Sentence Comprehension (I)
- The structure of a sentence
- Theories of parsing and interpretation: The
garden-path theory and minimal attachment
- Memory and sentence comprehension
- Vocabulary or lexical semantics
- Brain damage and aphasia
1. The structure of a sentence
Comprehension of language is a complex activity that involves
many processes. To start with, we will simplify things by considering
just two types of processes, parsing and interpretation. One thing
we do as we comprehend is to analyze the structure of the language
we hear, a process called parsing. This structure is conveyed
by word order from left to right in English and by the hierarchical
connections between words, to form something vaguely like a family
Figure 1. Tree diagram
The listener then needs to link up this structure (syntax) to
the meaning of the sentence: what is being described or reflected
upon by the speaker. This process of determining the meaning of
spoken language is called interpretation; it is the second process
we will consider, shortly. To give you an idea of the first aspect
of language we will cover, parsing, take a look at the following
(1) It was the dorry that the lutter prated.
Even though this sentence does not make sense, you should still
be able to pick out the words representing the participants, dorry
and lutter, and the word for the action, prated. These clues come
from the appropriate endings attached to the words, and from the
small connecting words, like "the" and "that."
So, it is these word endings (or inflections, such as -ed on prated),
and grammatical words (or functors/function words, such as "it,"
"was," "the" and "that") called,
in psycholinguistics, "the closed class vocabulary,"
which determine the structure of the sentence. The major lexical
items in the above sentence (1), replacing nouns, verb and adjective
(called content words), are nonsensical, made-up words, which
is why you cannot determine what was actually happening in the
[Back to top]
We will consider now how our brains work out the structure of
a sentence. In parsing, the brain has to determine how the words
are grouped into ever higher and sparser layers, like a pyramid.
The groupings adhere to a set of rules, specific for each language,
which most speakers seem to have internalized as young children
from the language around them. People can use these rules without
being able to specify what they are; so we call them tacit knowledge.
To see how you, as a comprehender, are aware of rules without
knowing what they are exactly, take this sentence:
(2) The horse raced past the barn fell.
Up to the second-last word, barn, you perhaps expected it to
continue something like this: "The horse raced past the barn
and into the yard." Probably the last word, "fell,"
came as a surprise, though. Did you go back and try again to work
out a different way of saying the sentence? The way it should
have read was as though there had been commas after The horse
and barn, to mean "the horse that was raced past the barn."
This is a classic sentence devised by a famous psycholinguist,
Bever, in 1970. It illustrates several characteristics of parsing:
Firstly, the reason for your being misled, or garden-pathed, as
it is called, is that the word "raced" is grammatically
ambiguous. That means: it can be used two different ways in English.
One way is to signify that the event (the action entailed in the
main verb) was completed in the past; this is the sense which
your brain first obtained. The other use is to signify a more
succinct version of "which was raced"
use of raced occurs in a reduced relative clause. It is "reduced"
because we have the option to leave out the conjunction "that"
altogether. A relative clause is a group of words with a main
verb which modifies, or says something about, the participant
in an event or state.
[Back to top]
3. Theories of Parsing: The garden-path
theory and minimal attachment
The fact that you arrived at one meaning at first and then had
to revise it, is a phenomenon which has lead to a famous theory
of parsing. It is called the garden-path theory and was proposed
by Lyn Frazier in her doctoral dissertation in 1978. Put simply,
this theory (based on earlier work by Bever, 1970, and another
person called Kimball, 1973), claims that the brain forms the
simplest structure it can at first (respecting the principle of
minimal attachment, such that there are the fewest phrases in
the structure). It is thought that, when a syntactically ambiguous
word signals that the simple structure could not be correct, the
brain backtracks. More consciously now, it makes a second run
through the parser. On this second pass, it reclassifies the words
for grammatical class (main verb, subject noun, etc.) and follows
the grammatical rules of the alternative, more complex, but correct,
structure. The garden-path theory is called a two-stage model
of parsing: stage 1 where the simplest structure is formed; stage
2 where the revised, correct structure is found.
To give you an idea of what an experiment was like
that demonstrated this effect of garden-pathing,
try the word-by-word reading comprehension experiment.
3.1 Characteristics of modular systems in relation to parsing
What characterizes the garden-path model
is that it operates according to Fodor's (1983) idea of a modular
cognitive system (a cognitive system is what we call a mental
system which processes information). A modular system has components
or modules working independently of others (such as the module
for determining the meanings of words), each passes its output
on to the next module in line. So, information passes along, a
bit like on a conveyer belt, with stations or stages that do something
to the information before allowing it to pass on to the next stage.
You can understand what a module is in language processing if
you think of the codes each one would use. The garden-path theory
considers that the parser is handling information coded for the
grammatical class of words. What codes do you think the interpreter
would be using to work out the roles participants have in the
meaning of a sentence? To find out, look at these two sentences:
(3) The boy kissed the girl.
(4) The boy envied the girl.
In the two sentences, "the boy" has two different roles.
In the first, he is an agent of the action of kissing: it is he
who kissed the girl. What is his role in the second sentence?
Is he the agent again, doing some action, or do you think he could
be the experiencer of the emotion of envy? He is not doing anything
to the girl in envying her, so he is not an agent; but he can
certainly experience envy, and so his role is called experiencer.
It is these thematic roles which form the information that the
interpreter component makes use of in a modular language system.
