Estonian has no word for “bread”. This can pose a problem for the unwary translator. In particular, it can pose a problem for the makers of a flashcard program I bought, who translated the word “bread” as “leib” and illustrated it with a photograph of a loaf of sliced white bread.
“Leib” is the Estonian word for black bread. The Estonian word for white bread is “sai”. But if you look up bread in a typical English-Estonian dictionary, it may not explain that distinction. When an Estonian sees the English word “bread”, he thinks of leib—that’s what they eat. He knows, intellectually, that the word also encompasses sai, but the concept of bread as an American understand it, of leib-and-sai, doesn’t come to mind as readily or as naturally to an Estonian. So the poor maker of the flash card program, who is after all just adapting the same English words and photos to dozens of different languages, looks up “bread” in the dictionary and gets the wrong answer. Or maybe he gives a list of words to be translated to an Estonian consultant, but without photos attached, and gets the wrong answer. And even if he got the right answer, he won’t be able to convey that right answer on a flashcard.
Conceptually, flashcards are based on the presumption that a word in one language has a unique translation into another language. That presumption is a useful approximation, in the same way that a spherical point mass is a useful approximation of a cow. And understanding why that presumption often fails illuminates why good machine translation is so difficult, and so far off. After all, simply to translate a sentence containing the word “bread” into Estonian requires information that may not be in the text—anywhere.