Many studies suggest that the parser makes predictions for upcoming linguistic structure (i.e.,

1
Disentangling incremental and top-down parsing
Experimental evidence from VO/OV word order
patterns in code-switching
Many studies suggest that the parser makes
predictions for upcoming linguistic structure
(i.e., look-ahead), but it remains
controversial how much (empty) structure is
projected ahead at which point in a sentence do
specific elements function as triggers for the
projection of rigid syntactic templates? Are
templates projected on the basis of any incoming
first word of a sentence? Or do we only predict
one step ahead, i.e. does the parser generate
structure on a node-by-node basis? This study
presents results from a Sentence Matching
experiment on English (VO) - Dutch (OV)
code-switches to discriminate between these
approaches.
1. top-down / depth-first structure generation
This hypothesis predicts that a fully-fledged
language-specific syntactic frame is built as
soon as the first element of a given sentence
becomes available. In terms of RTs, we would thus
expect (1a / 2b) lt (1b / 2a), because in (1a) and
(2b) the word order in the switched clause is
congruent with the word order proper to the
language in which the subject is expressed.
2. top-down / selective structure generation
Under this hypothesis, we expect a
language-specific syntactic frame to be projected
as soon as certain features, associated with the
linguistic content of a specific node, become
available. Conditions (3 4) test whether these
features are associated with the auxiliary
(INFL). If so, we should find the RT-pattern (3a
/ 4b) lt (3b / 4a), as illustrated in the trees
below.
3. incremental/node-by-node structure
generation A node-by-node approach to structure
building entails that the parser adjusts its
predictions for the nature of an upcoming element
with each incoming word, as illustrated in steps
1-4. Thus, in the present study, the Dutch -
English switches (5a 5b) should yield faster
RTs than the English - Dutch switches (6a 6b)
results The language of the subject did not
affect the preferred word order in the switch
(reverse effect p lt .001 / no
interaction) These results are incompatible with
a top-down/ depth-first approach to sentence
comprehension
results (1) \ There was no significant effect of
the directionality of the switch, i.e. Dutch -
English switches did not yield faster RTs than
English - Dutch switches. These results are
incompatible with a strict node-by-node approach
to sentence comprehension.
1
2
results The language of the auxiliary DID
affect the preferred word order in the switch (p
lt .01 / no interaction)
THE BOY WILL V / O De jongen zal O
/ V
3
THE BOY WILL lezen O De jongen zal
het boek V De jongen zal
READ O
4
De jongen zal het boek READ De jongen zal
READ het boek
These results are compatible with a top-down /
selective approach to sentence comprehension.
results (2) \ There was an interaction between
word order and switching pattern, approaching
significance (p .06) OV was preferred for
Dutch - English switches (5) and VO for English -
Dutch switches (6). This effect pairs with the
results obtained for the (3-4) conditions (see
2). Since there was no effect of the language
of the subject (see 1), it must be an effect of
INFL.
Thus, the combined results support a top-down /
selective approach to structure building.
method sentence matching Participants were
asked to indicate whether two sequentially
appearing sentences formed an identical pair or
not. Pairs of ungrammatical sentences have been
shown to yield slower reaction times (RTs) than
pairs of grammatical sentences. There was a 500ms
interval between each two sentences that formed a
pair. Both sentences remained on the screen for
2400ms.
sentence 1 500ms
sentence2
pilot study monolingual To make sure that
the Sentence Matching method is sensitive to the
types of VO/OV word order violations under
investigation, a monolingual pilot was carried
out with a group of L1 Dutch speakers and a group
of L1 English speakers (i.e. 7a lt 7b and 8b lt 8a).
conclusions This study shows that code-switching
data can provide insight into parsing mechanisms
the results suggest that (empty) syntactic
templates are projected ahead, in a top-down
fashion these projections, however, are
triggered by features associated with the
(phonological) content of specific syntactic
nodes only (here the inflected verb). These
findings are incompatible with narrow
interpretations of both depth-first top-down
approaches and node-by-node incremental accounts
of sentence comprehension.
results English natives did indeed show faster
RTs for VO word order and Dutch natives were
faster on OV word order. This effect was stronger
for the English group (English p lt .0001 /
Dutch p lt .01)
participants Pilot 21 native speakers of
English 22 native speakers of
Dutch Code-switching experiment 16 Dutch -
English balanced bilinguals