Skip to content

World's first open-source enterprise world model.

56.9 KB
Raw Blame

What prior work taught us

The three layers of [0044] and the time resolution of [0045] were not invented here. Each stands on a line of research that is fifteen to forty years old, and each of those lines has already met the failures we meet on our own corpora. This file places every layer in that literature and lists the pitfalls we took from it: who found it, what we do about it, and what is still missing. It changes when a design file changes for a reason the literature had already given, or when a new pitfall is taken. A paper is cited as [Fader 2011]; the list at the end gives venue, year and a link for every one, and every entry was checked against its publisher page before it was written here.

Where each layer stands

The open graph is extractive Open Information Extraction. A statement in the document's own words, with the relation phrase as written and no schema, is the programme Banko et al. opened [Banko 2007] and the Etzioni group carried through ReVerb [Fader 2011] and OLLIE [Mausam 2012]; ClausIE [Del Corro 2013], Stanford OpenIE [Angeli 2015], the supervised turn [Stanovsky 2018] and OpenIE6 [Kolluru 2020] are the same programme with better parsers, and three surveys cover it to 2024 [Niklaus 2018, Zhou 2022, Liu 2024]. Pei et al. draw the line we sit on: extractive OpenIE keeps surface forms, abstractive OpenIE may infer [Pei 2023]. For Chinese the reference points are CORE [Tseng 2014], ZORE [Qiu 2014], the SAOKE corpus behind Logician, whose annotation scheme has a class of descriptive facts that matches our described things [Sun 2018], and Title2Event, the largest manually annotated set of Chinese open triples [Deng 2022]. Our departures from that tradition are three: a language model reads the chunk under a JSON contract instead of a parser; every statement carries its own verbatim quote located by character offsets; the run is at temperature 0 with a judge that counts what the text did not say (extraction). Two recent designs converge on the same shape from the other side: TRACE-KG keeps a raw and a canonical predicate, qualifiers and a justification excerpt on every relation [Abolhasani 2026], and Han and Lai argue for natural-language relations over a minimal structural backbone [Han 2026].

Qualifiers beside the edge are MinIE's attributes and Wikidata's qualifiers. OLLIE first separated attribution and condition from the triple [Mausam 2012]; MinIE moved polarity, modality, attribution and quantity out of the relation phrase into annotations [Gashteovski 2017]; NestIE kept nested propositions as such [Bhutani 2016]; DisSim's split of a sentence into core and context linked by a rhetorical relation is the same cut made by rule [Niklaus 2019]. In the ledger the idea is Wikidata's qualifier on a statement [Vrandečić 2014] and the hyper-relational fact of HyperRED, whose 44 qualifier types are the comparison point for our role words [Chia 2022]; Text2NKG lists the four incompatible ways an n-ary fact is serialised [Luo 2024], and ours is the hyper-relational one. statement_qualifiers keyed by the document's role word is that model with the role word left as written [0037, 0044].

The ontology as a view, with alignment later, is open-KB canonicalization plus pay-as-you-go integration. Canonicalizing the noun phrases and relation phrases of an open knowledge base after extraction is the problem Galárraga et al. posed [Galárraga 2014] and CESI [Vashishth 2018a], CUVA [Dash 2021], JOCL [Liu 2021] and CMVC [Shen 2022] took up; Yang and Curry survey that line and find relation-phrase canonicalization its least solved part [Yang 2024]. Mapping open extractions onto a fixed schema was done for domain relations [Soderland 2010], for KBP slots [Soderland 2013, Angeli 2015], for DBpedia [Dutta 2015, Gashteovski 2020] and for ConceptNet [Romero 2023]. The LLM-era pipeline closest to [0044] is EDC, which extracts open triples, writes a definition for each extracted relation and then lets the model select a schema element or reject them all [Zhang 2024]; KGGen [Mo 2025], iText2KG [Lairgi 2024], AutoSchemaKG [Bai 2025], SAC-KG [Chen 2024b], Docs2KG [Sun 2025] and ODKE+ [Khorshidi 2025] are the same order of operations at other scales, and Bian's survey names the schema-based and schema-free paradigms they sit between [Bian 2025]. The stance that data is stored first and semantics added as it pays for itself is the dataspaces programme [Franklin 2005, Halevy 2006, Madhavan 2007]; an ontology that is a view over stored data through mappings is ontology-based data access [Poggi 2008, Xiao 2018]. The knowledge-graph surveys treat an emergent schema as a normal third mode beside a prescribed one [Hogan 2021, Weikum 2021].

