(Photo by Markus Winkler on Unsplash)

Text analysis is not always easy.

Computers can't always make sense of a pile of words. Newer-age neural networks and genAI are a big leap forward from traditional natural language processing (NLP) but they, too, can get tripped up. And it gets worse with short-form text like social media posts – often a mess of abbreviations, slang, and typos – which play hell on content moderation systems.

Emoji should make this easier, right? Your system could use those icons as context for the message at hand. "A picture is worth a thousand words," and all that.

Yes. Sort of.

But … which thousand words? When you're designing such a system, you get to decide what each emoji means. That will shape what the underlying model makes of that emoji. Unless your decision aligns with that of the person who wrote the original text, though, you're headed for trouble.

Consider these three examples:

1/ Cultural context: "😘" – In the US, this is seen as "throwing a kiss" or even a "kissing face." Among my French friends, though, this represents the "bisou" – the light cheek-kiss, often done in pairs, used as a greeting between close friends and family.

The US interpretation might be neutral or it might be romantic. So already there's room for confusion there. But the French bisou carries zero romantic interpretation. An automated system that assumes either meaning of "😘" will mis-classify messages that were sent with the other meaning in mind.

2/ Personal taste: "😜" – This emoji mostly carries a light-hearted "wtf?" or "haha that's crazy" meaning in my social circles. But to some people this carries a slightly sexual vibe.

There's no right or wrong answer here, just a strong potential for misunderstanding. If your system treats "😜" as a joke, it will accept messages that some people find inappropriate. Or if your system treats that emoji as a sexual expression, it will mistakenly flag a lot of jokes.

3/ Different images: "🔫" – Emoji appear as icons but they are technically characters behind the scenes*,* defined by a Unicode standard. Individual platforms interpret the standard's descriptions and decide which icons to display. In other words, different systems may show different icons for a single emoji character-code.

Unicode character U+1F52B is described as a pistol. In 2016, iOS 10 replaced their image of a gun with one of a water pistol, because a gun was considered too violent. Fair enough! Except that other platforms still rendered this emoji as a plain old gun. "Water pistol" and "firearm" are nowhere near the same thing. Yet again, a system that relies on "🔫" for any decision will stumble.

In fact, depending on how your browser or OS interprets the Unicode standard, all three emoji I have included above may render differently for you than they did for me. And that rendering may weigh in your interpretation.

Notice, those were just three examples of potentially troublesome emoji. I didn't mention the host of emoji that have taken on new life as slang. (I will not mention those emoji here, in part to make this a timeless post and also because I don't want this post to be flagged by content moderation systems…)

The takeaway lesson: for anyone who does text mining or other text analysis, beware how you interpret emoji. Those interpretations live on in every analysis and model you build, and can lead to unintended downstream effects.