Interpreting Eomji

Today I learned about an attempt to interpret emoji … via emoji. This paper has all the details.

The trick is to train word2vec as you would normally, including vectors for the emoji, and to then look at the differences between two emoji that are “polar opposites” of eachother. This spans a plane which can be used to compare other emoji.

The polar opposites, as far as I could tell, were manually selected and are summarised in the table below.

The paper even goes on and does a human study and concludes that the human interpretation is inline with the method from the paper.

It makes a bunch of sense, though I imagine that a lot of this depends on the dataset that was used to train the emoji embeddings. The paper seems to re-use earlier work but I would imagine the interpretation of emoji to vary across demographies. Still, fun paper to read!