Lesan turned offline print into a Horn-of-Africa language dataset

Content moderation in Amharic and Tigrinya faced a basic data problem: the languages were underrepresented online, so there was little digitized text to train language-specific systems. Multilingual models transferring lessons from English could miss cultural meaning, and some models did not even contain characters in the Tigrinya alphabet.

Lesan chose to work from offline resources instead of waiting for a one-model solution. The Berlin-based startup coordinated with Tigrinya-speaking communities to scan books and magazines, then built custom character-recognition tools to make the print readable by machines.

The grassroots collection produced a benchmark dataset for languages across the Horn of Africa. Lesan then partnered with the Distributed AI Research Institute to develop an open-source tool that identifies languages spoken in Ethiopia and detects harmful speech in them.

Lesan created a Horn-of-Africa benchmark dataset and partnered with DAIR on an open-source tool for Ethiopian language identification and harmful-speech detection.

Low-resource languages had too little digitized text to support reliable language-specific moderation models.

Cross-lingual transfer through English could erase cultural nuance and produce errors or false alarms.

Tigrinya-speaking communities could supply relevant print materials and language knowledge that global datasets lacked.

Custom character-recognition tools converted scarce offline material into reusable machine-readable data.

When a language has too little digital data for a general model, the decisive move may be to digitize community-held print with local language expertise before scaling the tool.

Lesan used its grassroots data-collection method to create a benchmark dataset for languages across the Horn of Africa, then joined with DAIR on an open-source language-identification and harmful-speech tool for Ethiopia. The article describes early progress but gives no adoption, revenue, or accuracy metric, so the result is best understood as a data and partnership breakthrough rather than a quantified commercial turnaround.