impenetrable (e.g. Fodor and Pylyshyn 1988). H ere, the labeling of the con-
stituents is handled by information that can come from the morphophonological
or conceptual components, with Merge subserving only the function of hierar-
chizing that content. This also means that structure building can be affected
by interference phenomena, which I will return to below.
There are at least two different ways of interpreting the labeling process.
Merge can here take just roots and hierarchize them at the syntactic level,
generating an unordered set that is then serialized by a unification operation,
which would add morphosyntactic, morphophonological and phonological infor-
mation in three parallel steps that would constitute the operation of linearization,
giving rise to labeling those categories as a serial structure (Boeckx et al. 2014).
A diff erent interpretation would consider words as chunks in long-term memory
that were originally acquired as roots that were fed into the merging operation,
but that in normal language use are retrieved automatically without the need
to reconstruct them through unification unless they become the object of atten-
tion, in which case they could be flexibly manipulated as any other chunk. The
difference between these two views also entails a bigger difference that has to
do with the question of whether there really is a system of linguistic knowledge
that is separate from a system of linguistic processing, that is, whether there is
one single tree or two independently generated trees during the moment of
processing (Phillips 2013). The value o f an account based on two trees is that
it allows syntax to be separate from the morphosyntactic differences of each
independent language, constituting a universal conceptual structure. An account
based on a single tree, on the other hand, would have to see syntax, not as the
linking component between the three interfaces, but as the builder of conceptual
structure, which can operate with or without language-specific features depend-
ing on the nature of the information that gets integrated at each cognitive cycle.
Regarding the forgetting step, it signals the closing of a cognitive cycle and
of the structure-building operation: the information that is forgotten is actually
transferred to long-term memory, and cannot be modified anymore as part of
that cognitive cycle. Only through an operation of refreshing in subsequent
cognitive cycles can it be retrieved and manipulated. Thus, Merge amounts to
a cyclic rhythm in which the structure-building process is delineated by phases
that always include a head and a complement.
2.2 Working memory and grammar
The time dimension in language is, of course, not foreign to classic generative
linguistics, which recognized that natural languages displayed non-local depen-
dencies that can be captured only by mildly context-sensitive grammars (Chomsky
1963, Joshi 1985). A simple sen tence like “Kate builds it up” (Pulvermüller
2010) shows us how frequent non-local dependencies are in language, with the
subject connecting with the verb’s inflection, and the verb connecting with it
and up. This takes us to a problem that any account of language based on WM
has to be able to answer: how does Merge know where to go? Or in other
Language and working memory 119