Skip to content

Should ten and oude be particles? (Jan ten Brink gives middle ten, family Brink) #1455

Closed

Description

@derek73

Neither word is listed, so each reads as a middle name (measured on master c878443):

Jan ten Brink         → middle 'ten', family 'Brink'
Piet Oude Luttikhuis  → middle 'Oude', family 'Luttikhuis'

Measured 2026-11-18, each added as kind="preposition" in a lexicon of its own, scored against the filed boxes (config_diff.py score default 'expr:Parser(lexicon=Lexicon.default().add("ten", kind="preposition"))' --cap 20000, ORCID, PatentsView and PSC, every group):

Word fixed broken where
ten 86 15 fixed: ORCID NL 31, PatentsView NL 28, PSC Dutch 7. Broken: PatentsView NL 4, Singapore 3, Taiwan 2, one each in six more groups
oude 16 1 ORCID NL 10, PatentsView NL 5; broken PatentsView NL 1

History. ten was tried as a global particle in PR #191 and withdrawn with to, in, an and then for swallowing a middle syllable or initial (AGENTS.md, Extension Patterns). That was decided on test names; the count above is the first on filed boxes.

Questions.

  1. Is 86 to 15 enough for the default, or does ten go in nl.toml as default = false, on under a Dutch choice only? The broken names outside the Netherlands are East and Southeast Asian (Ten as a given-name syllable).
  2. Does ten take a given reading, so that it reports particle-or-given where it opens a name?
  3. oude stands before a second word that is itself the surname's start (Oude Luttikhuis, Oude Elferink). Is it a particle, or should those surnames be read some other way?

Not done: the words' other readings and bearer counts (tools/name_usage.py), and the committed-corpus and local-corpus moves, as nameparser/vocab/AGENTS.md asks before a word is added.

Related: #1239 (vanden), #1284 (aus), #1059 (dels), the other particles waiting on a measurement.

Activity

  1. added this to the 3.2 milestone on Nov 18, 2026
  2. added
    vocabularyFixed by a word-list entry: per-word vocabulary and language packs
    on Nov 18, 2026
  3. self-assigned this
    on Nov 18, 2026
  4. derek73 commented on Nov 18, 2026

    @derek73
    OwnerAuthor

    Word vetting for ten and oude (nameparser/vocab/AGENTS.md steps 1–2, plus the mid-name test)

    Dated snapshot: 2026-11-18, master c8784432, py3.11, every command run from a clean worktree. No file changed; nothing below is a decision. Sections 1–6 report what the tools printed; the last section, Reading (mine), is my interpretation and is labelled as such.

    1. Other readings and earlier decisions

    • uv run python tools/vocab/query.py --word ten and --word oude: no reading and no excluded entry in any file (exit 1, nothing printed). Neither word appears in nameparser/vocab/*.toml, and none of docs/design/decisions/ (vocabulary-exclusions, vocabulary-collisions, P1, P2) or docs/design/rules/ decides either word.
    • What PR Add German/Dutch prefixes and German title/degree suffixes #191 recorded. Its body lists to, in, an, then, ten under "Deliberately not included … as global prefixes — these are common Korean/Vietnamese given-name syllables in the middle-token position (e.g. Park In Hwan), and adding them regresses a currently-correct parse for those names". That rests on test names. The PR gives no corpus count and no example for ten itself. The same PR did add te (since given a given reading, nl.toml) and op.
    • The rule that now binds such a word: decisions/vocabulary-collisions, 2026-11-12 Is ben/bin/בן a particle, a patronymic marker, or a given name? (Karim Ben Salah, בן כספית, Youssef Ait Taleb) #1257, item 2: "A Latin spelling is on by default only where it passes both the mid-name test … and bearer counts showing it rarely a given name. Otherwise it is written default = false, for a language choice to turn on." Item 2 names Add German/Dutch prefixes and German title/degree suffixes #191's test as the mid-name test. A decided exception to it exists (el, the same entry).

