I assume you mean "correctly predict", not just make a prediction.
If you impose certain requirements on the input data, it could. First, the corpus would have to not only have the language string, it would have to have a systematic English translation (e.g. "mouth" is translated as "mouth" and not as "beezer; piehole; cakehole; yap"). Second, it would have to contain enough paradigmatically related forms, for example "my N; I found an N; the N disappeared", so that the puzzle-solver can hold constant all of the the variables and solve for "my", "I", "found", etc. and then of course "eye". It would also have to contain at least one token of the word "eye" in a frame that is unambiguously parsable at least w.r.t. the slot containing the word "eye".
On the other hand, if you have an exhaustive corpus of every French word, phrase and sentence ever uttered / written but excluding any instance of "oeil", you have no way of predicting that the French word for eye is "oeil" (even if you know that the word for eyes is "yeux").