Using a previous Stockfish scorer they trained an NN without any labelling effort. This is also similar to how unsupervised translation is done in some methods.

Using a previous Stockfish scorer they trained an NN without any labelling effort. This is also similar to how unsupervised translation is done in some methods. They start from word->word dictionary results and iteratively train lang1->lang2 and lang2->lang1 models feeding on each other's output.