How can we prevent AI from confidently making things up during translation?
“AI may arbitrarily introduce, omit, or alter information during translation—these types of 'hallucinations' can be critical errors in translation.”
Root Cause Analysis
Extended Context Window Protection
Shangyi AI employs an extended context window technology. When translating each paragraph, the AI can access several thousand words of surrounding content, ensuring that pronouns (such as he, she, or it) are referenced accurately and not guessed out of context.
Validation Prompt Strategy
When invoking foundational models (such as GPT-4o), we employ a proprietary prompt framework that explicitly requires the AI to produce faithful translations and prohibits it from introducing creativity, particularly regarding data, dates, and proper nouns.
Final Solution Summary
Through multiple technical constraints, Shangyi AI has reduced the incidence of AI hallucinations to an industry-low level.