Sushant Sinha <sushant354@gmail.com> writes:
>> Doesn't this force the headline to be taken from the first N words of
>> the document, independent of where the match was? That seems rather
>> unworkable, or at least unhelpful.
> In headline generation function, we don't have any index or knowledge of
> where the match is. We discover the matches by first tokenizing and then
> comparing the matches with the query tokens. So it is hard to do
> anything better than first N words.
After looking at the code in wparser_def.c a bit more, I wonder whether
this patch is doing what you think it is. Did you do any profiling to
confirm that tokenization is where the cost is? Because it looks to me
like the match searching in hlCover() is at least O(N^2) in the number
of tokens in the document, which means it's probably the dominant cost
for any long document. I suspect that your patch helps not so much
because it saves tokenization costs as because it bounds the amount of
effort spent in hlCover().
I haven't tried to do anything about this, but I wonder whether it
wouldn't be possible to eliminate the quadratic blowup by saving more
state across the repeated calls to hlCover(). At the very least, it
shouldn't be necessary to find the last query-token occurrence in the
document from scratch on each and every call.
Actually, this code seems probably flat-out wrong: won't every
successful call of hlCover() on a given document return exactly the same
q value (end position), namely the last token occurrence in the
document? How is that helpful?
regards, tom lane