greedy decoding

Topics

• decoding strategies
• greedy algorithms

 The simplest decoding method which selects the most probable token at each step $t$.
In an encoder-decoder architecture where we condition on input $X$ and previous token outputs ${\hat{y}}_1,\dots ,{\hat{y}}_{t-1}$

$${\hat{y}}_t=\underset{w\in V}{\mathrm{argmax}}{P}_{\theta }({\hat{y}}_t=w\mid {\hat{y}}_{1:t-1},X)$$

Note that in practice, we don’t directly condition on input $X$, but use the context variable $\mathbf{c}$ produced by the encoder to represent the input sequence.

Related

• beam search
• contrastive search