How to understand DIET entity extractor

Hi! I’m trying to understand how to work DIET entity extractor. I watched the Algorithm Whiteboard series but I still don’t understand how diet performs entity extraction and how the decoder part is used in the transformer layer. If the training phrase is “play ping pong” whose entity is game_name, and the value is “ping pong”. The outputs of the transformer for “play” “ping” and “pong” are three vectors of dimension 256 entering a feedforward layer (CRF). The outputs of this layer are combined with the input of tag 0, B-game_name, and L-game_name to make the prediction? Could someone clarify for me how the entity extractor works or recommend me documentation for dummies? Thanks and sorry for my ignorance!

1 Like