RE: proposal: replace 'rank' property with score'

Great idea. In fact, I read "relevance rank" as being a
score, not an ordinal ranking from 1 to N. Renaming
rank to score would prevent such confusion, and
would also be more convenient to implement in some
cases.

However, there are a few issues about scores to
consider.

(1) For some systems, the score of a document
relative to a query depends upon the collection in
which the document exists. If exactly the same
document were in another collection, the exact same
query would give a different score. On the
other hand, for some systems, the score is
dependent only on the document and the query,
not on the collection. Which case pertains matters
for the case where the query is across multiple
document collections: When merging
multiple result streams ordered on score,
the scores across collections need to be
comparable.

STARTS makes a partially successful attempt to
address this issue by having two scores, a raw
score, and a normalized score. The raw score
is good enough for one collection. For multiple
collections, the normalized score is needed.
The normalized score is less sensitive to the
collections, but not completely independent of 
them.

So, do we need more than one score property?
Maybe not. Maybe we can just have one, and assume 
that it is the normalized score or a collection
independent score if more than one collection
is involved. For the single collection case,
it doesn't matter. The datatype (integer or real)
can be obtained from the QSD. I believe DASL is 
only strictly concerned with the single collection
case on release 1.0.

(2) We should still say words to the effect that
"a larger number for score means the document
is judged to be more relevant to the query"
so there is no ambiguity.


Alan Babich

> -----Original Message-----
> From: Jim Davis [mailto:jdavis@parc.xerox.com]
> Sent: July 24, 1998 2:30 PM
> To: www-webdav-dasl@w3.org
> Subject: proposal: replace 'rank' property with score'
> 
> 
> I propose that we replace the dav:rank property with a score.
> 
> As far as I know, full text servers compute a score and use 
> that to rank
> order the results.  It's the score that's the more primitive.  
> 
> While it's true that there's no standard about the form of 
> the score (for
> some it's an integer 0 to 1000 for others a float between 0 
> and 1.0) it's
> still useful, at least to UIs.
> 
> By contrast rank is pretty useless.  You already get the rank 
> just by the
> order of records in the record list anyway.  It it 
> superfluous to sort on
> rank ascending and perverse to sort on it descending.
>

Received on Friday, 24 July 1998 20:05:00 UTC