Fitting a regular grammar to strings from a PCFG: how big does it get?

Question

Let $G=(V, \Sigma, R, S)$ be a (non regular) probabilistic context-free grammar, and $u_1, \ldots, u_n$ a set of $n$ strings generated by $G$.

For finite $n$, it is always possible to find a regular grammar $\hat G=(\hat V, \Sigma, \hat R, S)$ which generates the strings $u_1, \ldots, u_n$.

Intuitively, as $n$ goes to infinity, we expect $\hat G$ to get larger: my guess is that the cardinal of $\hat R$ (and maybe also the cardinal of $\hat V$?) would need to go to infinity.

Are there results which formalize this, e.g. by giving a lower bound on these cardinals as a function of $n$?

Are you referring to strictly regular grammars or are extended ones also allowed? — dkaeae, Dec 20 '18 at 13:08
@dkaeae I am mostly interested in strictly regular grammars, but would also welcome answers related to extended ones. — Robin Ryder, Dec 20 '18 at 13:38
If we replace probabilistic context-free grammar by (usual) context-free grammar in the question, does it make any difference? If it does, please clarify. — John L., Dec 21 '18 at 08:01
@Apass.Jack I would be equally happy with an answer about usual CFGs. Using PCFGs means that it might be possible to get a probabilistic statement, e.g. about the expected value of the cardinal of $\hat R$. — Robin Ryder, Dec 21 '18 at 09:24

score 1 · Answer 1 · answered Dec 26 '18 at 10:16

1

The question, if understood in the simplest naive way, might be uninteresting.

Here is a simple example. Let $G$ be the regular language $\{a^n\mid n\ge0\}$ over the alphabet $\{a\}$.

Consider the strings $\epsilon, a, a^2, \cdots, a^n$ in $G$. What are the regular languages that contains those strings?

The minimal such language, i.e., which contains no other strings, will need $n+1$ generation rules $S\to a^i$ for $0\le i\le n$.
The language with the least generation rules is $G$ itself, which has two generations rules $S\to \epsilon$ and $S\to aS$.

However, once we start to twiddle with the ways how to approximate context-free grammar by regular grammar, there are tons of research.

Here is a related question that links to many related stuff, Is there a known method for constructing a grammar given a finite set of finite strings?.

This paper considers approximating CFG from above by a regular grammar.

You can browse the google search result for approximate context-free grammar by regular grammar or the google scholar.

answered Dec 26 '18 at 10:16

John L.

38,985
4
33
90

Practical Experiments with Regular Approximation of Context-Free Languages seems interesting. – John L. Dec 26 '18 at 10:55
Thank you for the links! However, I don't understand your example, since your $G$ is regular, so of course $\hat G=G$ works. – Robin Ryder Dec 28 '18 at 08:11
That is exactly the point. I am saying there are ways to interpret the requirement of $\hat G$ that makes the solution uninteresting. – John L. Dec 29 '18 at 09:17
I tried to phrase the question in a way that would specifically avoid this case, with $G$ a non-regular (P)CFG. – Robin Ryder Dec 29 '18 at 10:39
I noticed your attempt to emphasize non-regular-ness. However, you can just attach another piece to $G$, such as ${a^nb^n\mid n\ge0}$ to make it non-regular. – John L. Dec 29 '18 at 12:24
Basically, the first question is how to estimate/measure/evaluate how close is a CFG to a regular language. There are many ways. My example tries making it obvious that a simplistic measurement might not work perfectly. – John L. Dec 29 '18 at 12:34

Fitting a regular grammar to strings from a PCFG: how big does it get?

1 Answers1