Most image-similarity tools only notice how things look (color, shape, class) and miss deeper, human-like connections.
This paper teaches computers to understand words by also looking at the smaller pieces inside words, like 'un-', 'play', and '-ing'.