Recursion and Aphasic Sentence Comprehension

1
Recursion and Aphasic Sentence Comprehension David
Glenn Clark, MD University of Alabama at
Birmingham
Abstract
Methods, contd
Results, contd
Methods, contd
The model presented here is a parser that
relies on a neural network for constructing tree
structures. Use of a neural network makes it
possible to give the parser lesions of variable
severity (graceful degradation). The parser uses
lexicalized tree adjoining grammar1 (TAG) to
compute a structure for an input sentence, which
is then translated compositionally into a
semantic interpretation. The introduction of
noise into the network results in progressive
distortion of elementary trees as they are built
from the bottom up. The ability of the parser to
combine trees through the operations of
Substitute and Adjoin (Fig.1) is impeded due to
this distortion. Since distortion is greatest
in the higher parts of the tree, this is where
failure of combinatory operations is most likely
to occur. Compositional interpretation of the
resulting trees reflects comprehension deficits
in aphasia.
The sentence structures generated underwent
compositional semantic interpretation. BRING THE
NOISE Introduction of random noise into the
neural network computations resulted in
distortion of the numerical code representing the
new constituent. The code for this constituent
then acted as input to the neural network during
projection of the next (higher) syntactic node.
Therefore, the amount of distortion of any given
node increased in proportion to the nodes height
in the elementary tree.
Figure 6. Comprehension at increasing levels of
noise for the object-relative sentence The boy
that the girl hugged is happy.
context-free rules was encoded. These rules are
of the form X ? Y Z (e.g., S ? NP
VP) where Y and Z are syntactic constituents that
combine to form a constituent X. Syntactic node
labels were represented in the network as arrays
of numbers, with grammatical knowledge stored as
a distributed memory in the network that mapped
pairs of constituents to a higher constituent.
The network had 11 input nodes for the left
constituent, 11 input nodes for the right
constituent, and 11 output nodes for the new
constituent. BUILDING ELEMENTARY TREES Prior
to syntactic parsing, elementary trees had to be
constructed for all words in the input string.
The process of constructing each tree meant
assigning a numerical code to each node of the
tree. These codes were derived from the neural
net and were used (rather than text labels) by
the parser for building syntactic structures.
The entry for each word in the lexicon consisted
of a sketch of a tree structure containing slots
for all the arguments of the word. The input
layer of the neural network was first clamped to
the numerical code for the words category and
the code for the words complement. In Fig. 2,
these correspond to V and DP. The parent of
these two nodes was then assigned the numerical
code output by the neural network.
Figure 3a. In a low-noise tree, the neural
network approves substitution. 3b. In a tree
distorted by noise the same operation fails.
Figure 7. Comprehension of the subject-relative
sentence The boy that hugged the girl is happy
at increasing levels of noise. Note incorrect
assignment of the thematic role to the matrix
verb (lower solid line on right graph).
Figure 1. Substitution entails replacing a
frontier node with an initial tree (a,b).
Adjoining entails replacing an internal
non-terminal node with an auxiliary tree (c,d).
Results
Figure 8. Comparison of frequency of thematic
role errors in OR and passive sentences. This
pattern was predicted by the model and is present
in the literature.
The proportion of appropriate logical forms
generated for each sentence type reflected
patterns of aphasic comprehension described in
the literature Active Subject Relative gt
Passive gt Object Relative This pattern was
confirmed with ANOVA, followed by a post-hoc
Tukey test.
Methods
Figure 2. Building the tree fragment V'. Codes
for V and DP are entered into the neural network
(bottom layer of lower panel). The node labeled
V is assigned the code output by the neural net.
This process was repeated for the entire tree
(see Fig. 3). Therefore, each elementary tree
used by the parser consisted only of words
(terminal nodes) and numerical codes
(non-terminal nodes).
LINGUISTIC ASSUMPTIONS Structures for four
sentence types were taken from the linguistics
literature. These four sentence types are
considered to be the core data of agrammatic
comprehension. The target structures were X-bar
theoretic trees with a split IP node. Relative
clauses were analyzed as DP adjuncts containing a
trace that was c-commanded by the moved
antecedent DP. Passive voice sentences were
analyzed as adjectival passives with an
optionally adjoined agentive by-phrase. All
trees generated by the parser were subjected to
compositional semantic interpretation according
to a deterministic algorithm. The output of the
interpretation algorithm was a sentence from
formal logic. If the available structure
assigned the verb only one argument, this
argument assumed the Patient role. THE NEURAL
NETWORK The construction of elementary trees
for use by the parser relied on a backpropagation
neural network in which a grammar composed of
Figure 4. Appropriate assignment of the patient
role to the verb in the active sentence The boy
hugged the girl occurs with high accuracy
(vertical axis) even at high levels of noise
(horizontal axis).
Conclusion
This model accounts for the core data of
agrammatism, as well as for errors in
subject-relative comprehension. It has the
advantage of using distributed representations
for syntactic constituents and a parallel network
for general grammatical knowledge. This approach
may reconcile working-memory models4 with the
Tree Pruning Hypothesis3 and the Structural
Prominence Hypothesis.2
Figure 5. Assignment of the patient role to the
DP in the agentive by-phrase in the passive
sentence The boy is hugged by the girl.
References
THE PARSER Once the elementary trees were
constructed, the parser combined them through the
TAG operations of Substitution and Adjoining
(Fig.1). The neural network was accessed for each
TAG operation, to ensure that the operation was
consistent with the context-free grammar.
1. Abeillé, A., Rambow, O. (2000). Tree
Adjoining Grammars. CSLI Publications Stanford,
CA. 2. Friederici, A.D., Gorrell, P. (1998).
Structural prominence and agrammatic theta-role
assignment a reconsideration of linear
strategies. Brain and Language, 65, 253-275. 3.
Friedmann, N. (2001). Agrammatism and the
psychological reality of the syntactic tree.
Journal of Psycholinguistic Research, 30, 1,
71-90. 4. Haarman, H.J., Just, M.A.,
Carpenter, P.A. (1997). Aphasic sentence
comprehension as a resource deficit a
computational approach. Brain and Language, 59,
76-120. 5. Hickok, G., Avrutin, S. (1995).
Representation, referentiality, and processing in
agrammatic comprehension two case studies.
Brain and Language, 50, 10-26.