Singapore's government translator learns local terms, and beats a general AI model at them

Singapore has launched SG Translate 2.0, an updated national machine translation system for public services built by the Ministry of Digital Development and Information (MDDI) and GovTech, according to an MDDI factsheet reported by OpenGov Asia. It combines large language models with neural machine translation to help agencies communicate consistently in English, Chinese, Malay and Tamil.

The aim is to capture local terminology, expressions and intended meaning rather than translate word for word. The system draws on more than 7,000 official terms in the Government Terms Translated database and was trained on translations of past government communications.

Tested against a general model

Public officers, professional translators and volunteer Citizen Translators gave feedback during development. In early testing, MDDI says the system handled Singapore-specific terminology and context four times better than a general-purpose large language model. A pilot with 177 Citizen Translators compared its output with human translations.

Why a national tool

Government translation is a quiet but important service. Health advisories, tax letters and emergency messages need to be understood by everyone, and a mistranslated official term can confuse or mislead. General AI tools are good at everyday language but weak on the vocabulary of a particular state's laws and schemes.

Singapore's approach, a shared tool trained on the government's own glossary and checked by the communities who use the languages, is one other multilingual governments can copy. The essential ingredient is not the model but the curated terminology and the human reviewers behind it.