How does InjType work in the test phase? #1
Description
As I understand it, In training phase, Injtype seperates the text generation goal into three steps: (1) generating entity indicator or contextual words, (2) generating entity mention using entity indicator and hidden state, (3) adding a GRU module in a reversed direction.
But what confused me is, what's the generative process in test phase? Does Injtype generate entity indicator first and then generate the mention, with the use of beam search strategy?
Please help. Thanks. @wyu97