acl-anthology-accepted-venue

# ACL Anthology Accepted Venue Detection ## Scope This is an embedded support skill for `paper-reliability-verifier`, same level as `subskills/dblp-accepted-venue/SKILL.md`. Use it when a paper may belong to venues indexed by ACL Anthology, especially: - ACL, EMNLP, NAACL, EACL, AACL, COLING, CoNLL; - Findings of ACL/EMNLP/NAACL/EACL/AACL; - TACL or Computational Linguistics; - ACL Anthology workshops and shared-task proceedings. Its job is narrow: 1. Use the `acl-anthology` Python package to locate/download the official ACL Anthology metadata repository. 2. Search local ACL metadata by DOI or conservative normalized title/year match. 3. Extract the Anthology ID, URL/PDF URL, volume title, venue ID/acronym/name, and paper metadata. 4. Fill the parent verifier's `accepted_publication` object. 5. Let the parent verifier use local `paper_ranking/*.csv` for CCF/CORE/中科院分区/JCR ranking. This skill answers “is this paper present in ACL Anthology and under what ACL venue/volume?” It does **not** replace local ranking snapshots. ## Documentation Read Primary docs: ```text https://acl-anthology.readthedocs.io/py-v1.1.0/ ``` Relevant documented API concepts: - Install with `pip install acl-anthology` on Python 3.10+. - Instantiate from the official repository: ```python from acl_anthology import Anthology anthology = Anthology.from_repo() ``` - The first call clones/updates roughly 120 MB of metadata; later calls reuse/update the local checkout. - Instantiate from a local data directory when needed: ```python anthology = Anthology(datadir="/path/to/acl-anthology/data") ``` - `anthology.get("2022.acl-long.220")` returns a `Paper`; IDs are hierarchical: collection, volume, paper. - `anthology.get("2022.acl")` returns a `Collection`; `anthology.get("2022.acl-long")` returns a `Volume`. - `Paper` metadata includes `full_id`, `full_id_tuple`, `title`, `authors`, `year`, `doi`, `bibkey`, `pages`, `abstract`, `pdf`, `web_url`, `venue_ids`, `volume_id`, `collection_id`, and `parent`. - `MarkupText` fields such as title/abstract should be converted with `str(...)` to strip Anthology XML markup. - `PDFReference.url` gives `https://aclanthology.org/<id>.pdf`; `Paper.web_url` gives the public paper page. - `Volume` metadata includes `full_id`, `title`, `year`, `venue_ids`, `venues()`, `web_url`, and `get_events()`. - `Event` objects are addressed as `{venue}-{year}`, e.g. `acl-2022`, and can group main conference, Findings, workshops, and colocated events. - `Venue` metadata includes `id`, `acronym`, `name`, `is_acl`, `is_toplevel`, `url`, and associated volume IDs. ## Implementation Notes for Hermes The preferred semantic route is the documented Python package: ```python from acl_anthology import Anthology anthology = Anthology.from_repo(verbose=False) paper = anthology.get("2022.acl-long.220") ``` For verifier-wide title search, the current implementation uses the package to locate `anthology.datadir`, then scans the official XML files under `data/xml`. This is intentional because package-level global indices such as venues/events may require loading the entire latest metadata tree and can fail if the installed `acl-anthology` schema is older than the downloaded data. XML scanning preserves the documented data model while avoiding brittle full-index construction. If a caller already has an Anthology ID, use `anthology.get(id)` directly. If the caller only has a title or DOI, scan XML records. ## Matching Rules Prefer explicit user-provided accepted venue over ACL Anthology. Use ACL Anthology before DBLP when the paper is likely NLP/CL or when title/DOI search in ACL metadata succeeds. Accept a hit only when one of these is true: 1. DOI matches exactly after normalization. 2. Normalized title similarity is high enough, default `>= 0.90`, and year matches when year is known. 3. Normalized title similarity is high enough and it is the best hit when year is unknown. If hits are low-similarity or ambiguous, do not guess. Return `acl_anthology_not_found_or_ambiguous` and keep the top few hits for diagnostics. Absence from ACL Anthology is not evidence that the paper was not published elsewhere. ## Venue Type Mapping Use volume venue metadata: - venue IDs `acl`, `emnlp`, `naacl`, `eacl`, `aacl`, `coling`, `conll`, etc. -> `conference`; - `tacl` and `cl` -> `journal`; - workshop/shared-task volumes are still `conference`-like proceedings for indexing, but must carry non-main-track warnings from the volume title; - ACL Anthology `Findings` volumes are accepted ACL-family publication records; map them to their parent ACL/EMNLP/NAACL/EACL/AACL/COLING/CoNLL venue for ranking; - `short`, `demo`, `student research workshop`, `system demonstrations`, `tutorial abstracts`, `workshop`, `shared task`, `poster`, and `extended abstract` are not main-track full-paper evidence. For ranking lookup, pass: - `venue_name`: the full venue name from venue YAML when available, e.g. `Annual Meeting of the Association for Computational Linguistics`; - `acronym`: ACL/EMNLP/NAACL/EACL/AACL/TACL; - `venue_type`: `conference` or `journal`. ## Output Contract to Parent Verifier Return an `accepted_publication`-compatible object: ```json { "status": "acl_anthology_confirmed | acl_anthology_not_found_or_ambiguous", "venue_name": "Annual Meeting of the Association for Computational Linguistics", "venue_type": "conference", "acronym": "ACL", "evidence_source": "ACL Anthology local metadata: title_year_match", "interface_version": "accepted-publication-v1", "acl_anthology": { "title": "Learned Incremental Representations for Parsing", "title_similarity": 1.0, "year": "2022", "doi": "10.18653/v1/2022.acl-long.220", "anthology_id": "2022.acl-long.220", "bibkey": "kitaev-etal-2022-learned", "url": "https://aclanthology.org/2022.acl-long.220/", "pdf_url": "https://aclanthology.org/2022.acl-long.220.pdf", "venue_id": "acl", "venue_acronym": "ACL", "venue_name": "Annual Meeting of the Association for Computational Linguistics", "volume_title": "Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)", "source_xml": "2022.acl.xml" } } ``` The parent verifier then decides whether to call `match_conference_ranking` or `match_journal_ranking`. ## Operational Caveats - The package may emit `SchemaMismatchWarning` if its bundled schema differs from the latest downloaded data; this is not automatically fatal for lookup. - Avoid building global venue/event/person indices unless needed; they can require parsing all XML files. - ACL Anthology includes workshops and Findings. Findings is strong evidence of publication and can be strong venue evidence after parent-venue rank lookup; workshops still need non-main-track caveats. - Use the volume title and Anthology ID to preserve track information. - Do not scrape the website for metadata when the package/local XML data already provides it. ## Verification Checklist - [ ] `acl-anthology` package can instantiate `Anthology.from_repo()` or `ACL_ANTHOLOGY_DATA_DIR` points to a valid `data/` directory. - [ ] Selected hit has DOI match or high title similarity. - [ ] Year was checked when known. - [ ] `venue_name`, `acronym`, and `venue_type` were mapped into `accepted_publication`. - [ ] ACL Anthology Findings signals from `volume_title` were mapped to the parent venue for ranking; workshop/short/demo/non-main-track signals were preserved as warnings. - [ ] The accepted venue was passed to local ranking CSV lookup; ranking was not inferred from ACL Anthology alone.