Can anyone help me find an efficient implementation.

If the cost of a subroutine call is cheaper than scanning the list (which I suspect is the case), then you can assemble all the target patterns into one, perform a single match, and then dispatch to the sub that gives you what you want to substitute:

#! /usr/local/bin/perl -w

use strict;
use Regexp::Assemble;

my $ra;

my %dispatch = (
    'food'   => sub { 'pizza' },
    'water'  => sub { 'beer'  },
    'like'   => sub { 'enjoy' },
    '(\d+)F' => sub { int(($ra->mvar(1) - 32 ) * 5/9 ) . 'C' },
);

$ra = Regexp::Assemble->new( track => 1 )->add( keys %dispatch );

while( <DATA> ) {
    while( $ra->match($_) ) {
        my $m = $ra->matched;
        s/$m/&{$dispatch{$m}}/e;
    }
    print;
}

__DATA__
I'd like a glass of water with my food, it's 92F in here!
[download]

... produces...

I'd enjoy a glass of beer with my pizza, it's 33C in here!
[download]

Generating the dispatch table from a data file is left as an exercise to the reader (but a pretty fun one, I might say).