Figure 4.
![Total Framework. In the encoder, besides the relation-aware self-attention layer, we add a pruned self-attention layer to abandon the irrelevant elements in the sequence. The relation selection mechanism is used in both type layers. For the decoder, we have the same implementation with RAT-SQL [5].](https://mitp.silverchair-cdn.com/mitp/content_public/journal/dint/5/2/10.1162_dint_a_00194/1/dint_a_00194.figure.4.jpeg?Expires=1716561736&Signature=EjvX2ZPzBnxSUaozzUWWWaGUdljEpsrFnXM8PvjG2gVdijm~T8X2tFT~itFHuSrAj4hWb-osOJ85YZEXgaEpO8NGjr5kZuQphN~ktrx-tjCjWvNIs-QcDMVLhupgjlx3DLdV8i7UljpE6vOHTsgBZq6e6ijIBcOia3o-WH8Dfp8WpvrG3qL1MHoparDNoO-AKp3ajQuEIr~QXxkedsEiBvkLM9D9VdlY7YYExpOcfcH2ba7wsiFFTpBIUoj1VYIWtvCSmia6UBRmc~WGwSNpJR25elXh~tOKijWwJ4SItECk6M1d8FOT6~ZxbMHUG3FQ3B2i7u2lXj~Vco9Ozl5G4w__&Key-Pair-Id=APKAIE5G5CRDK6RD3PGA)
Total Framework. In the encoder, besides the relation-aware self-attention layer, we add a pruned self-attention layer to abandon the irrelevant elements in the sequence. The relation selection mechanism is used in both type layers. For the decoder, we have the same implementation with RAT-SQL [5].
