Machine Learning for Emoji Trends

instagram-engineering.tumblr.com

46 points·mjs··7 comments

If anybody else is confused why the first usage of the heart emoji (️) in text renders as a black heart (), it's because the page lost the 0xFE0F VARIATION SELECTOR-16 invisible character that comes after the U+2764 HEAVY BLACK HEART character to turn it into the emoji form.

Edit: What the, HN deleted both instances of U+2764 HEAVY BLACK HEART in the text, and left the VARIATION SELECTOR-16 behind. That's super weird.

Pretty cool, I like how they figured out the French flag emoji is associated with italy, #montreal.

It reminds me of the Android app dango which does a pretty good job at suggesting emoji associated with full sentences

This seems like it took very heavy influence from http://getdango.com/emoji-and-deep-learning/

Or actually it could be the opposite, seeing as this post is about a year old

I like this app that applies deep learning to Emojis: http://getdango.com/emoji-and-deep-learning/

This was posted in 2015, but now that fasttext (https://github.com/facebookresearch/fastText), just released by Facebook and can scale to Instagram-sized datasets, can create word vectors better than word2vec which account for context (https://arxiv.org/pdf/1607.04606v1.pdf), this type of analysis will only improve in the future.

I imagine that a natural language AI which uses emojis would be better at passing a Turing test

Deep learning representations for anything, be them words, phrases, emojis, musical taste or images are not yet there. They form only the "receptive" part of an AI agent. What sits on top of them is the behavioral part, which is implemented with reinforcement learning (like AlphaGo which learned to "behave" in the world of Go)

Reply on news.ycombinator.com