Koen - I'm glad you came back on this one.

On Sat, 19 Jul 1997, Koen Holtman wrote:

> Martin J. Duerst:
> >
> [....]
> >Another question that remains to me is that currently, the
> >spec assumes that each document has assigned exactly one
> >language-tag, 
> ???  The spec does not assume this, see the section on the
> Content-Language header.  Is there some text in the spec which
> contradicts the Content-Language section?

"exactly one language tag" was probably wrong. What I mean is
something on a somewhat higher level, which I at the moment
have difficulties to express.

Assume we have a very sophisticated server, which knows a lot
about the documents it keeps or produces, and the languages
they are kept in. Assume some of this knowledge cannot be
expressed in one or more language tags attached to the
documents, including q values.

The question then is whether such a server is allowed to
cheat, i.e. whether it is e.g. allowed to peek at the
Accept-Language header field sent in before it calculates
the language(s) and q values it assigns to a certain

As an example, assume that a server has an English document,
as Japanese document, and an English-Japanese bilingual
document. The bilingual document is the source and contains
all text in the original, and therefore is the preferred
document. However, if a reader has only English in her
Accept-language header field, the English document should
be served, and so on. Currently, such behaviour could only
be modelled by "cheating" as explained above, i.e. by
analysing the Accept-language header field and e.g.

	English document:		en; q=1.0
	Japanese document:		ja; q=1.0
	Bilingual document:		en; q=0.5, ja; q=0.5

in case the Accept-language header field contains
only English or only Japanese, but

	English document:		en; q=0.5
	Japanese document:		ja; q=0.5
	Bilingual document:		en; q=1.0, ja; q=1.0

in case the Accept-language header field contains both
English and Japanese (with a reasonably high q value each).
Note that many other q values in the above example will
have the desired effect, but there is no single combination
of q values that will have the desired effect in both cases.

I'm not proposing here to make the algorithm more complicated,
as this is probably a rare case. But I wonder whether such
"cheating" is allowed or disallowed according to the spec,
or whether it is just an implementation issue.

Regards,	Martin.

Received on Monday, 21 July 1997 06:41:21 UTC