Compress data with context free grammars

Question

It is possible to compress some kinds of data, such as human text or source code, with straight-line grammars. You basically create a grammar whose language has exactly one word – the uncompressed data. In this task, you have to write a program that implements this method of data compession.

Input

The input is a string of not more than 65535 bytes length. It is guaranteed, that the input matches the regular expression [!-~]+ (i.e. at least one printable ASCII character excluding whitespace).

An example input is

abcabcbcbcabcacacabcabab

Output

The output is a set of rules that form a grammar that describes exactly one word (the input). Each nonterminal is denoted by a decimal number greater than 9. The start symbol is symbol number ten. An example output corresponding to the example input is given below; its syntax is described further below:

10=11 11 12 12 11 13 13 11 14 14
11=a 12
12=b c
13=a c
14=a b

Each rule has the form <nonterminal>=<symbol> <symbol> ... with an arbitrary whitespace-separated number of symbols on the right side. Each output that obeys the following restrictions and derives exactly the input string is valid.

Restrictions

In order to stop people from doing strange things, a number of restrictions are taking place:

• Each nonterminal must appear at least twice on the right side of a rule. For instance, the following grammar for the input abcabc is illegal because rule 12 appears only once:

    10=12
    11=a b c
    12=11 11
  

• No sequence of two adjacent symbols may appear more than once in all right-hand sides of all rules, except if they overlap. For instance, the following grammar for the input abcabcbc is illegal since the sequence bc appears twice:

    10=11 11 b c
    11=a b c
  

  A valid grammar would be:

    10=11 11 12
    11=a 12
    12=b c
  

• Your program must terminate in less than one minute for each valid input that is not longer than 65535 bytes.

• As usual, you may not use any facility of your language or any library function that makes the solution trivial or implements a large part of it.

Sample input

Generate sample input with the following C program.

#include <stdlib.h>
#include <stdio.h>

int main(int argc, char **argv) {
  unsigned int i,j = 0,k;

  if (argc != 3
     || 2 != sscanf(argv[1],"%u",&i)
      + sscanf(argv[2],"%u",&k)) {
    fprintf(stderr,"Usage: %s seed length\n",argv[0]);
    return EXIT_FAILURE;
  }

  srand(i);

  while(j < k) {
    i = rand() & 0x7f;
    if (i > 34 && i != 127) j++, putchar(i);
  }

  return EXIT_SUCCESS;
}

The sample input generated by the program above will usually not result in good compression results. Consider using human text or source code as example input.

Winning criteria

This is code golf; the program with the shortest source code wins. For extra credit, write a program that reconstructs the input from the output.

Answer 1

GolfScript, 111 108 characters

1/{.}{:^1<{^1$/,2>.{;,)^<.0?)!}*}do-1<.,1>{^1$/[10):10]*0+\+}{;^}if(\}while][0]%.,,]zip{))9+`"="+\~" "*+}%n*

This is a quite clumsy approach using GolfScript. The second version performs much better than the initial one. It is much longer than the intended code but my implementation had nested do-loops and this caused issues with the interpreter.

Examples:

> abcba
10=a b c b a

> abcabcbc
10=11 11 12
11=a 12
12=b c

> abcabcbcbcabcacacabcabab
10=11 12 12 13 14 14 c 11 15
11=15 13
12=c b
13=14 b
14=c a
15=a b