A kind word bound to a class is ultra-fine entity typing, decided by a judge that has been shown its own bias. Choi et al. defined types as the free noun phrases a text uses, with head words as the supervision [Choi 2018]; our kind word is that phrase kept verbatim, and the binding is its mapping to the ontology's class, the step Dai and Zeng treat as its own task [Dai 2023]. Typing by description alone [Obeidat 2019], by enriched classes [Ouyang 2024] and top-down through the hierarchy [Komarlu 2024] are the three moves the binding call makes. The candidate-then-judge shape, with an explicit "none", is the shape of LLM ontology matching [He 2023, Hertling 2023, Babaei Giglou 2024, Qiang 2024, Song 2025]. Deciding with two votes and the candidate order reversed on the second comes from the position bias measured in LLM judges and choosers [Zheng 2023, Wang 2024, Pezeshkpour 2024, Zheng 2024]; agreement across samples as a confidence signal is self-consistency [Wang 2023a].

Time resolution is temporal tagging over a bitemporal ledger. The words of a mention and their value are TimeML's TIMEX3 [Pustejovsky 2003]; computing the value against a reference time by rule is what HeidelTime [Strötgen 2012, Strötgen 2013] and SUTime [Chang 2012] do, and TempEval-3 measured how much harder the value is than the span [UzZaman 2013]. The split we made, a model that returns an interpretation and code that computes from it, is TimeNorm's grammar given an anchor [Bethard 2013], SCATE's compositional operators [Bethard 2016a] and, closest of all, the 2025 framing of normalization as generating a program that a deterministic library executes [Su 2025a]; the surveys are [Leeuwenberg 2019, Zhong 2023, Su 2025b]. Dating a document is its own task [Chambers 2012, Vashishth 2018b]; we transcribe the date the document states instead of classifying. Facts with validity intervals are YAGO2's model, which "never assigns an ill-defined time" [Hoffart 2013], and the temporal scoping line [Ling 2010, Wang 2011, Talukdar 2012]; the LLM-era temporal graphs separate observed time from valid time the same way [Rasmussen 2025, Lairgi 2025]. The two axes of [0019] are the valid time and transaction time of bitemporal databases [Snodgrass 1999, Jensen 1999, Kulkarni 2012], in the words of the consensus glossary [Jensen 1998]; the precision columns are time granularity [Bettini 2000]; attested_to is a now-relative value in the sense of Clifford et al. [Clifford 1997]. The clinical corpora are where relative time was studied hardest [Sun 2013, Styler 2014].

A statement with a located quote is KBP provenance and an atomic proposition. TAC KBP required a document offset as provenance for every slot fill; FActScore [Min 2023] and Dense X Retrieval [Chen 2024a] made the atomic proposition the unit of checking and of retrieval, and the quote-then-answer models [Menick 2022, Huang 2024, Zhang 2025] made the verbatim quote the unit of trust. Our statement is that unit with its provenance attached at write time [0001, 0044].

The pitfalls we took

Each item: where it was found, what we do, what is still missing. Items marked open are not yet built.

