In the Dynamic Key-Value Memory Network (DKVMN), the first step of its working mechanism is getting the attention weight. The knowledge component $q_t$ is extracted from the embedding matrix and used as a query $k_t$ to search the key memory matrix $\mathbf{M}^k$. This produces an attention weight vector that measures how much attention should be paid to each value in the value memory matrix:

$w_{ti} = \text{Softmax}(\mathbf{M}_i^k k_t)$

where $\sum_{i=1}^N w_{ti} = 1$, $\mathbf{M}_i^k$ is the $i$-th row vector of the key memory, and $w_{ti}$ is the $i$-th element of the attention weight vector.