    2. Bearers

    uv run python tools/name_usage.py ten oude --country NL --country BE --country ID --country SG --country TW --country KR --country VN --country CN --country MY (the --country flags imply --boxes). Lowercase and capitalized inputs gave identical rows.

    source ten oude
    US Census 2010 surnames 452 bearers (rank 47,006; 44.69% Asian/PI, 24.12% Hispanic, 23.23% white) under 100
    US SSA given names under 5/yr every year under 5/yr
    nam_dict given names not listed band 1 (rare), the Netherlands only
    Wikidata (lower bound) given 1+ (Japan 1); family 16+ (Spain 4, France 2, Russia 2, USSR 2, Canada 1) 0 / 0
    ORCID boxes (anywhere in box) given 40, family 610. Given by country: BR 3, BT 2, SG 2, IT 1, JP 1, NL 1. Family: NL 150 (0.18% of 85,119), US 28, RU 12, DE 11, BR 8, BE 2 given 1, family 78. Family: NL 29, BE/FR/GB/IE 1 each
    PatentsView boxes given 81, family 236. Given: US 20, NL 16, DE 13, TW 12, SG 3, CA 3. Family: NL 125 (0.19% of 66,001), US 46, DE 19, CA 8, BE 5 given 7 (NL 5), family 24 (NL 16)
    local corpora, word anywhere ctgov-officials 42 of 539,368 (ten 33, Ten 9), 0 after a comma; tse-donors 12 (all TEN); none in cz-candidates tse-donors 1

    These box counts include the word anywhere in the box. AGENTS.md asks for boxes the word opens, so I scanned the extracts read-only: a scratch script over ~/.cache/nameparser-tools/local/{orcid-researchers,patentsview-inventors,psc-individuals}/raw.jsonl, matching the word as a whole token. These extracts are smaller than the box index name_usage.py reads, so their counts are lower.

    • ten opening a family box: ORCID NL ten 28 / Ten 3, BE 1, elsewhere 6. PatentsView NL Ten 20 / ten 8, BE 1, elsewhere 11. PSC Dutch 7, elsewhere 2.
    • ten as the whole family box (a one-word surname): ORCID 9, PatentsView 8, PSC 2. These are Korean-Russian (Koryo-saram) bearers (Russian given names with Ten), Malaysian/Singaporean Chinese bearers, and Turkish bearers. None is Dutch.
    • ten in a given box: in three ways.
      • Opening a given name (SG, TW, BT; e.g. Ten Kao / Liu): ORCID 3, PatentsView 5.
      • Inside one, as a Chinese given-name syllable (MY, SG).
      • A Dutch ten typed into the given box by the filer: PatentsView NL 4, NO 1, and 1 with no country.
    • oude: every family box that holds it writes Oude, capitalized, as the first of two words (ORCID NL 10 of 10; PatentsView NL 5 plus 1 elsewhere). It never stands alone as a family box and never opens a given box. It appears inside a family box only after another particle (De Oude) or another surname (a double Dutch name, X Oude Y). One filer typed it into the given box.

    3. What a particle reading moves (corpus diff)

    uv run python tools/differential/config_diff.py diff default 'expr:Parser(lexicon=Lexicon.default().add("ten", kind="preposition"))' 'expr:…add("ten", kind="preposition").add("ten", kind="given"))' 'expr:…add("oude", kind="preposition"))' 'expr:…add("oude", kind="preposition").add("oude", kind="given"))' --committed --local all --match '(?i)(^|[\s,.])(ten|oude)([\s,.]|$)'

    The --match filter reads only names holding a word, and no other name can move, so this is the full move set. It read 273 names. No committed corpus (tools/differential/corpus*.jsonl) holds either word, so snapshot.jsonl would not move.

    ten, particle only: 180 names move. By source: ctgov 35/40, fec 11/43, govuk 2/2, jrc-names 8/22 (not quoted), nepal 1/1, orcid 47/67, patentsview 54/73, psc 11/14, tse 12/13. By field: middle+family 168, given+middle+family 7, given+family 5. I read the 172 that are not jrc-names:

