I have an attached an updated patch with following changes:
1. Respects ShortWord and MinWords
2. Uses hlCover instead of Cover
3. Does not store norm (or lexeme) for headline marking
4. Removes ts_rank.h
5. Earlier it was counting even NONWORDTOKEN in the headline. Now it
only counts the actual words and excludes spaces etc.
I have also changed NumFragments option to MaxFragments as there may not
be enough covers to display NumFragments.
Another change that I was thinking:
Right now if cover size > max_words then I just cut the trailing words.
Instead I was thinking that we should split the cover into more
fragments such that each fragment contains a few query words. Then each
fragment will not contain all query words but will show more occurrences
of query words in the headline. I would like to know what your opinion
on this is.
-Sushant.
On Thu, 2008-06-05 at 20:21 +0400, Teodor Sigaev wrote:
> > A couple of caveats:
> >
> > 1. ts_headline testing was done with current cvs head where as
> > headline_with_fragments was done with postgres 8.3.1.
> > 2. For headline_with_fragments, TSVector for the document was obtained
> > by joining with another table.
> > Are these differences understandable?
>
> That is possible situation because ts_headline has several criterias of 'best'
> covers - length, number of words from query, good words at the begin and at the
> end of headline while your fragment's algorithm takes care only on total number
> of words in all covers. It's not very good, but it's acceptable, I think.
> Headline (and ranking too) hasn't any formal rules to define is it good or bad?
> Just a people's opinions.
>
> Next possible reason: original algorithm had a look on all covers trying to find
> the best one while your algorithm tries to find just the shortest covers to fill
> a headline.
>
> But it's very desirable to use ShortWord - it's not very comfortable for user if
> one option produces unobvious side effect with another one.
> `
>
> > If you think these caveats are the reasons or there is something I am
> > missing, then I can repeat the entire experiments with exactly the same
> > conditions.
>
> Interesting for me test is a comparing hlCover with Cover in your patch, i.e.
> develop a patch which uses hlCover instead of Cover and compare old patch with
> new one.
