Grammar extraction using pre-existing grammars.

Given a set of strings $s_1, \dots, s_n$ over $\Sigma$ it isn't clear what a good generalization of the strings would be using a regular language, that extends the set infinitely. This comes from the many interpretations you could have of the set of strings.

What about taking pre-existing grammars like the following two

DIGITS = [0-9]+
ID = (ALPHA | _ ) (ALPHA | _ | DIGITS)*

used in programming language parsing, call them $g_1, g_2$.

Consider permuting the terminals of a grammar, for instance $\phi : a \mapsto b, \ b \mapsto c$ and $g$ is: $$g \to b AB\\ A \to aaA + B \\ B \to abab$$

Then $\phi \circ g$ is $$g' \to c A B \\ A \to bbA + B \\ B \to bcbc$$

Then these alphabet permutations (aka terminal permutations) preserve the structure of the grammars. So now ask whether $s_i \in \phi\circ g_j$ (language membership), for some $j,\phi$, or even if $s_i \in (\phi_1 g_1) \cdot \dots \cdot (\phi_m g_m)$, for some given set of $g_i$ and for some permutations $\phi_i$.

Or what if we took $\Sigma' = \Sigma \cup \{g_1, \dots, g_m\}$ and considered permutations of $\Sigma'$.

Any grammar containing the $\{s_1, \dots, s_n\}$ we construct using these methods has structure derived from the given grammrs $g_i$, and this essentially answers the question as to what a good generalization of a set of strings would be: it's the one that is built from specified, preferrable grammar structures. The specified grammars act like hints to the algorithm as to what grammars to choose. Has anyone investigated this?

Given any finite set $S$ of strings, you can define an extension just by saying that your language is the set of all strings that contain a member of $S$ as a prefix and/or suffix. Or since $S$ is finite, you can take the union of $S$ with any regular language and get a regular language. So it seems like your question is a bit ill-posed. Maybe the "minimal" such language that contains your strings in $S$ is the set you get from applying the pumping lemma criteria to $S$, trying to get as many strings as possible from $S$ into the same pumping class?