    • About 150 (153 by my hand count): a Dutch or German surname gains its particle (… ten X, … Ten X, both cases), mostly m:ten f:X → f:ten X.
      • 10 of these are FEC family-comma names (shape BRINK, JAN TEN → family TEN BRINK). This is rules/P6's trailing-particle attachment after a family comma.
      • 12 are one Brazilian family name, TEN CATEN (tse-donors).
      • Hyphenated and double names come out whole: ten Cate-Hoek, ten Bokkel Huinink.
    • 10: Ten opening the name becomes part of the family.
      • Nine are a given name: TW/SG Chinese (5; Ten Kao Liu → family Ten Kao Liu), Bhutanese (2), Nepal (1), and a PSC filer whose given box is Ten.
      • The tenth is an FEC family-comma name whose given name is the surname Ten Eyck (shape SMITH, TEN EYCK → family SMITH TEN EYCK).
      • All ten are wrong under the never-given particle. A given reading restores every one (the reading is given='Ten' again, and nine of the ten report particle-or-given).
    • East and Southeast Asian syllables inside the name: 3. Kheng Ten Choo (MY) → family Ten Choo, plus one SG and one Chinese name of the same shape. The filed boxes say these are wrong. A given reading does not help here, because it acts only at the opening position.
    • Spanish Ten as a first surname: 2–3. For example Amadeo Ten Esteve → family Ten Esteve (Valencian surname Ten; Wikidata's Spain 4). This is arguably more right than the default's family Esteve. PatentsView scored one ES record as broken (a filer with Ten in the given box).
    • A name with two particle runs: 1. Jutte van der Werff ten Bosch reads family van der Werff ten Bosch today. With ten listed it reads middle van der Werff, family ten Bosch. This is the existing shape for any listed second particle: … van der Werff ter Bosch and Jan van den Berg te Riele split the same way on master.
    • A filer's box disagrees with the name, whatever the parser does: 2. One NO name with ten Velden filed in the given box. One PSC name filed with a four-word given box.
    • No initial was swallowed. Every name with an initial before ten (shape J.M. ten Brink, R. W. ten Brink) keeps the initial as middle and gains ten X as family.

    ten, particle + given: 179 names move. The same moves, except that the openers stay given. 11 names gain only an ambiguity, and 10 FEC comma names gain particle-or-given beside the move.

    oude, particle (with or without given): 22 names move, all middle+family. 21 readable: every one a Dutch Oude X surname gaining Oude (one Brazilian-Dutch in tse-donors; plus 1 in jrc-names, not quoted). No other kind of move appears.

    4. Scores against the filed boxes (three alternatives)

    uv run python tools/differential/config_diff.py score default <the four expr: specs above> --cap 20000: ORCID, PatentsView and PSC, every group of 100+ records, 20,000 records per (source, group). The cap means each NL row reads about 20k of ORCID's 85k NL researchers, so absolute counts are a sample. Wall time was 6 minutes.

    alternative fixed broken where
    ten particle, never-given (the issue's run) 86 15 fixed: ORCID NL 31, PV NL 28, PSC Dutch 7, 20 elsewhere (US 3, GB 2, NO 2, BE 2, …). Broken: PV NL 4, PV SG 3, PV TW 2, and one each in ORCID BT, NO, SG, PV ES, MY and no-group
    ten particle + given 86 9 identical fixes. The 6 opener breaks (PV TW 2, PV SG 2, ORCID BT 1, ORCID SG 1) are gone. The 9 left: PV NL 4 + no-group 1 + ORCID NO 1 (Dutch ten typed in the given box), PV MY 1, PV SG 1 (mid-name Chinese syllable), PV ES 1
    oude particle, with or without given 16 1 ORCID NL 10, PV NL 5, PV DE 1. Broken: PV NL 1, a filer who typed Oude into the given box

    On only under a Dutch choice (what default = false in nl.toml would do).