Extraction

  1. Uninformative relation phrases. ReVerb named the two error classes of open extraction, incoherent phrases and uninformative ones, its example being "made a deal with" cut to "made" [Fader 2011]; the fact-level re-annotation of BenchIE rejected vague relation phrases outright [Lamarche 2024]; labelling chunks instead of tokens is one remedy [Dong 2023a]. Our "各地区 深入 企业" is ReVerb's second class. We keep the whole sentence as the quote, so nothing is lost; the contract states that a phrase is the verb with the words that belong to it, so that the statement reads on its own, and the judge reports "reads alone" beside "not stated" (extraction, #743).
  2. A triple the sentence did not assert. OLLIE showed how many extractions are conditioned, attributed or hypothetical in the sentence and false without that context [Mausam 2012]; speculation is a property of the tuple, not of the sentence [Dong 2023b]. Qualifiers hold the context [0037, 0044]; a statement the passage requires, plans, expects or makes conditional carries a mood qualifier in the passage's own words, a reporting verb is not a mood, and alignment never materializes a statement with a mood as a typed fact (extraction, #745). Open: the subject of a directive sentence in a notice whose addressee is outside the chunk; a rule naming the addressee was tried and withdrawn because it emptied a fifth of a filing's chunks.
  3. Minimal is not the same as self-contained. MinIE says minimization loses context in its own evaluation [Gashteovski 2017]; the neural systems went the other way and over-include [Fatahi Bayat 2022]; CaRB [Bhardwaj 2019], WiRe57 [Léchelle 2019], LSOIE [Solawetz 2021] and BenchIE [Gashteovski 2022] disagree on how compact a fact should be, and rankings flip with the matcher. The fact-checking side reached the same place: fully atomic facts are not the right unit, decontextuality and minimality pull against each other [Gunjal 2024, Choi 2021], and more decomposition is not monotonically better [Hu 2025]. Our rule: one statement is one complete proposition; coordinated objects split; a verb chain sharing one object does not (#743).
  4. The decomposition is itself a model output. Change the decomposition prompt and every downstream number moves [Wanner 2024]; a segmentation pre-pass can drop entities for strong models [Pommeret 2026]. The contract is versioned in code and each bench round names its contract extraction. Open: entity recall measured with and without any pre-pass before one is added.
  5. Coordination and verb chains are the yield on dense sentences. Splitting conjunctive sentences first gave up to 1.8× yield with a precision gain [Saha 2018]; a coordination analyser inside the extractor is what moved OpenIE6 [Kolluru 2020]; it can be done without a parser [Wang 2023b]; split-and-rephrase is the general task [Narayan 2017]; irrelevant context is a named failure of LLM OpenIE [Ling 2023]. The judge and the coverage script measure the dense NVDA sentences separately (#731). A structural count of the money and percentage figures that no statement of their chunk carries (scripts/bench/figures.mjs, #748) found 2 of 83 in prose on the 25-document batch, 2 of 35 on the NVDA releases and 0 of 451 in their tables, with reasoning on at the endpoint; a second-look pass waits for a corpus that loses more.
  6. Common nouns as entities. On Wikipedia text only 42% of open-extraction arguments are named entities [Gashteovski 2019]; canonicalization meets the same string with different referents first [Galárraga 2014]; the opposite failure, one referent under many strings, is the sparsity KGGen was built against [Mo 2025], and coreference before extraction was the largest lever on duplicate nodes in CORE-KG [Meher 2025]. "企业" merged across documents is a hub that means nothing. A described thing exists only when a statement points at it [#736]. A thing is named only by a proper name or a fixed term that means the same in any document; a role or a generic phrase whose referent the passage decides is described, scoped to its document and never merged by string; a generic phrase in an adjunct stays words in the qualifier (#743). The gate is a term gate, not a proper-name gate: a disease, a drug or an indicator merges across documents by design.
  7. A closed schema makes the model invent. Prompting with a fixed relation set causes hallucination [Jiang 2024]; a model forced onto an ontology conforms almost perfectly and is still wrong [van Cauter 2024]; the dominant "error" of LLM extraction against gold is spans the gold never annotated [Han 2023]; exact match penalises paraphrase, so the honest evaluations are human [Wadhwa 2023, Li 2023]. The open contract uses the document's words, the quote must occur in the chunk, the run is at temperature 0, and the judge asks whether the text states it rather than whether it matches a gold form; "not stated" stays under 2% on every corpus (extraction). Hallucination and omission are reported as two numbers, as Ghanem and Cruz ask [Ghanem 2024]: the judge and the coverage script.
  8. A quote that is a substring is not yet support. With frontier models 83 to 91% of required quotes were even substrings and only 48 to 79% of those supported the claim [Windisch 2026]; citation presence is not citation support [Gao 2023]; a three-way label (attributable, extrapolated, contradictory) is more useful than a binary one [Yue 2023]; abstention must be a first-class output [Menick 2022]. quote_not_in_chunk is a drop reason and the judge is a separate pass; since #749 a second pass marks each "misworded" verdict extrapolated or contradicted. The verdict pass itself is untouched, because the judge is sensitive to its own wording: the same 909 statements came out 0.8% misworded when the three verdicts were reworded as four and 2.1% when the kind was asked in the same call, against 5.0% and 3.6% from two passes of the unchanged judge.
  9. The contract's shape has a cost. Format restrictions degrade reasoning [Tam 2024]; guideline prompts are not reliably obeyed without training [Sainz 2024]; quoting before stating improves grounding [Huang 2024]; forcing citations finer than a sentence hurts [Wang 2026]. The compact contract was chosen by measurement (#731); the quote is the first slot of a statement, so the model copies the sentence before it writes from it (#748: on the four NVDA releases 909 statements against 901, misworded 5% in both, prose misworded 6% against 9% and prose that does not read alone 5% against 11%). The same four filings extracted twice under one configuration give 885 and 824 statements, misworded 3.7% and 4.2%, not stated 0.0% and 0.2%, reads alone 10.7% and 8.2%: a run is worth about half a point of misworded and seven per cent of its statements, which is the band any two numbers here have to clear. Open: clause-level quotes measured against minimal ones.
  10. Numbers need their own patterns. Open numerical extraction is its own problem [Saha 2017]; the dense-sentence figures we drop are that class. Measured, and the pass is not built. The four filings are the densest text we have, 25 figures per thousand characters against 7 for the next corpus, and a structural bench that reads no model counts every figure a passage writes against every statement drawn from it: over two runs, prose figures reaching no statement are 0 and 1 of 35, table figures 7 and 18 of 451. A second look would chase a dozen table cells for a second call on every chunk, and a chunk costs 150 seconds. Open: a corpus where prose figures are lost often enough for the pass to move the number.

