ablaut

A fast, correct German verb conjugator.

The name is the German term for the vowel gradation in singen, sang, gesungen: Jacob Grimm formalized its seven classes in 1819, and they still describe every strong verb in the language.

Highlights

• Measured correctness. Every generated form is validated against two machine-readable gold standards: 99.2% agreement with UniMorph deu (194k forms) and 98.0% with the Wiktextract extraction (794k forms, including analytic tenses and the perfect auxiliary). Where the two sources agree with each other, ablaut has zero known errors; every remaining disagreement is ruled on in a published adjudication log with references to standard grammars.
• Generalizes to unseen verbs. A rule engine with a curated exception lexicon, not a lookup table: novel verbs (googeln, downloaden) conjugate correctly. On verbs outside its lexicon, agreement with UniMorph is 99.55%.
• Fast and small. No I/O, no runtime data files, no dependencies in the core crate. A complete conjugation table (100+ forms) takes microseconds; the entire engine compiles to a 106 KB WebAssembly binary (about 45 KB gzipped).
• Permissively licensed. MIT OR Apache-2.0, including the lexicon: no ShareAlike obligations.

Usage

Rust

cargo add ablaut

use ablaut::{AnalyticTense, Mood, Number, Person, Tense, Verb};

let v = Verb::from_infinitive("aufstehen")?;

v.conjugate(Tense::Present, Mood::Indicative, Person::First, Number::Singular);
// "stehe auf"
v.conjugate(Tense::Preterite, Mood::Indicative, Person::Third, Number::Singular);
// "stand auf"
v.analytic(AnalyticTense::Perfect, Mood::Indicative, Person::First, Number::Singular);
// "bin aufgestanden"
v.past_participle();          // "aufgestanden"
v.zu_infinitive();            // "aufzustehen"
v.imperative(Number::Singular); // Some("steh auf"); None for modals

The full paradigm as one struct: ablaut::table::Table::build(&v).

Python

pip install ablaut

import ablaut

c = ablaut.conjugate("aufstehen")
c.present[0]        # "stehe auf"
c.perfect[0]        # "bin aufgestanden"
c.auxiliary         # "sein"
c.zu_infinitive     # "aufzustehen"
c.imperative        # [None, None] for modals

Wheels are abi3 (Python 3.11+); no Rust toolchain needed.

WebAssembly

wasm-pack build --target web -- --features wasm

import init, { conjugate } from "./pkg/ablaut.js";
await init();
const c = conjugate("aufstehen");
c.present[0];       // "stehe auf"
c.zuInfinitive;     // "aufzustehen"

Coverage

All synthetic and analytic forms of standard German:

• Tenses and moods: Präsens, Präteritum, Perfekt, Plusquamperfekt, Futur I/II, Konjunktiv I and II in all tenses (the würde-form is Futur I Konjunktiv II), processual and statal passive
• Non-finite and imperative: both participles, zu-infinitive, perfect infinitive, imperatives including the adhortative (stehen wir auf!) and polite (stehen Sie auf!) forms
• All five inflection classes: weak (the productive default), strong (ablaut, with 2/3sg stem alternation: sprichst, hältst), mixed (dachte), preterite-present (the modals and wissen: ich kann, no imperative), and suppletive (sein, werden, tun)
• The prefix system: separable (stehe auf, aufgestanden, aufzustehen), inseparable (verstanden), fused compounds that never split (lobsingen: lobsang, lobgesungen), multiword lemmas (Rad fahren: fahre Rad, Rad gefahren), and per-lexeme rulings for dual-behavior prefixes (umarmen vs untertauchen)
• Orthographic surface rules: e-epenthesis (arbeitest, bewillkommnest), s-coalescence (du lässt), schwa elision (ich sammle), the Latinate -ieren participle (studiert but geschmiert), and per-lexeme perfect auxiliaries (aufstehen: sein)

The exception lexicon (data/verbs.tsv, about 900 human-readable rulings) is compiled into the binary.

How correctness is verified

src/bin/golden.rs diffs every generated form against the gold data (scripts/fetch_unimorph.sh, scripts/fetch_kaikki.sh) and CI fails if accuracy regresses below pinned thresholds. scripts/cross_oracle.py cross-examines the two oracles against each other: they agree on 98.9% of their 195k shared slots, and triaging ablaut's mismatches against that agreement separates genuine bugs from gold-data errors.

The method has found real errors in the gold data itself, documented in docs/adjudications.tsv: UniMorph gives the Konjunktiv II of nennen as nannte (correct: nennte) and marks wisse! as nonexistent; Wiktionary assigns wachen the auxiliary sein and conjugates entgelten weak. One caveat: both oracles derive from Wiktionary, so their agreement is strong but not fully independent evidence.

Design

The architecture is a productive default rule plus a finite table of stored exceptions, which is precisely the dual-mechanism model of inflection (Marcus, Pinker et al., 1995) for which German morphology is the canonical test case. Features follow the UniMorph schema. docs/design.md maps the whole domain in four layers and doubles as the design spec.

License

MIT OR Apache-2.0.

Data provenance: the exception lexicon is independently curated factual data about standard German (principal parts, prefix behavior, auxiliaries), hand-verified row by row. UniMorph (CC BY-SA 3.0) and Wiktextract (CC BY-SA) are used as test oracles only: fetched at development time, never shipped.