What prior work taught us

[WaylandYang]

Stacked on #738. A new design page, docs/design/prior-work.md, places each layer of 0044 and 0045 in the research it stands on and lists the pitfalls taken from that research: where each was found, what the design does about it, and what is still open.

What lands

Where each layer stands. The open graph is extractive Open IE (TextRunner, ReVerb, OLLIE, ClausIE, Stanford OpenIE, OpenIE6; CORE, ZORE and SAOKE for Chinese); qualifiers beside the edge are MinIE's attributes and Wikidata's qualifiers; the ontology as a view with alignment later is open-KB canonicalization (Galárraga, CESI, CMVC) plus pay-as-you-go integration (dataspaces, OBDA), with EDC, KGGen, AutoSchemaKG and ODKE+ as the LLM-era pipelines in the same order; a kind word bound to a class is ultra-fine entity typing decided by a judge with a measured position bias; time resolution is temporal tagging (TimeML, HeidelTime, SUTime, TempEval-3, SCATE, the 2025 program-generation framing) over a bitemporal ledger (Snodgrass, Jensen, SQL:2011, the consensus glossary, Clifford's semantics of "now"); a statement with a located quote is KBP provenance and an atomic proposition.

Twenty-six pitfalls, ten in extraction, nine in alignment, seven in time, each with the paper that found it, the design's answer and an open item where the answer is not built. Among the open items: a phrase must read on its own; a modality qualifier per statement; a proper name as the gate for named entities; the quote before the statement in the contract; a class without a definition is not a binding target; classes that name roles are not kind-word candidates; a rejection-rate check in the alignment bench; the queue card shows parents and children; trial-relative mentions kept as offsets from a named local anchor; fiscal periods as their own granularity with the SEC period codes; the cross-precision comparison written down.

What has no precedent. Five decisions held up only by our own benches, and three things the survey looked for and did not find.

References. 150 entries, each verified against its publisher page (ACL Anthology, arXiv, ACM, Springer, CEUR, the SEC and Wikidata pages), grouped by layer. Corrections found on the way and applied: the four-domain anchoring strategies of HeidelTime are in the LREC 2012 paper, not the 2013 journal paper; Chambers 2012 is "Learning from their Time Expressions"; PRAVDA's authors; Dense X Retrieval is EMNLP 2024; the Chinese system is CORE, not COER.

Not in this change

None of the open items; they are listed so the next slices can cite the reason. The README gains one line pointing at the page.

🤖 Generated with Claude Code