This is because there are multiple things about this which are a problem for ideal recognition with these tools: it has high uncertainty because it is a different language, and language models aren’t designed to work with a single word. I expect changing the weight to affect this outcome, but if it doesn’t, that is the answer on whether this approach will work.