Shockingly inefficient PERL script on Google n-gram

Some colleagues developed a Perl script that compares the similarity of two sentences using Google n-grams. The n-gram files are huge, and without knowing Perl, we believe they have done nothing to optimize retrieval from the n-gram files. Each sentence comparison now takes an average of 7 minutes, and since we have about 500,000 sentence pairs to compare, this task would take almost 7 years to run. We need the speed improved by two orders of magnitude, to an average of 4.2 seconds per comparison. We suspect a simple initial indexing of the n-gram files to at the start of the process may take care of the problem. It would be ok for the system to take up to an hour at the startup to do any indexing and storing in memory. Up to 20GB of memory may be used to store the indexed data.