In this paper: Unsupervised Machine Translation Using Monolingual Corpora Only
the authors proposed a novel method.
Intuitively it is an autoencoder, but the Start Of Sentence token is set to be the language type.
One other advanced method is to use the pre-training model. In this paper: Cross-lingual Language Model Pretraining researchers proposed an ...
This is called "clustering" , If the network is already trained with data that has similar features as of the "symbols", you can use that network with its last classification layer removed , then run a clustering algorithm like "k-means" on top of the vectors obtained from the last layer of the network.