MIT Solve

Solution Overview

Solution Name:

Lesan AI

One-line solution summary:

Lesan builds machine translation systems for languages in emerging economies to unlock the web’s content for millions of people.

Pitch your solution.

Millions of people around the world have access to the web but there is limited content to consume in their native language. This holds communities back from fully participating in opportunities on the web. Machine translation can solve this problem by instantly translating content to their native language. In the past it has been very challenging to build datasets large enough to serve accurate translations for many of these languages. We found a way to build datasets that are large enough to serve accurate machine translation systems and make them publicly available. For our first two languages in Ethiopia we have had over one hundred thousand users and have served over 9 million translations. Scaling this solution around the world will unlock the webs content for millions by making the web more relevant and consumable for these communities.

Film your elevator pitch.

What specific problem are you solving?

There are hundreds of millions of people around the world that can access the web with their smartphone or computer but have limited relevant content in their native language. These communities are in different parts of Africa, South East Asia and India. This means they miss out on many of the educational and economic opportunities the web can offer because they cannot consume web content. The way to solve this problem historically has been with machine translation. Machine translation instantly translates text from one language to another. Recently there have been large advancements in this field of machine learning but large machine translation service providers like Google and Microsoft do not serve accurate machine translation systems for many low resource languages. A low resource language is a language where it is very hard to build a large dataset to build accurate models. We have found a way to build large datasets with millions of translations for these languages and serve accurate machine translation models for these communities for the first time.

What is your solution?

Our solution is focused on building high quality datasets for low resource languages to empower our machine translation systems. We found a way to build large scale translation datasets for these low-resource languages and train models to serve millions of accurate translations for thousands of people. Building these datasets to train machine learning models has been the main challenge to open up the web's content for these communities because there is very limited online content in these languages and their exact translations in English to build parallel corpora. The first step for our solution is building large datasets of millions of past translations that are accurately aligned for each translated sentence. The next step is to train machine learning models off of these millions of translations so it can be served to the public for free. The end result is an accurate and publicly available machine translation system for anyone to use on our website.

Who does your solution serve, and in what ways will the solution impact their lives?

Millions of people in different countries around the world have access to the web through a smart phone or computer but have limited content they can consume in their native language. These communities can be university and high school students trying to read through an article for their paper, they can be software engineers trying to translate how-to guides in Stack Overflow to their native language. They can also be diaspora communities in Sweden translating immigration documents into their native language. The possibilities for content consumption are very broad. Any internet user can go online to our website and translate content for free. We also provide an API to developers to build tools off of like web plugins or news feeds.

Which dimension of the Challenge does your solution most closely address?

Equip everyone, regardless of age, gender, education, location, or ability, with culturally relevant digital literacy skills to enable participation in the digital economy.

Explain how the problem you are addressing, the solution you have designed, and the population you are serving align with the Challenge.

Millions of people, all around the world, are being left out of the digital revolution that is happening online. This cuts them off from important opportunities that are accessible to those with an internet connection, a device and the ability to comprehend content in a language they can understand. We work to make web content that already exists online available for consumption by these communities. This ensures that when these communities come online, they have a meaningful and rewarding experience. We strive to make sure the communities we serve are active participants in the digital world.

In what city, town, or region is your solution team headquartered?

Berlin, Germany

What is your solution’s stage of development?

Growth: An organization with an established product, service, or business model rolled out in one or, ideally, several communities, which is poised for further growth.

Explain why you selected this stage of development for your solution.

We have launched 6 language pairs so far: Tigrinya <> English, Amharic <> English and Amharic <> Tigrinya. These two low resource languages are popular languages in East Africa and are spoken by 60 million people. We have completed more than nine million translations and have had over one hundred thousand users on our website try out the translator. We are looking to scale out our technology for new languages across Africa and eventually the world. Our users come from all over the world but primarily Ethiopia and North America. We also have document translation clients like Dalberg and McKinsey who use our translation service to translate documents for their teams.

Who is the Team Lead for your solution?

Adam Beaudoin