HermitCrab analysis: an affix-rule index cut port work 11.8% on Cebuano; C# gain unmeasured

Aberta
  • Algo errado?
#599 1 comentário 0 reações 0 responsáveis Ver no GitHub

Mantenedores costumam responder em até 1 dia

Ninguém assumiu esta issue ainda.

Avaliação

Dificuldade
4/5
Tempo estimado
3-5 dias
Facilidade para iniciantes
45/100
Tipo de issue
Funcionalidade
Clareza
Razoavelmente clara
Status de atividade
Ativa
Stack de tecnologia
csharp
Domínio
backend

Direção de pesquisa

Start by reading PR #574 for the existing per-allomorph prefilter and issue #490 for the index analysis; the payload names no source files or test paths. Then trace the C# analysis cascade and template-slot batches, including where the syntactic-feature gate runs. Done means the rule-level index preserves identity on the conformance fixtures and five sample grammars, with the reported work reductions measured.

Escrita pelo modelo de indexação a partir do texto da issue.

Descrição

Hermit Crab

Claude-drafted.

An affix-rule index, built once at grammar load, cut analysis work by 11.8% on a 7,206-type Cebuano held-out set in the Rust port, and by 0 to 4.9% on five sample grammars. The C# gain is unmeasured. C# still visits every affix rule on every word and rejects allomorph by allomorph (#574).

Status: built and merged in the port (sillsdev/PanGloss PR #95); not built in C#.

What we know (port, work steps; parses identical on every completed word):

  • Cebuano 7,206 held-out types: 5,664,386 to 4,996,598 (-11.8%). Aweti -4.9%, Sena -3.2%, Indonesian -0.4%, Amharic 0 (measured before the final syntactic-gate change, which cannot raise them).
  • 141 conformance fixture grammars: 1,501,384 to 1,489,476 (-0.8%). No grammar or word went up.
  • Memory: 0 to 4 KB per grammar.
  • The index maps each rule to the mandatory first and last edge segments of its allomorphs. A rule is skipped before its visit only when #574's edge predicate rejects every allomorph, so results cannot change. Optional, unknown, modified, deleted and featural edges take the old path.
  • It is consulted at every cascade recursion and template slot batch, against the current intermediate shape, and only after the syntactic-feature gate.

Not known: the C# gain. #574 leaves C# census steps unchanged (ledger row 42), so a C# gain would show as matcher work, not steps.

Proposed change: port the index to C# after #574 merges, behind the same identity gate (conformance suite plus the five sample grammars).

Related:

  • PanGloss: PR sillsdev/PanGloss#95 (merged); umbrella issue sillsdev/PanGloss#90 (open).
  • sillsdev/machine#574 (PR, open) and #564 (issue, open): the per-allomorph prefilter this extends.
  • sillsdev/machine#610 (open): the PanGloss#90 clitic measurement that led here.

Next: after #574 merges, measure matcher work on Sena with and without the index on hc.dll.

Evidence -- numbers, method, commits

Work steps from PanGloss PR #95 (squash-merged 2026-10-30) and its lane report. Counts via --stats --work-columns --threads 1 --step-cap 200000 for Cebuano.

Grammar Work before Work after Change
Cebuano held-out, 7,206 types, round 2 5,664,386 4,996,598 -11.8%
Cebuano with enclitic g, held-out 6,623,154 5,840,844 -11.8% (g's added cost 958,768 to 844,246)
Sena (round 1) 182,297,249 176,401,068 -3.2%
Indonesian (round 1) 1,617,711 1,611,898 -0.4%
Amharic (round 1) 32,238,244 32,238,244 0
Aweti, 194 valid words and 12 invalid shapes, 30M-attempt cap (round 1) 1,237,499,929 1,176,371,185 -4.9%
Mbugwe, 200-word check (round 1) 11,449,056 11,047,179 -3.5%
141 fixture grammars, round 2 1,501,384 1,489,476 -0.8%

Round 1 charged one step per visit that rejected a rule; round 2 charged only rules that pass the syntactic gate, which cannot raise the round 1 totals. Sample grammars were not re-run in round 2. Eleven Mbugwe words never complete on either build and are unverified.

Why the index must see the current shape: on Cebuano, earlier rules expose intermediate forms that end in a clitic's segment, so a check against the surface word alone misses them. Why it comes after the syntactic gate: a rule that gate rejects for free would otherwise pay for the lookup.

Paths not taken -- the environment pre-check

An input-word pre-check of the allomorph environment (E5 in the #490 round) is a dead end (see #601). On Cebuano it could save at most about 4% of an added enclitic's cost: 83% of the g clitic's added steps fall on words ending vowel + g, where its environment [V] _ is satisfied and a host search runs. The index is the general lever. Source: PanGloss#90 comments.

Sources -- pinned
Linguagem predominante
C#
Estrelas
29
Forks
16
Merge médio
4d 11h
PRs com merge (30d)
14

Preparar o ambiente

  • Sem Dockerfile nem arquivo Docker Compose
  • Tem um modelo de pull request
  • Sem guia de contribuição

Primeiros passos

  1. Leia a issue inteira e depois o guia de contribuição do projeto.
  2. Comente na issue dizendo que vai assumir — evita que duas pessoas façam o mesmo trabalho.
  3. Faça um fork do repositório e trabalhe em uma branch.
  4. Abra um pull request que referencie o número da issue.

Mais de sillsdev/machine

Todas as issues de sillsdev/machine

Issues semelhantes

Mais issues de C#

Receba novas issues na sua caixa de entrada

Um resumo curto de issues do GitHub para quem está começando.