[Paper Review] :telephone::person::sailboat::whale::okhand:; or "Call me Ishmael" - How do you translate emoji?

Will Radford, Andrew M. Chisholm+2 more|arXiv (Cornell University)|Nov 7, 2016

Digital Communication and Language11 references3 citations

TL;DR

This paper analyzes Emoji Dick, a crowdsourced translation of Moby Dick into emoji, using NLP techniques like statistical word alignment and POS tagging to study how emoji function as a communication system. It finds that emoji favor nouns and verbs, show high repetition rates, and preserve semantic content despite low lexical diversity, suggesting emoji may operate like a simplified, pidgin-like language in constrained contexts.

ABSTRACT

We report on an exploratory analysis of Emoji Dick, a project that leverages crowdsourcing to translate Melville's Moby Dick into emoji. This distinctive use of emoji removes textual context, and leads to a varying translation quality. In this paper, we use statistical word alignment and part-of-speech tagging to explore how people use emoji. Despite these simple methods, we observed differences in token and part-of-speech distributions. Experiments also suggest that semantics are preserved in the translation, and repetition is more common in emoji.

Motivation & Objective

• To investigate how emoji are used in a high-stakes, context-free literary translation task.
• To examine whether emoji preserve semantic content when translating complex literary text.
• To compare statistical properties of emoji and text corpora, focusing on distribution, repetition, and part-of-speech alignment.
• To explore whether emoji exhibit structural parallels to pidgin languages through POS distribution and alignment patterns.

Proposed method

• Used IBM Model 1 for statistical word alignment between English tokens and emoji, trained on 9,734 aligned sentence pairs.
• Applied POS tagging to both text and emoji to analyze syntactic role distribution, focusing on NN (nouns), VB (verbs), and other parts of speech.
• Computed frequency distributions and rank-frequency plots (log-counts vs. rank) to compare lexical diversity and Zipfian characteristics.
• Measured emoji repetition via bigram analysis, comparing repeat frequencies in emoji vs. text corpora.
• Filtered out stopwords and punctuation, and used NLTK’s default tokeniser and POS tagger for text preprocessing.
• Evaluated alignment confidence by ranking top emoji-token alignments and analyzing POS-specific probabilities.

Experimental results

Research questions

• RQ1 How do emoji distributions and frequencies compare to those of natural language in a high-contrast, context-free setting?
• RQ2 To what extent is semantic content preserved in emoji translations of a complex literary text?
• RQ3 Do emoji exhibit syntactic patterns resembling those of pidgin languages, particularly in part-of-speech distribution?
• RQ4 What role does repetition play in emoji communication, and how does it differ from natural language?

Key findings

• Emoji corpora have significantly lower lexical diversity, with only 470 unique emoji types compared to 16,454 in the text corpus.
• Emoji sentences are roughly half the length of text sentences, with a mean of 9.4 emoji tokens versus 20.8 text tokens per sentence.
• The most frequent emoji are associated with people (e.g. 🧍‍♂️), whales (e.g. 🐋), and actions (e.g. 🛶), with strong alignment to nouns and verbs.
• Emoji show a 3.2% repeat bigram rate, compared to only 0.4% in text, indicating systematic use of repetition as a linguistic device.
• POS distribution analysis reveals that 70% of the highest-probability emoji alignments are to nouns (NN), and 56% to verbs (VB), suggesting a strong syntactic role.
• Despite high noise and subjectivity, the alignment model identifies consistent mappings—e.g., 🐋 aligns strongly to 'whale', 'sperm', and 'whales'—indicating shared semantic reference.