Doccano

Provides a unified annotation interface supporting three distinct NLP task types (text classification, sequence labeling/NER, sequence-to-sequence) within a single project management system. Uses a Django REST Framework backend with task-specific serializers and Vue.js frontend components that dynamically render annotation UIs based on project type configuration. Label schemas are defined per-project and enforced at the API layer, enabling teams to switch between annotation paradigms without data migration.

Implements a three-tier permission model (project admin, annotator, viewer) with Celery-based asynchronous task assignment and progress aggregation. Uses Django's authentication system to enforce access control at the API endpoint level, while the frontend tracks per-user annotation state and completion metrics. Example assignment logic distributes documents to annotators with optional overlap for inter-annotator agreement measurement, storing assignment state in the database for resumable workflows.

Supports both single-label (mutually exclusive) and multi-label (independent) text classification annotation. The frontend renders classification labels as buttons (single-label) or checkboxes (multi-label), with the backend storing annotations as label references. The annotation UI prevents invalid state transitions (e.g., selecting multiple labels in single-label mode) through client-side validation.

Supports sequence-to-sequence (seq2seq) annotation where annotators provide target text outputs for source documents (e.g., summaries, paraphrases, translations). The frontend provides a text input field for annotators to enter the target sequence, with the backend storing source-target pairs. Export formats include JSONL with source and target fields, compatible with seq2seq model training frameworks.

Supports annotation in multiple languages including right-to-left (RTL) languages (Arabic, Hebrew, Persian) with proper Unicode text handling and bidirectional text rendering. The frontend uses CSS flexbox with direction properties to render RTL text correctly, while the backend stores all text as UTF-8 without language-specific processing. Language selection is per-project, affecting UI language and text rendering direction.

Provides a pluggable auto-labeling system that integrates with external ML services (OpenAI, Hugging Face, custom REST endpoints) via a template-based request/response mapping system. The backend stores auto-labeling configurations per-project, including service credentials, request templates (with variable interpolation), and response parsers. Celery tasks execute auto-labeling asynchronously on imported datasets, with results stored as pre-filled annotations that annotators can accept, reject, or modify.

Supports importing datasets from multiple formats (CSV, JSON, JSONL, plain text files) with automatic format detection and schema mapping. The import pipeline uses Celery tasks to process large files asynchronously, parsing each row/object and creating Example records in the database. Users can map CSV columns or JSON fields to document text and optional metadata fields, with validation errors reported in a summary log rather than blocking the entire import.

Exports annotated datasets in multiple formats (JSONL, CSV, CoNLL for sequence labeling, JSON for seq2seq) with configurable field selection and filtering. The export pipeline uses Celery to serialize annotations asynchronously, transforming the internal annotation representation into task-specific formats. Users can filter exports by annotator, completion status, or label type, with the resulting file generated as a downloadable artifact or streamed to cloud storage.

Exposes a Django REST Framework-based API with endpoints for CRUD operations on projects, examples, annotations, and labels. The API uses token-based authentication (Django Token Auth) and implements pagination, filtering, and ordering on list endpoints. Clients can programmatically create projects, import examples, submit annotations, and export datasets, enabling integration with external ML pipelines and custom annotation workflows.

Allows users to define custom label taxonomies per-project with support for single-label (classification) and multi-label (sequence labeling) scenarios. Labels are stored in the database with project foreign keys, enforcing that annotations can only use labels defined for their project. The backend validates annotation submissions against the project's label set at the API layer, rejecting invalid labels before persistence.

Provides a Vue.js-based single-page application with task-specific annotation components (text classification buttons, sequence labeling span selection, seq2seq text input). The frontend implements real-time validation of annotations before submission, keyboard shortcuts for rapid annotation, and a responsive design supporting desktop and mobile devices. The UI maintains local state for unsaved annotations and syncs with the backend via REST API calls.

Provides Docker Compose configuration for containerized deployment of Doccano with separate services for frontend (Node.js), backend (Django), database (PostgreSQL), and task queue (Redis/Celery). Environment variables control database credentials, secret keys, and service endpoints, enabling deployment across development, staging, and production environments without code changes. Docker images are published to Docker Hub for easy distribution.

Implements sequence labeling (NER) annotation with a token-level UI where annotators select text spans and assign entity labels. The backend tokenizes documents on import and stores annotations as character-offset spans, which are converted to BIO (Begin-Inside-Outside) tags during export for compatibility with standard NLP tooling. The frontend highlights selected spans in real-time and prevents overlapping annotations.