Markov-chain based text generator for Raku language

SYNOPSIS

    use Text::Markov;
    
    my $mc = Text::Markov.new;
    
    $mc.feed( qw{Easy things should be easy and hard things should be possible.} );
    $mc.feed( qw{People who live in glass houses should not throw stones.} );
    $mc.feed( qw{Live and let live.} );
    
    say $mc.read( );
    # People who live in glass houses should be easy and let live.

METHODS

Markov chain is a mathematical system.

To understand terminology used below read OPERATING PRINCIPLE paragraph first.

new( order => 2 )

Order (optional, default 1) controls how many past states determine possibe future states.

feed( 'I', 'like', 'pancakes.' )

Add transitions of states.

State can be represented by any object that can be identified by nonempty String.

feeder( $sequence )

Stream version of feed( ), allows to load transition of states from Sequence. Useful for feeding large inputs on the fly, like for example whole book word by word.

read( ) / read( 128 )

Generate chain of states up to requested length (optional, default 1024).

reader( ) / reader( 'I', 'like' )

Stream version of read( ). Returns lazy Sequence that will provide states. Useful for generating large (potentially infinite) outputs on the fly.

Accepts initial states, provided list must be no longer than chain order.

OPERATING PRINCIPLE

Let's put abstract hat on and imagine that each word represents state.

Therefore sentence made of words can be represented as transitions between states.

For example sentence I like what I see is expressed by the following graph:

                            4     +------+
                     +------------| what |<----+
                     |            +------+     |
                     |                         |
                     v                         | 3
+-------+    1     +---+    2     +------+     |
| START |--------->| I |--------->| like |-----+
+-------+          +---+          +------+
                     |
                     |
                     |      5     +-----+    6     +-----+
                     +----------->| see |--------->| END |
                                  +-----+          +-----+

It may be surprising but transition number is not important for feed and can be discarded.

Instead of that transitions counters are stored (in this example each transition occured only once):

                           1x     +------+
                     +------------| what |<----+
                     |            +------+     |
                     |                         |
                     v                         | 1x
+-------+    1x    +---+    1x    +------+     |
| START |--------->| I |--------->| like |-----+
+-------+          +---+          +------+
                     |
                     |
                     |     1x     +-----+    1x    +-----+
                     +----------->| see |--------->| END |
                                  +-----+          +-----+

Next sentenceNow I see you like cookies when passed to feed will simply add new transitions or increase counters of already existing ones in the same graph:

                           1x     +------+
                     +------------| what |<----+
                     |            +------+     |
                     |                         |
                     v                         | 1x
+-------+    1x    +---+    1x    +------+     |
| START |--------->| I |--------->| like |-----+
+-------+          +---+          +------+
    |               ^ |            ^    |
    |               | |         1x |    |
 1x |            1x | |            |    | 1x
    |   +-----+     | |        +-----+  |        +---------+
    +-->| Now |-----+ |        | you |  +------->| cookies |
        +-----+       |        +-----+           +---------+
                      |            ^                  |
                      |            | 1x               | 1x
                      |            |                  v
                      |    2x     +-----+    1x    +-----+
                      +---------->| see |--------->| END |
                                  +-----+          +-----+

Markov chain is generated by making transitions from the current state to one of the next possible future states with respecting probability assigned to each transition. The higher the counter the more probable transition is.

Let's generate:

• From START transition can be made to I [50% chance] or Now [50% chance] - I is rolled.
• From I transition can be made to like [33.(3)% chance] or see [66.(6)% chance] - like is rolled.
• From like transition can be made to what [50% chance] or cookies [50% chance] - cookies is rolled.
• From cookies transition can be made only to END [100% chance].

New sentence I like cookies is generated!

Note that it is not subpart of any sentence that was used by feed to create graph, yet it has correct grammar and makes sense.

Improving output quality

Default setup will produce a lot of nonsense. From sentences...

• I was tired.
• It was snowing.
• Today I was going to do something useful.

...new sentence I was snowing. may be generated.

It happens because single was word does not give enough context to make rational transitions only. Param order => 2 in constructor restricts possible transitions to those which appears after two past states. So from I was only two transitions are possible and more reasonable Today I was tired. sentence may be generated.

This is called Markov chain of order m.

The higher the order the more sensible output but more feed is also required. You have to experiment :)