If your file contains paragraphs (ie. blank lines between reasonable sized block of text, something like this might work. It is pretty quick, processing a 587 MB test file below (replicated several dozen times to get a file of comparable size) in under a minute.

You can adjust the definition of a 'word' to suit. I've specified 1 to 5 words either side to account for the word being at the start or end of a paragraph.

#! perl -sw
use strict;

our $WORDS ||= 1;
our $KEYWORD || die "-KEYWORD=word needed";

local $/ = ''; # Paragraph mode

my $re_word = qr[\S+\s+];

my $re_5w_key_5w = qr[
    (
        $re_word {1,$WORDS}
        \Q($KEYWORD\E[,.;:!?]*\s+
        $re_word {1,$WORDS}
    )
]ix;

open IN, '<', $ARGV[ 0 ] or die $!;

while( <IN> ){
    if( $_ =~ $KEYWORD ) {
        while( $_ =~ m[$re_5w_key_5w]g ) {
            print "'$1'\n---\n";
        }
    }
}
[download]

Results:

[ 0:58:39.10] P:\test>404751 -WORDS=1 -KEYWORD=poverty "Rhetoric - Ari
+stotle.txt"
'or poverty; it '
---
'or poverty but '
---
'or poverty, of
'
---
'or poverty or '
---
'his poverty, a '
---
'in poverty or '
---
[ 0:58:50.09] P:\test>404751 -WORDS=3 -KEYWORD=poverty "Rhetoric - Ari
+stotle.txt"
'to wealth or poverty; it is
of '
---
'to wealth or poverty but to appetite.
'
---
'of wealth or poverty, of
being lucky '
---
'by sickness or poverty or love or
'
---
'disregard of his poverty, a man aging '
---
'by us in poverty or in
banishment, '
---

[ 0:58:53.10] P:\test>404751 -WORDS=5 -KEYWORD=poverty "Rhetoric - Ari
+stotle.txt"
'action due to wealth or poverty; it is
of course true '
---
'due not to wealth or poverty but to appetite.
Similarly, with '
---
'the sense of wealth or poverty, of
being lucky or unlucky. '
---
'are afflicted by sickness or poverty or love or
thirst or '
---
'man by disregard of his poverty, a man aging war by '
---
'who stand by us in poverty or in
banishment, even if '
---
[download]

use strict;
use warnings;

my $fname = 'test.dat';
my $csize = 102400;
my $word = 'bingo';

my ($handle, $length, $end, $pos, $c);

open($handle, $fname);
while () {
    $length = read($handle, $_, $csize);
    if (index($end.$_, $word) != -1) {
        $pos = tell($handle) - $length + index($end.$_, $word) - 100;
        $pos = 0 if $pos < 0;
        seek($handle, $pos, 0);
        read($handle, $_, 220);
        for (@_ = split(/\W+/)) {
            if ($_ eq $word) {
                print join(' ', @_[(($c > 4) ? $c - 5 : 0)..(($c < $#_
+ - 3) ? $c + 5 : $#_)]);
                last;
            } $c++;
        }
        last;
    }
    last if (!$length);
    ($end) = m/[\W](\w*$)/;
}
close($handle);
[download]

I'd hate to code the search algorithm myself efficiently, so I might shell out to grep -A 1 -B 1 keyword filename, join the lines together, and do the five word windowing on the results.

I'm guessing your kwicionary is rather constant. If so, you can build an index (word => file position) outside of the CGI script, and store it to disk. There must be algorithms out there to build index files which don't need to be completely loaded into memory, but I'm not familiar with them. On second thought, a database would be perfect to store the index.

The CGI script could

1. Locate the position of the word using the index file/table.
2. Open the dictionary file.
3. Seek to the position found in the index file.
4. Read past the word.
5. Read five words. Those are your Next Five Words.
6. Reopen the dictionary file with File::ReadBackwards.
7. Seek to the position found in the index file.
8. Read five words. Those are your Previous Five Words.