The concat (concatenation) attention scoring function is one of the three attention scoring functions proposed by Luong et al. (2015). It calculates the alignment score between an encoder hidden state $\mathbf{h}$ and a decoder hidden state $\mathbf{h}'_t$ as:

$\text{score}(\mathbf{h}, \mathbf{h}'_t) = \mathbf{v}_a^T \tanh(\mathbf{W}_a [\mathbf{h}; \mathbf{h}'_t])$

where:

• $\mathbf{h}$ is the encoder vector (source hidden state).
• $\mathbf{h}'_t$ is the decoder vector (target hidden state at time step $t$).
• $\mathbf{W}_a$ and $\mathbf{v}_a^T$ are learnable weight parameters.
• $[\mathbf{h}; \mathbf{h}'_t]$ represents the concatenation of the encoder and decoder states.

In a neural network, this is implemented by concatenating the encoder and decoder states, applying a dense layer with a $\tanh$ activation, and projecting the result to a single scalar to produce the final score. During training, this mechanism learns which source words are most influential for generating target words.