Контекстуальные универсальные встраивания

AbstractWe propose a framework to modularize the training of neural language models that use diverse forms of context by eliminating the need to jointly train context and within-sentence encoders. Our …...