This is a follow-up to question Completeness vs Compactness in logic 68788. One common theme was that compactness in logic is a purely semantic notion, so should have no need of completeness.
The definition of compactness seems to depend in an irreducible way on the concept of a sentence, which appears to be a syntactic notion. So my question is: Is there any purely semantic definition of a sentence in first-order logic?
One way to completely avoid syntax is to use Ehrenfeucht–Fraïssé games. If $\mathfrak{A}$ and $\mathfrak{B}$ are two structures with the same signature, then two $k$-tuples $\bar{a}$ and $\bar{b}$ from these respective structures satisfy the same type if and only if duplicator wins every finite length EF game between $(\mathfrak{A},\bar{a})$ and $(\mathfrak{B},\bar{b})$. This is a perfectly semantical way of defining the space $S_k$ of $k$-types for a given signature. One can show directly that the resulting space $S_k$ is a compact zero-dimensional space. The $k$-ary formulas of the language basically correspond to clopen sets in $S_k$. The compactness of $S_0$ is basically a restatement of the Compactness Theorem. Similarly, the Omitting Types Theorem is basically a restatement of the Baire Category Theorem for these spaces $S_k$.
This is the approach used by the Fraïssé school of model theory. One recent book which tries to promote this approach is Bruno Poizat's A Course in Model Theory.
I suspect you've misunderstood what you quoted from the earlier question, since the semantics discussed there was the semantics of sentences, presupposing a notion of sentence. Nevertheless, your present question makes sense, and the (probably rather uninformative) answer is yes. The idea is to mimic semantically the definitions of formulas and their satisfaction, without ever mentioning the strings (or trees, etc.) of symbols that syntax deals with. Fix a vocabulary $L$ and (to avoid set-versus-class problems) consider only $L$-structures included in some fixed large set. "Formulas" will, in this approach, be certain collections of pairs $(A,s)$ where $A$ is a structure and $s$ is a finite tuple of elements of $A$. (The length of the tuple $s$ is to be the same for all pairs in a specific formula, but it may be different for different formulas). To define which collections of pairs count as formulas, one uses an inductive definition whose main clauses look like
If $R$ is an $n$-ary predicate symbol of $L$ then there is a formula consisting of exactly those pairs $(A,s)$ where $s$ is an $n$-tuple satisfying $R$ in $A$.
If $\phi$ and $\psi$ are two formulas in which the $n$-tuples have the same length, then $\phi\cup\psi$ is a formula.
The complement of any formula $\phi$, within the collection of all pairs $(A,s)$ whose tuples $s$ have the same length, is again a formula.
If $\phi$ is a formula whose tuples have non-zero length, say $n+1$, then there is a formula consisting of those pairs $(A,s)$ in which $s$ is an $n$-tuple and $(A,(a)^\frown s)\in\phi$ for some $a\in A$.
If $\phi$ is a formula whose tuples have length $n$ and if $f:\{1,2,\dots,m\}\to\{1,2,\dots,n\}$ then there is a formula consisting of exactly the pairs $(A,s\circ f)$ with $(A,s)\in\phi$.
The first clause says that atomic formulas are formulas; the next three say that the class of formulas is closed under the usual first-order constructors (disjunction, negation, and existential quantification), and the last allows trivial manipulations of the components in tuples (permutation, identifications, and adding dummy components).
