45 — ☎️👨⛵️🐋👌; or “Call me Ishmael” – How do you translate emoji?

This paper uses the 2010 Emoji Dick project — which crowd-sourced the translation of Moby Dick into emoji — as a translation corpus. The authors note obvious differences between the “languages” — emoji sentences, at 9.4 tokens per sentence, are about half as long than English sentences (20.8); there are nearly three times as many unique words as unique emoji tokens; English follows a Zipfian distribution (an exponential falloff of word frequency; the most popular word is twice as popular as the second most common word and three times as popular as the third most common word) while emoji approximate this curve but do not match it. (This is likely a sign of loss-of-specificity; it’s unlikely you could ever translate the emoji text back into English.)

Whereas english tokens follow a smooth part-of-speech probability distribution, emoji follow a very different distribution (that again could look like the same distribution, but with some drastic flaws).

One interesting phenomenon the authors found was the syntactical use of emoji duplication: There was a 3.2% probability of emoji duplication, versus 0.4% in English. This suggests that duplication like “💪💪” is used to convey semantic meaning: “very strong”, instead of just “strong”. In other words, authors seem to find that they cannot demonstrate information as well with emoji and so they resort to other ways of conveying that same meaning.