    • How I simulated it. The comparator was plain:nl, i.e. parser_for(Locale(code="choice", lexicon=Lexicon.empty(), languages={"nl"})). The candidates were that parser's lexicon with the reading added as a caller edit: expr:(lambda p: Parser(lexicon=p.lexicon.add("ten", kind="preposition")[.add("ten", kind="given")], policy=p.policy))(parser_for(bare("nl"))), and the same for oude.
    • Why that stands in for a real nl.toml reading. A caller edit is not a real default = false reading. It behaves the same only because neither word has a reading in any other file, so choose has nothing to turn off. I did not measure a strict choice (strict:nl).
    • What it shows.
      • plain:nl agrees with default on every scored record in every group (0 groups differ).
      • The added reading gives the same 86/15, 86/9 and 16/1 as above, but only for a caller who chose Dutch.
      • The default parse is unchanged, so default callers get 0 fixed / 0 broken.
      • A caller who chooses Dutch only for NL/Dutch records sees ten: 66 fixed / 4 broken (ORCID NL 31/0, PV NL 28/4, PSC Dutch 7/0) and oude: 15 / 1.
      • The other 20 fixed for ten (and 1 for oude) are Dutch-surnamed people filed under other countries. They stay unfixed unless the caller chooses Dutch for them.

    5. Opposite-case sweep (the mid-name test)

    A scratch script parsed 23 shapes × 3 orders (GIVEN_FIRST, FAMILY_FIRST, FAMILY_FIRST_GIVEN_LAST) × {default, particle, particle+given}. The shapes covered the word in the middle slot (Park Ten Hwan, Nguyen Ten Nga, Tan Ten Hock, John Ten Smith, J. Ten Smith) and at the head (Ten Brink, Ten Wei Ming, Ten K. Smith, Ten John). The comma shapes were: none; a full name before the comma (Brink, Jan Ten, Brink, Ten Jan); one word before (Ten, Jan); nothing before (, Ten Brink); and a suffix comma (Ten Smith, Jr.). Synthetic names; oude behaves identically. What reads wrongly (by my reading of the intended name):

    • Middle slot, given-first, under either particle alternative: Park Ten Hwan → family Ten Hwan. The same happens for every three-word name with Ten second. This is Add German/Dutch prefixes and German title/degree suffixes #191's failure, and the given reading does not prevent it.
      • Under the two family-first orders the middle Ten instead joins the following word into the given name (Tan Ten Hock → given Ten Hock). For a two-syllable Chinese given name that may be the intended reading.
    • Head, never-given only: Ten Brink → family Ten Brink (right for Dutch). Ten Wei Ming, Ten K. Smith and Ten John → all family (wrong). The given reading restores given='Ten' with particle-or-given under every order.
    • After a family comma, never-given only:
      • Brink, Ten Jan → family Brink Ten Jan.
      • , Ten Brink → family Ten Brink, given empty.
      • The given reading restores both.
    • Before a suffix comma, both alternatives: Ten Smith, Jr. → family Ten Smith with no ambiguity reported, even with the given reading. This is a gap in where the fork is reported, not specific to ten.
    • Initials are safe in every order: J. Ten Smith keeps J. as given, and Jan ten Brink with an initial before ten keeps the initial.
    • Case repair (Parser.capitalized, plain and force=True) under the particle reading:
      • JAN TEN BRINK → Jan ten Brink, matching ORCID NL's family boxes (ten 28 : Ten 3).
      • PIET OUDE LUTTIKHUIS → Piet oude Luttikhuis. Every family box holding oude writes it Oude, so oude as a particle would need a cased mask (AGENTS.md, "A word that becomes a particle: its case").

