342 Dik Bakker and Anna Siewierska
past that has been remembered and that is being told. However, in sponta-
neous conversation completely ‘new’ things may be said that speakers do
remember later on. Also, telling ‘old’ stories may have a restructuring ef-
fect on the way they are remembered and told on a later occasion. Thus, it
seems to be necessary to have feedback from the semantics of an utterance
to semantic memory, too.
In the light of the above we would suggest that the FDG model be re-
fined in such a way that what resides in working memory and what in
permanent storage is determined more precisely. A possible scenario is the
following. Assuming that all basic material – both building blocks and
rules – are located in semantic memory, an utterance is then constructed on
three more or less independent levels: IL, RL and EL. For all three levels
construction starts out from semantic memory, which sends (parts of) the
contributions of pragmatics, semantics and morphosyntax to working
memory at certain intervals. These intervals and the amount sent are related
to the capacity of working memory and to what is already present in it.
This process follows the hierarchy pragmatics > semantics > form. Accord-
ingly, some bits of the pragmatics are first sent to working memory, say
some aspects of the Move and the Act. When this has arrived, semantics
starts filling in the relevant parts, while pragmatics moves further on in the
process, e.g. to the first subordinate Act. When semantics is finished with
the first chunk and is ready to start working on the second chunk it gets
from pragmatics, the expression rules kick in and start working on the first
chunk of RL material. Apart from the fact that a great deal of the material
to be uttered by the speaker is in semantic memory in some form or other
before any utterance is made at all, we assume that it is IL material and
above all RL material that is fed back to semantic memory. There it is used
by an autonomous process that constructs a version of the total discourse,
to the extent that what is being said is a complete paraphrase or a more or
less new version of what was already residing there in some form or other
anyway. Only very little of the EL material is fed back for permanent stor-
age. Only this, and what is left in working memory, is available for
monitoring processes.
As to the shape and size of the contents of working memory at any one
time, we should probably best think of entities such as Chafe’s intonation
unit (Chafe 1987; and see Crystal 1975 for an earlier version of it) or Hal-
liday’s information unit (Halliday 1994).^14 Indeed, in studies of spoken
language it has been frequently observed that (spontaneous) speech gener-
ally proceeds via relatively short contributions, which may contain one or
more given elements and rarely introduce more than one element that is