Verified, then logged.
Sun 22 Nov 2026 · 00:34 KST / 15:34 UTC한국어 | ENSearch /

01 / AI · Analysis

Translation AI favours new meanings despite ignore orders

A controlled Mandarin-to-English experiment found that reliance on supplied word meanings varied with word frequency and sentence context.

A researcher seen from behind compares blurred source passages, an external glossary, and contrasting AI translation outputs on two monitors in a language technology lab. AI-generated image

Researchers reported that seven of nine translation models favoured newly supplied word meanings even when explicitly instructed to ignore them in a controlled Mandarin-to-English task.

The paper's abstract explained that external vocabulary resources can help translation but can also conflict with a model's existing word knowledge. The study examined which meaning models followed when these sources of knowledge disagreed.

The researchers assigned new meanings to familiar Mandarin words and asked models to translate them into English. The task included meaning changes within a semantic category and across categories. These conditions provided a way to examine how models chose between existing and supplied meanings.

Word frequency and context shape meaning selection

The researchers defined meanings supplied in the prompt as contextual knowledge (CK). They examined how word frequency, sentence context and the way instructions were framed affected translation outputs.

According to the abstract, models relied more on supplied meanings for less frequent source words. Reliance was also greater when the surrounding sentence supported the supplied meaning.

Larger models rely more on supplied meanings

The researchers reported that larger models within the same family used supplied meanings more often than smaller models. Changing how the prompt was framed could reduce that reliance.

The abstract reported that presenting the supplied meaning less categorically also reduced reliance on it. The effect of repeating information depended on its placement in the prompt.

The researchers found that the effect of reasoning varied across models. The abstract did not disclose whether reasoning increased or reduced reliance on supplied meanings for individual models.

The paper called for methods to detect and evaluate conflicts between existing word knowledge and supplied meanings, and to guide models' choices. Its focus was the choice between those meanings during translation.