reference is only embedded to S with the Glove embedding. and input to LSTM. I think it just makes the task less difficult, or even the learning ability of graph networks less . For example, without reference description of target, GCNN needs to learn cup usually on the desk, but with reference description of target, GCNN doesn't need to learn it.
reference is only embedded to S with the Glove embedding. and input to LSTM. I think it just makes the task less difficult, or even the learning ability of graph networks less . For example, without reference description of target, GCNN needs to learn cup usually on the desk, but with reference description of target, GCNN doesn't need to learn it.