This can also be interpreted as selecting the action according to the policy $\boldsymbol{p}_t = (p^1_t,\ldots,p^N_t)$ where $p^i_t = \dfrac{w_t^i}{W_t}$.
Submit our decision $\hat{y_t}$ and suffer a loss $l(\hat{y_t}, y)$.
Note Hedge needs full information, i.e., the whole vector $\boldsymbol l_t$ should be observed. The procedure above, however, only observes $l^{i_t}_t$ which is the loss for the selected action. Hence, to apply hedge, we need to find an estimated loss vector $\tilde{\boldsymbol{l_t}}$ by exploiting the observed loss $l^{i_t}_t$.
Estimation Method: Importance Weighting (IW)
$\tilde{\boldsymbol{l_t}} = (\tilde{l}^1_t,\ldots,\tilde{l}^N_t)$, where $\tilde{l}^i_t = \frac{l^{i_t}_t}{p^i_t}$ if $i=i_t$ and $\tilde{l}^i_t = 0$ if $i\neq i_t$
Property: $\tilde{\boldsymbol{l_t}}$ is the unbiased estimator of $\boldsymbol{l_t}$
Proof:
$$\mathbb E [\tilde{l}^i_t] = \mathbb P(i=i_t) \mathbb E [\tilde{l}^i_t|i=i_t] + \mathbb P(i \neq i_t) \mathbb E [\tilde{l}^i_t|i\neq i_t]$$
$$=\mathbb P(i=i_t) \mathbb E [\tilde{l}^i_t|i=i_t]$$
Then we get
$$\mathbb{E} [\Phi(t+1)-\Phi(t)] \geq \boldsymbol p_t^T \boldsymbol{l_t} - \frac{\eta}{2}\sum_{i\in N}p^i_t \mathbb{E}[(\tilde{l}^i_t)^2]$$
We have
$$\mathbb{E}[(\tilde{l}^i_t)^2] = \mathbb{E}[(\tilde{l}^i_t)^2|i=i_t]\mathbb P(i=i_t) + \mathbb{E}[(\tilde{l}^i_t)^2|i\neq i_t]\mathbb P(i\neq i_t)$$