    6. oude specifically

    • Shape: always Oude + one surname word (Oude Luttikhuis, Oude Elferink, Oude Aarninkhof, …). These are Twente/Achterhoek farm names.
    • Never alone: no box holds Oude as a one-word surname, and no corpus name ends in it.
    • Never before a one-word surname, other than as the head of such a compound.
    • Never a given name in any source: nam_dict band 1 in NL, no Wikidata bearers. The single PSC/PV given-box sighting is a filing slip.
    • Which reading fits: the particle reading produces exactly the filed family in all 21 readable moves. No other kind of move appeared. The kind is an open question. preposition is what the issue measured, but oude is an adjective ("old"), and nl.toml's other kinds are preposition and article. Whatever kind is chosen, the reading needs a cased mask.

    Not measured

    • A strict Dutch choice.
    • two_family_names or other policies.
    • The ctgov/ukec/govuk comma-credential interplay beyond the diff above.
    • ten's Spanish surname reading as its own question.
    • Groups under 100 records.
    • Records past the 20,000 cap.
    • Bearer counts' ends-with/opens figures from the full ORCID box index: I used the smaller raw.jsonl extracts.

    One side effect to note: name_usage.py rebuilt ~/.cache/nameparser-tools/local/ctgov-officials/wordindex.json on its first run ("built from other records or by another fold"). That is the tool's own derived index in the shared cache. I wrote nothing else there.

    Reading (mine, not a measurement)

    oude

    ten

    • What the data supports:
      • The Dutch gain is real and large: about 150 corpus moves, and 66 fixed / 4 broken on NL/Dutch boxes.
      • A given reading removes most of the non-Dutch cost: 15 → 9 broken, and every opener restored.
      • Of the 9 breaks that remain, 6 are Dutch filers who put ten in the given box.
    • What it fails: the mid-name test as Is ben/bin/בן a particle, a patronymic marker, or a given name? (Karim Ben Salah, בן כספית, Youssef Ait Taleb) #1257 item 2 states it. In real data that failure shows up as 3 Malaysian/Singaporean/Chinese names (Kheng Ten Choo) and every synthetic X Ten Y. The given reading cannot reach that position.
    • What the bearer counts show:
      • As a surname, ten is mostly Asian in the US Census (452 bearers, 44.7% Asian/PI), and Koryo-saram, Chinese and Turkish where it stands alone.
      • As a given-name syllable it appears in TW, SG, MY and BT.
    • How the two placements compare: by the rule as written, default = false in nl.toml (with the given reading wherever the reading lands) is the placement the rule points to. Default-on would be an exception in the way el was: it would add 20 fixes outside NL at the cost of the few breaks above.

    That trade is not mine to weigh.

  5. derek73 commented on Nov 18, 2026

    @derek73
    OwnerAuthor

    Decided 2026-11-18, on the vetting above:

    • oude: a particle in the default. It needs a cased mask so case repair writes Oude, and its kind is the builder's to propose (it is an adjective; nl.toml's particle kinds are preposition and article).
    • ten: a particle in nl.toml with default = false, on under a Dutch choice, with a given reading beside it. This follows decisions/vocabulary-collisions' 2026-11-12 Is ben/bin/בן a particle, a patronymic marker, or a given name? (Karim Ben Salah, בן כספית, Youssef Ait Taleb) #1257 item 2: the word fails the mid-name test (Kheng Ten Choo), so it is not an exception as el was. The 20 fixes for Dutch-surnamed people filed under other countries are given up for the default caller; a caller with Dutch names is expected to choose Dutch.

    Left for the build to settle: no shipped locales.NL constant exists, so a caller chooses the language by its tag today. Whether ten is reason enough for one is a question for the PR.

  6. derek73 commented on Nov 18, 2026

    @derek73
    OwnerAuthor

    Decided 2026-11-18: the change adds a shipped locales.NL pack, so a caller reaches ten the way the other packs are reached. Without it the reading would be on only for a caller who builds the language choice by its tag.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

Labels

particlesParticles and surname prefixes (rules/P)vocabularyFixed by a word-list entry: per-word vocabulary and language packs

Projects

No projects

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions