In a May 29, 2025 paper, a group of researchers from the Harbin Institute of Technology proposed a method to improve AI translation performance in large language models (LLMs) by addressing a common issue: semantic ambiguity caused by context-sensitive words.
These include polysemous words — terms with multiple meanings depending on context — as well as domain-specific or culturally nuanced expressions.
They explained that LLMs still often struggle when a word has more than one possible meaning. For instance, the word “bank” can mean a financial institution or the side of a river. If the model lacks enough context or fails to reason correctly, the translation may be inaccurate or even nonsensical. In some cases, the model may avoid translating the sentence altogether.
To tackle this, they propose an approach called Dynamic Focus Anchoring (DFA). The idea is to help LLMs identify the most challenging parts of a sentence — words that are particularly sensitive to context — and guide the model to pay special attention to them during translation. Notably, this is done without modifying the model or requiring additional training data.
The method works in two steps. First, it identifies the context-sensitive words using both external bilingual dictionaries and the LLM’s own internal knowledge. The system focuses on three types of terms: those with multiple meanings, technical or domain-specific terms, and culturally specific vocabulary.
2024 Slator Pro Guide: Translation AI
The 2024 Slator Pro Guide presents 20 new and impactful ways that LLMs can be used to enhance translation workflows.
Once these words are identified, the second step involves modifying the prompt given to the LLM. The enhanced prompt instructs the model to make sure these specific words are translated accurately — without providing the correct answer — prompting the model to reason through the context using its own knowledge.
Focused Prompting
This process helps the LLM retrieve and apply relevant background information more effectively, leading to more accurate translations. It also avoids overwhelming the model with long lists of possible meanings, which can confuse rather than help.
Tests on WMT22 datasets showed that adding this type of focused prompting improved translation quality across both similar (English-German) and distant (English-Chinese) language pairs.
The researchers also found that the method works best when focusing on a limited number of terms — up to eight — per sentence. Including too many context-sensitive words made the prompts too long and less effective.
Additionally, removing any of the three categories of challenging words led to a drop in translation quality, highlighting the importance of considering all types.
In conclusion, the researchers noted that although the proposed method is designed for AI translation, the underlying idea — guiding LLMs to focus on key terms — may also prove useful in other natural language processing tasks where context matters.
Authors: Qiuyu Ding, Zhiqiang Cao, Hailong Cao, and Tiejun Zhao