The knowledge retriever takes the context-aware question and question-response pair embeddings $\hat{x}_{1:t}$ and $\hat{y}_{1:t-1}$ as input and outputs a retrieved knowledge state $h_t$ for the current question. $h_t = f_{kr}(\hat{x}_1, \dots, \hat{x}_t, \hat{y}_1, \dots, \hat{y}_{t-1})$ The input to the prediction model is a vector that concatenates both the retrieved knowledge $h_t$ and the current question embedding $x_t$; this input goes through another fully-connected network before finally going through the sigmoid function to generate the predicted probability $r_t \in [0, 1]$.