[guided]The purpose of this step is to translate the moment into cumulants, because the hypothesis defining a semicircular family is a statement about cumulants. We fix $m \in \mathbb{N}$ and indices $i_1,\dots,i_m \in I$, and we write $NC(m)$ for the set of noncrossing partitions of $\{1,\dots,m\}$.
For each noncrossing partition $\pi \in NC(m)$, the contribution of $\pi$ is built block by block. If $V = \{v_1 < \cdots < v_r\}$ is a block of $\pi$, define
\begin{align*}
\kappa_V(s_{i_1},\dots,s_{i_m}) := \kappa_r(s_{i_{v_1}},\dots,s_{i_{v_r}}).
\end{align*}
The order $v_1 < \cdots < v_r$ is the inherited order from $\{1,\dots,m\}$, so the variables are fed into the cumulant in the same order in which they occur in the word $s_{i_1}\cdots s_{i_m}$. The full contribution of $\pi$ is the product over all blocks:
\begin{align*}
\kappa_\pi(s_{i_1},\dots,s_{i_m}) := \prod_{V \in \pi} \kappa_V(s_{i_1},\dots,s_{i_m}).
\end{align*}
The free moment-cumulant formula states that the moment of a word is the sum of these partition cumulants over all noncrossing partitions. It applies here because $s_{i_1},\dots,s_{i_m}$ are elements of the noncommutative probability space $(\mathcal{A},\varphi)$ and $\kappa_r$ denotes the corresponding joint free cumulant maps. Applied to the word $s_{i_1}\cdots s_{i_m}$, it gives
\begin{align*}
\varphi(s_{i_1}\cdots s_{i_m}) = \sum_{\pi \in NC(m)} \kappa_\pi(s_{i_1},\dots,s_{i_m}).
\end{align*}
This is the exact point where noncrossing partitions enter the argument. The remaining work is to use the semicircular hypothesis to determine which terms in this sum can be nonzero.[/guided]