Skip to main content

Browser-based Multilingual Translation

Periodic Reporting for period 1 - Bergamot (Browser-based Multilingual Translation)

Reporting period: 2019-01-01 to 2020-06-30

We are making a system to translate between languages that runs directly on desktops and laptops, preserving your privacy. Most other providers run in the cloud, which exposes browsing habits to the provider and, in at least one case, leaked corporate documents. While some apps can translate on phones without Internet, these are generally limited to translating one sentence a time for speed; our research enables translating entire web pages in reasonable time.

Partnering with Mozilla means the project will extend Firefox with private translation functionality.

The system also estimates its own confidence and displays it to users, warning of potentially incorrect translations. This functionality can also help you send an e-mail or fill out a form in a language you do not speak. If the system is not confident about translating part of your e-mail to another language, it will point to the problematic part and invite you to rephrase.
"We have a working demo that translates on Linux and Mac in Firefox; Windows effort is underway. The system currently supports Czech, Estonian, German, and Spanish translation. On a range of consumer laptops and desktops dating from 2012 to 2019, it translates 1000-9000 words per second. The system displays its confidence using colors based on our research on efficiently estimating quality; we are conducting user studies on the best way to present this information. Initial experiments on translating to a language the user doesn't speak show that prompting to rephrase the input improves the understandability of the translation.

Our project featured in over 40 news articles and made #1 on Hacker News."
Free local translation preserves privacy and empowers users who work with sensitive data. Extending Firefox means the project will be widely available. Open source enables extension and the possibility for the community to contribute translation models for underserved languages. The form filling assistant enables users to respond in languages they do not speak, going beyond ordinary web translation.

Our project has pushed boundaries of efficient neural network models, making them faster or supporting larger models on consumer devices. These methods have implications the speed and quality of other systems like captioning and summarization.

Estimating the system's own quality ensures trust in algorithms, a rising societal issue, is placed more appropriately. Moreover, it requires less labeling, so the technology can be applied more widely for languages with less data available.
Screenshot of browser offering to translate