1st Workshop on Computational Developmental Linguistics (CDL) → Workshop on Computational Developmental Linguistics
• ACL 2026 Workshop on Evaluating Evaluations (EvalEval) → Workshop on Evaluating Evaluations
• Proceedings of the 1st Workshop on X (X 2026) → Workshop on X

2d. "Nth workshop" new venue → search before creating

If a new venue looks like "the 9th workshop on X", search the Anthology first:

• (a) wrong venue id: an existing venue already covers X under another slug → correct anthology_venue_id to that slug (don't create a new venue).
• (b) prior instances under old-style IDs (e.g. W19-xxxx) that were never <venue>-tagged → tag those old volumes with the venue, and still remove the ordinal from the new venue name. Only if neither holds is it genuinely new (then just strip the ordinal, 2c).

# search for an existing venue / prior instances of "Event Extraction"
grep -rilE "event extraction" data/yaml/venues/
grep -rlE "Event Extraction" data/xml/ | tail

2e. Volume-name consistency

1 or main is fine for a single volume. If a venue has multiple volumes this year (e.g. a workshop + its shared task, often two repos sharing one venue), they must be consistently named: either both numbers (1, 2) or both names — never a number mixed with a name. Also volume_name must match ^[a-z0-9]+$ (lowercase): Workshop is invalid → use 1/main/etc.

2f. Copyright forms must NOT become <attachment> tags

Organizers sometimes attach the signed copyright-transfer form as a paper attachment. These must never be published as an <attachment>. bin/ingest.py already skips any attachment whose type contains the lowercase substring copyright (if "copyright" in attachment["type"]). The gap: that test is case-sensitive and type-based, so a form typed Copyright/CTA/License, or mislabeled (e.g. Supplementary Material), slips through. Pre-flight, scan the attachment types/filenames and normalize anything that is really a copyright form:

grep -niE "copyright|transfer|cta\b|license" inputs/papers.yml

For each true copyright form, either drop the attachment entry from papers.yml or set its type: to lowercase copyright so the built-in filter removes it. Then confirm post-ingest (§4).

3. Ingest + commit, one at a time

Pick the mode based on what you're ingesting:

3a. Main / standalone event (e.g. the conference itself)

Do not pass -w or --parent-event.

cd {{REPO}}
{{PY}} bin/ingest.py "<ROOT>" \
  --event-title "64th Annual Meeting of the Association for Computational Linguistics" \
  --event-location "San Diego, California, United States" \
  --event-dates "July 2–7, 2026" \
  --event-website "https://2026.aclweb.org" \
  --event-handbook "/path/to/acl-2026-handbook.pdf"
git add data && git commit -m "Ingested <name>"     # only on success

The event options create or update the collection's <event> block, which populates the event page and gives colocated workshops something to attach to. --event-handbook copies the supplied PDF to ~/anthology-files/handbooks/<venue>/<collection-id>.handbook.pdf by default and adds a handbook link; override the root with --event-files-dir. Colocated workshops later append <colocated> entries without replacing the event metadata.

3b. Colocated workshop

Pass -w (adds the ws venue tag) and --parent-event (colocates under the parent's <event>).

cd {{REPO}}
{{PY}} bin/ingest.py "<ROOT>" -w --parent-event {{PARENT_EVENT}}
git add data && git commit -m "Ingested ws_<name>"  # only on success

--parent-event colocates the volume: the parent collection XML (data/xml/YYYY.<parent>.xml) gains/extends an <event id="{{PARENT_EVENT}}"><colocated><volume-id>…</volume-id> block. (If the parent event isn't yet explicit, the script promotes it so the block is written.)

Both modes

Stage only data/. A driver that ingests a list and commits each on success, logging failures without committing (and git clean -fdq data to drop a stray venue file from a failed run), is reproducible — but keep the pre-flight fixes (§2) ahead of it.

4. Post-ingest fixes (MANDATORY — do not skip, and ACT, don't just inspect)

• Joint venue tag: add the secondary <venue> to the in

---

Content truncated.