Suppose you train two neural networks on English and French. Each gleans the structure of its respective language, developing an internal representation known as a latent space. Essentially, it is aword cloud: a map of all the associations that words have in that language, built by placing similar words near one another and unrelated words farther apart. The cloud has a distinctive shape. In fact, it is the same shape for both languages because, for their all their differences, they ultimately refer to the same world. All you need to do is rotate the English and French word clouds until they align.[…] “Without having a dictionary, by looking at the constellation of all the words embedded in the latent spaces for each language, you only have to find the right rotation to align all the dots,” Kanai says.