Alignment

  1. Arguments before relations. Galárraga et al. cluster noun phrases first and relation phrases on a semi-canonicalized base; four coarse argument types raised pairwise precision from 0.946 to 0.997 because polysemous phrases split by signature [Galárraga 2014]; CESI inherits the order [Vashishth 2018a]; the mapping of an open relation to a DBpedia property depends on the argument types [Dutta 2015]; Angeli's KBP mapping is conditioned on the type signature [Angeli 2015]; typed markers lift relation extraction [Ling 2012, Zhong 2021]; ODKE+ exposes only the type's slice of the ontology [Khorshidi 2025]; iText2KG lists typing as its missing ingredient [Lairgi 2024]. Kind words bind before relation phrases [0044 cut 2]; the phrase binding reads the classes at both ends and the (class, class) signature is part of its key (#751).
  2. Precision is lost in the binding, not in the extraction. Of the mapping errors in the three-hour KBP system, 31% were definition mismatch and 23% over-generalized rules against 15% open-extraction errors [Soderland 2013]; the top hundred predicates cover only 57 to 82% of an open base's triples [Romero 2023]; an open statement often maps to a conjunction of ontology facts, not one property [Gashteovski 2020]. A word with no fitting class stays untyped and its statements stay open; the only output is a suggestion to add a class (ontology, #741). Open: a phrase may bind to a formula, not only a property.
  3. Nothing aligns to an undefined class. EDC's gain comes from definitions written before matching and a retriever whose recall bounds what can be bound [Zhang 2024]; definitions help both retrieval and the judge [Song 2025]; a class is introduced by its description alone [Obeidat 2019] and improved by example names [Ouyang 2024]; written guidelines are what make zero-shot extraction work [Sainz 2024]. The signature the aligner reads is the kind word, its spellings, example names and the relation phrases its things take part in; candidates come from the ontology's embedding index when the base has one, else the whole list when it is small (#741). Open: a class without a definition is not a binding target, and the ontology page says so.
  4. The characteristic error is the wrong level, not the wrong branch. Naive LLM matching binds a subclass to its parent [Norouzi 2023]; across ten models the errors are align-up, align-down and disputed, with few outright false mappings [Qiang 2025]; order sensitivity concentrates where the model is torn between its top options [Pezeshkpour 2024]; resolving top-down through the hierarchy helps [Komarlu 2024]. Open: the queue card for an undecided word shows each candidate with its parent and children; a candidate list includes the parent of every candidate.
  5. The judge prefers the first option. Swapping two answers changed GPT-4's verdict in a third of cases and Claude-v1's in three quarters [Zheng 2023]; reordering alone reversed a leaderboard [Wang 2024]; selection bias comes from the prior on option-ID tokens [Zheng 2024]. Two votes, the second with candidates reversed; candidates are named by class key, never by a letter; disagreement is a queue item, not a binding (#741). Spending the model's votes only on borderline cases is the next economy [Taboada 2025].
  6. Conformance is not correctness. Ontology conformance near 1.0 with hallucinated arguments [van Cauter 2024]; the standard metric names are conformance and hallucination rate [Mihindukulasooriya 2023]; the standard check for the no-match case is a rejection rate over concepts that have no counterpart [He 2023], now part of the OAEI Bio-ML evaluation [OAEI 2024]. Open: the alignment bench seeds kind words that fit no class and counts how often the aligner binds them anyway.
  7. Candidate recall is the ceiling. Whatever the retriever misses can never be bound [Hertling 2023, Zhang 2024]; the similarity threshold is a precision dial tuned per ontology [Qiang 2024]; the number of clusters is estimated, not set [Shen 2022]. The whole list goes to the model when the base has at most sixty classes. Open: candidate recall measured on the batch base before the embedding index is trusted above sixty.
  8. A role is not a kind. A class that names the role a thing plays in a relation cannot be assigned from the thing's own kind word; Soderland et al. hit it with BombingAgent [Soderland 2010]. Open: classes declared as roles are excluded from kind-word candidates.
  9. A pipeline that cannot be re-run drifts. Mappings are partial and revisable by design [Madhavan 2007]; error propagation from extraction into fusion is the survey's standing warning [Bian 2025]; the surveys ask for provenance and re-runnable stages [Weikum 2021, Hogan 2021]. Every binding carries its two votes and a decided_at compared with the class's updated_at; only stale bindings are re-decided; every statement keeps its quote and offsets; the typed rows are recomputed per changed signature and each carries the statement it was computed from (#752).

