Technical Terminology Verification for Neural MT

Anonymous

TL;DR: We propose a method for dealing with hallucinations that uses fact checking to address the problem of inaccurate machine translations of technical acronyms

Abstract: Translating technical acronyms is a problematic task for MT systems, with an error rate around 50% for Google Translate and around 65% for Opus-mt. Incorrect acronym translation is a fatal error. We present a turnkey solution for translating long form (LF)–short form (SFs) pairs and verifying their use by the scientific community. Since MT models perform better on LFs than SFs, our proposed method takes advantage of this observation to improve translations of SFs, by introducing a novel verification process. This process is motivated by standard practice in professional translation.

Paper Type: short

Research Area: Machine Translation

Contribution Types: Position papers

Languages Studied: French, English