inception-project/inception

What it solves

It simplifies the creation of high-quality annotated text corpora. It addresses the difficulty of manually labeling large amounts of text for machine learning training data, the need to ground annotations in specific ontologies, and the need to maintain consistency across multiple annotators.

How it works

The platform allows users to define custom annotation schemes and layers (such as entities, relations, and coreference) directly in the browser. It integrates with knowledge bases (RDF, OWL, SPARQL endpoints) to link entities to existing ontologies. To speed up the process, it uses machine-assisted recommenders that learn from the user's current annotations in real-time and active learning to suggest labels for the remaining text.

Who it’s for

It is designed for researchers and institutions building text datasets for NLP, particularly those who need complex, multi-layer annotations grounded in formal knowledge bases.

Highlights

• Knowledge-Base Integration: Supports linking to Wikidata, SNOMED CT, and other ontologies via RDF or SPARQL.
• Machine-Assisted Labeling: Recommenders train on the fly and use active learning to prioritize uncertain cases.
• Multi-Layer Support: Handles overlapping layers for syntax, frames, and document labels.
• Quality Control: Includes tools to measure inter-annotator agreement and curate gold standards.
• Flexible Deployment: Available as a desktop installer or a server deployment with single sign-on.
• Extensible: Provides a REST API and support for external recommenders via custom models.