Time

  1. The document type chooses the anchor, and the anchor is where the value goes wrong. With the same extractor, anchoring narrative text to the document's date instead of the previous mention drops the value score from 89.8 to 64.9, and anchoring clinical-trial text to a local time point zero raises it from 74.4 to 88.7 [Strötgen 2012]; in clinical narratives the anchor choice is the least accurate component, at 74.7% [Sun 2015], and anchor errors are the top cause of wrong values [Olex 2022]; a discharge summary has two legitimate document times, admission and discharge [Sun 2013]. The document context is read once, mention anchors are resolved in a second pass, and the anchor is stored in the interpretation, so a value always says what it was computed from [0045]. Open: the strategy is chosen per document type; a trial-relative mention ("week 4") is kept as an offset from a named local anchor in the manner of HeidelTime's time point zero and THYME's event-anchored expressions [Styler 2014], so that it becomes grade B the day the anchor is dated, instead of grade C forever.
  2. Normalizing is harder than finding. In TempEval-3 the best system found mentions at 90.3 but gave the right value at 77.6, and every participant normalized by rules [UzZaman 2013]; the neural parsers score operators such as "last" and "next" at a third of the rest [Laparra 2018]; relative and underspecified mentions stay the hardest for prompted models [Gautam 2024]. The model gives only an interpretation; code computes the value; nothing is written when it cannot [0045]. The strongest recent version of that split generates a program and executes it [Su 2025a]; the vocabulary of the code side is the ten operations of ARTime [Ding 2021]. Open: the time bench scores a resolved interval by overlap with the reference, not by string equality [Laparra 2018]; an ambiguous mention may return alternatives rather than one value [Bethard 2013].
  3. A wrong document date poisons every relative expression. No study manipulates the document date alone; the anchor numbers of item 20 are the quantification. The document's own stated dates are the best evidence of its date [Chambers 2012]; year-level dating is useless for a quarter [Vashishth 2018b] and dating from implicit cues is decades coarse [Kumar 2012]. The agent-memory graphs of 2025 anchor relative dates to the message timestamp and collapse a date to midnight [Rasmussen 2025], the two errors this ledger forbids; ATOM separates observed time from valid time and measures run-to-run stability [Lairgi 2025]. Upload time and file modification time are never a document date [0045, #714]; anything computed from the document date is grade B. Open: a person changing the document date triggers re-resolution (0045 cut 4); run-to-run stability of the interpretation call is a bench number.
  4. A fiscal period is its own granularity. Fiscal quarters and 52/53-week years are non-regular granularities, and truncation to a calendar month is only defined when the target groups the source [Bettini 2000]; a 52/53-week quarter truncated to the month is off by up to six days. SEC filings carry the fiscal year end and a period code (FY, Q1 to Q4, H1, H2, M9, T1 to T3) in their XBRL metadata [SEC FSDS]. Fiscal quarters are computed from the fiscal year end the opening states [0045]. Open: half-years and nine-month periods are period codes the resolver accepts; the fiscal year end is read from the filing's metadata when the source is EDGAR; fiscal periods are one granularity system beside the calendar, not a calendar month.
  5. Mixing the two axes corrupts history. Snodgrass's book is one long warning [Snodgrass 1999]; transaction time is system-assigned and append-only, valid time is asserted [Jensen 1999]; record time is the commit instant, assigned once and never back-dated [Torp 2000]; SQL:2011 periods have neither a granularity nor an unknown end [Kulkarni 2012]. Two axes with separate predicates [0019]; the precision columns and the ended-unknown shape are stated as extensions beyond the standard. Open: time names the axes once in the glossary's words, valid and transaction [Jensen 1998].
  6. Unknown is three different things. Wikidata distinguishes a value, an unknown value that exists, and no value; an end time is "unknown value" when the thing ended on an unknown date and absent when it is ongoing [Wikidata Help]; the TimeBank anchoring corpus annotates a point, an interval, a lower bound, an upper bound or nothing [Reimers 2016]; YAGO2 leaves a fact without a time rather than give it an ill-defined one [Hoffart 2013]. Ours are until, ended_unknown with attested_to, and an open end with attested_to [0022, 0045], and grade C writes nothing. attested_to means true as of the last attestation, a now-relative value that is stored as an explicit timestamp and evaluated when asked [Clifford 1997]; a fact whose last attestation is old is stale, not false [Soulard 2025].
  7. A table is a structure, not a paragraph. The financial-report question-answering sets feed a table to the model one row at a time with its row and column headers written out [Chen 2021, Zhu 2021], and the pretrained table models flatten it without losing which row and column a cell sits in [Herzig 2020]; the standard for turning cells into one faithful sentence is ToTTo [Parikh 2020]; the statistical-report tables with hierarchical headers have their own dataset, whose difficulty is "complex hierarchical indexing" [Cheng 2022, Zhao 2022]; how a table is serialised, ordered and partitioned changes what a model reads from it [Sui 2024], a change of structure with the same content costs accuracy until the structure is normalised [Liu 2023], a whole table is not fed to a model at all when it is large [Chen 2024c], and the survey of the field is [Fang 2024]; which column is the subject and which are properties is the semantic table interpretation task [Cafarella 2008, SemTab]. Under #743 the earnings release lost its column headings between the HTML converter and the chunker and the model wrote the period as the phrase. Tables are now rendered from the DOM by structure, one row per line under its real headings, the section path folded into the label, the caption run kept with every chunk (sources, #744); a heading that names a time is when, a heading that names a change between periods stays the phrase; a phrase equal to the value or to the subject is kept and counted. DOCX tables, spreadsheets and CSV files are laid on the same grid by their parsers (#750): a DOCX cell keeps its span and the indent of its first paragraph, a sheet's or a file's first row of two or more cells is its header even when the headings are years, and a table with no figure in it is a header row and records. Open: the text layer of a PDF carries no table structure (a scanned PDF's tables arrive from MinerU as pipe tables already); the paragraph units of the NVDA cross-domain corpus are recut from rendered rows with their headings, and the reference facts carry the period; a check that a cell statement under a period column carries when.
  8. Precision is not confidence, and vague is not coarse. An indeterminate instant is a set of chronons with a distribution, and comparing two of them needs a stated semantics [Dyreson 1998]; confidence and bitemporality are separate axes [Chekol 2018]; "early 2019" is a modifier with a calibrated width, not a coarser precision [Tissot 2019]; grading a date by the provenance of its anchor is uncertainty typing, which the survey prefers to a scalar [Jarnac 2025]. A precision column on every bound, with CHECK constraints that truncate the value [0024]; the grade is a letter beside the value, never a probability. Open: the cross-precision comparison used by the temporal engine, overlap or containment, is written down in time; a vague modifier is a qualifier on the mention, not a precision.

What has no precedent

These decisions have no paper behind them and are held up only by our own benches (extraction, the batch review of #731, #740 and #741):

  • A described thing becomes an entity only when a statement points at it [#736].
  • The contract names things by their words, not by ids; ids mixed on dense passages [#731].
  • Temperature 0 and a verbatim quote as the condition of entry; "not stated" ≤ 2% on every corpus.
  • The three grades of a resolved time as a letter beside the value [0045]. YAGO2's refusal to write an ill-defined time and the survey's preference for typed uncertainty are precedents in spirit, not in form.
  • Every intermediate layer stays in the ledger and is re-run by a job, instead of a pipeline that computes once.

Three things the survey looked for and did not find: a controlled comparison of schema-first and schema-late extraction with the same extractor; a paper that studies generic-noun hubs as a named phenomenon; a study that manipulates the document date alone and measures what it does to relative mentions. All three are ours to measure.

References

Every entry was verified against the linked page. Grouped by the layer that cites it; a paper cited by two layers is listed once.

Open extraction

Granularity, qualifiers, propositions

Grounding and evaluation of model extraction

Canonicalization and alignment

Typing

Matching and judging

Integration and surveys

Tables

Time: tagging and normalization

Time: document dates, validity, clinical relative time

Time: the ledger