%left and precedence

[Bernd Prager]

Hi,
I want words which are combined with a '-' character
interpreted with a higher priority then uncombined words.
So this is the grammar I tried:
---- snip ----------------

%type <str> object word sequence term
%token <str> WORD
%left '-'

%start object
%%

object: word
       {
          printf( "object: (word >%s<)\n", $1);
          $$ = strdup($1);
       }
    |  sequence
       {

    ;
          printf ("object: (sequence >%s<)\n", $1);
          $$ = strdup($1);
       }
    |  term
       {
          printf ("object: (term >%s<)\n", $1);
          $$ = strdup($1);
       }
    ;

word:  WORD
       {
          printf( "word: >%s<\n", $1);
          $$ = strdup($1);
       }
    ;

sequence:  object '-' object
       {
          printf( "sequence: >%s<->%s<\n", $1, $3);
          $$ = (char *)malloc(strlen($1)+strlen($3)+2);
          sprintf($$, "%s-%s", $1, $3);
       }
    ;

term:   object object
       {
          char buf[1024];
          printf( "term: (word >%s< object >%s<)\n", $1, $2);
          $$ = (char *)malloc(strlen($1)+strlen($2)+2);
          sprintf($$, "%s %s", $1, $2);
       }
    ;

---- snip ----------------
I thought the precedence rule (%left '-') would take care
of this. But when I try my grammar with the teststring "New-York is fun",
I get:
---- snip ----------------
word: >New<
object: (word >New<)
word: >York<
object: (word >York<)
word: >is<
object: (word >is<)
word: >fun<
object: (word >fun<)
term: (word >is< object >fun<)
object: (term >is fun<)
term: (word >York< object >is fun<)
object: (term >York is fun<)
sequence: >New<->York is fun<
object: (sequence >New-York is fun<)
---- snip ----------------
I do not want the sequence >New<->York is fun<.
I want the sequence >New<->York< ...

I do understand that as soon as "York" is pushed from the 
stack, bison thinks it is a part of the term "York is fun" and removes it.

Is there a way to achieve what I want with precedence rules or 
is that one of the look-ahead things bison can't do?

Thanks,
-- Bernd