UDPipe Flutter

NLP tokenization, POS tagging and dependency parsing using UDPipe 1 UDPipe 1 github, built with Flutter.

Native FFI on desktop and mobile. WebAssembly via Emscripten on the web.

Platforms

Platform	Status
Android	✅
Web	✅
Windows	✅
Linux	✅
macOS	🚧 Built but untested
iOS	🚧 Built but untested

Features

Tokenization, lemmatization, POS tagging and dependency parsing
Runs fully on-device (no server required)
Separable verb detection for German
50+ pre-trained language models available from ÚFAL
Async batch processing keeps the UI thread free

Live Demo

https://konyshevgmbh.github.io/udpipe_flutter/

pub.dev

https://pub.dev/packages/udpipe_flutter

Getting Started

git clone --recursive https://github.com/konyshevgmbh/udpipe_flutter.git
flutter pub get
flutter run -d windows         # Windows desktop
flutter run -d linux           # Linux desktop
flutter run -d macos           # macOS desktop
flutter run                    # Android (device or emulator)

Web (WASM)

Build the WebAssembly module first, then run on Chrome:

make run-web                   # builds WASM via Docker, then launches Chrome

Requires Docker Desktop.

Model files

Download models from ufal.mff.cuni.cz/udpipe/1/models and place the .udpipe file in assets/models/ (or any path inside your Flutter assets).

All ~94 models from UDPipe 1 (UD 2.5) are listed in kUdpipeModels and in assets/models/README.md.

A few popular ones:

Model ID	Language	File
`german-gsd`	German	`german-gsd.udpipe` (~20 MB)
`german-hdt`	German	`german-hdt.udpipe` (~60 MB)
`english-ewt`	English	`english-ewt.udpipe`
`russian-syntagrus`	Russian	`russian-syntagrus.udpipe`
`french-gsd`	French	`french-gsd.udpipe`

Library API

import 'package:udpipe_flutter/udpipe_flutter.dart';

final svc = UDPipeService();

// 1. Load from kUdpipeModels catalog (file must be in assets/models/)
await svc.init(modelId: 'german-gsd');

// 2. Load from any asset path
await svc.initFromAsset('assets/models/russian-syntagrus.udpipe');

// 3. Load from bytes (e.g. downloaded at runtime)
final bytes = await downloadModel();
await svc.initFromBytes(bytes);

// Process text (synchronous, call from any isolate)
final result = svc.process('Er steigt aus dem Bus aus.');

for (final sentence in result.sentences) {
  for (final token in sentence.tokens) {
    print('${token.form} → ${token.lemma}  [${token.upos}]');
    // Er → er  [PRON]
    // steigt → steigen  [VERB]
    // Bus → Bus  [NOUN]
  }
  for (final sv in sentence.sepVerbs) {
    print('sep.verb: ${sv.particle}+${sv.verbForm} → ${sv.fullLemma}');
    // sep.verb: aus+steigt → aussteigen
  }
}

// Async batch — keeps UI thread free (background isolate on native, micro-tasks on web)
final results = await svc.processAllBlocksAsync(['First paragraph.', 'Second paragraph.']);

Architecture

lib/udpipe/
  udpipe_flutter.dart          — public API
  src/
    conllu_parser.dart         — CoNLL-U parser (pure Dart)
    udpipe_types.dart          — data classes and result builders
    udpipe_bindings.dart       — dart:ffi bindings
    udpipe_service.dart        — conditional export (native vs web)
    udpipe_service_native.dart — FFI service (Windows / Linux / macOS / Android)
    udpipe_service_web.dart    — WASM/JS service (web)

native/
  CMakeLists.txt               — builds udpipe_flutter.dll / .so / .dylib
  udpipe_ffi.cpp               — thin C wrapper around UDPipe
  udpipe_src/                  — git submodule: ufal/udpipe
  build_web.sh                 — Emscripten build script

Tech Stack

Component	Library
UI	Flutter 3.x + Material 3
Native bindings	ffi
NLP engine	UDPipe 1 (C++, MPL 2.0)
Web engine	Emscripten WASM