In the garden-path theory, the second process of interpretation
happens after the first run of the parser when an initial expectation
of the structure has been determined. Parsing on this first run
does not consider the roles the participants are taking. (Variations
on this theory allow earlier use of role information but nevertheless
keep the mechanisms handling the different types of information
3.1.2 Information encapsulation
The idea of separate parsing
and interpretation, handled by different mechanisms rather than
a common one, is an example of information encapsulation, one
of the properties of cognitive systems specified by Fodor (1983).
3.2 Constraint-based theories of parsing and interpretation
Another group of theories proposes that comprehension is not
so sequentially organized (i.e., first in-depth parsing then interpretation).
Instead, the two happen in an interactive way from the start.
These constraint-based theories consider that all sources of information
are brought to bear in comprehending a sentence, particularly
the information that is stored in memory concerning each word
in our vocabulary. Some of the types of information stored in
the lexical entry of each word in our long-term memory for words
(semantic memory) include the meaning of the word, its role in
grammar, the structures it can appear in, and the role it plays
in sentence meaning. The constraint-based theories assume that
the type of information used depends on the probability of finding
it in the language and the language-user's experience with it
3.3 Connectionism and parsing-interpretation
theories still keep the idea of separate sources of information,
unlike yet another theory called connectionism, proposed by McClelland
and his collaborators (amongst others). In this third theory,
there is no separation in memory of the representations of the
different types of information, just different connections between
3.4 The concurrent model of parsing-interpretation
there is now a hybrid theory, which amalgamates two of the ones
we have considered. It was proposed by Boland in 1997 and called
a concurrent model because it comprises aspects of the garden-path
and constraint-based theories, operating simultaneously rather
than sequentially. Multiple possible structures, rather than one
(as in the garden-path model), are generated by the listener at
each incoming word. The constraints as to which one wins out come
from lexically stored information, as described in constraint-based
theories (and any other relevant information). Because of this
parallel processing going on, meaning may precede syntax at different
points; but in the end syntax prevents the acceptance of ungrammatical
structures, even if they are plausible.
[Back to top]
4. Memory and sentence comprehension
While what has been said so far seems no doubt complex, it is
not the end of the story! Comprehension is not just parsing and
interpretation. The brain must use memory capacity for the different
types of information in determining the structure and interpretation
of the language to which we attend. This memory system is now
commonly called working memory (a term attributed to Baddeley,
one of the most well-known of memory researchers). Working memory
is presumed by some researchers to be a system that holds information
while it is being processed. Various theories as to the nature
of working memory exist amongst researchers who believe in modularity.
We will consider four basic types here, because of their relevance
to comprehension. Connectionists, on the other hand, see working
memory as just a product of the network in operation, not anything
with a capacity for storage of information.
4.1 Baddeley's model of working memory
Baddeleys model of working memory has special storage
systems for visual and auditory-sound information. The auditory-sound
information memory capacity is called phonological, because it
is thought not to be acoustic (i.e., sound based) or articulatory
(i.e., related to pronunciation) but some more-abstract code,
nonetheless based on the sound patterns of a language. Lots of
research, however, indicates that semantic information is also
coded in short-term memory, and must be held temporarily in memory
while comprehension is going on (e.g., words are recalled better
than made-up nonwords, in the short term). To experience this
difference for yourself, try the
STM recall task comparing word and nonword list recall.
From what you have learned about parsing, it seems logical that
the brain must hold partial syntactic analyses while it is comprehending
sentences in a discourse. So, another group of theories has more
than just the two storage systems that Baddeley and Hitch proposed
4.2 Caplan and Water's model of working memory for language
differ as to how many of these additional storage systems there
are: Caplan and Waters (1999) propose one separate system for
all the processes to do with language (such as parsing and interpretation)
and a third system for handling the output of the language processor
4.3 Carpenter, Just and Miyake's model of working memory for
Yet another idea, from the team of Carpenter, Just
and Miyake, is that there is just one general capacity system
of memory for both the short-term recall of lists and all the
different aspects of language processing. Gibson (1998, 2000)
has a general memory capacity for language processing in his model,
too, but he has been more specific in describing how its capacity
is depleted in terms of parsing and interpretation.
4.4 Martin's model of working memory for language
The fifth and last hypothesis that we will deal with here
is that from Martins lab (e.g., Martin and Romani, 1994).
She believes that there are storage systems in memory for each
and every type of information involved in comprehension. Thus,
there would be separate ones for phonology, for the semantics
of words (lexical semantics), for syntax, for interpretation,
and so forth. Her research with patients shows that there are
indeed separate systems for phonology and semantics, at least.
The evidence comes from the different areas damaged by strokes
and the behavioral differences. Evidence for a separate memory
system for syntax has yet to be demonstrated conclusively. One
team, Miyake, Carpenter and Just (1994), has tried to show that
the complexity of syntactic analyses affects working memory. There
are objections to their work, however, so we need better evidence
of a syntactic storage buffer.
4.5 Working memory as a cognitive system
Consdier how these ideas on memory systems are like the disputes
over modular systems in comprehension: some have a general storage
facility undifferentiated for the type of information; others
have different degrees of specialization in memory storage, depending
upon the type of code or information involved. Such versions exemplify
the modularity of memory systems and their information encapsulation.
[Back to top]
5. Vocabulary or lexical semantics
Probably the first thing that came to mind, when you started
to read this topic of comprehension, was vocabulary. When we talked
earlier about the constraint-based theories of sentence comprehension,
we met this topic of vocabulary, or lexical knowledge. Vocabulary
is portrayed in various ways in psycholinguistics, depending upon
its different uses. When you hear someone speak, the sounds of
the words you hear have to be recognized as those words in your
English vocabulary. This dictionary of heard words from your language
is called a lexiconæan input phonological lexicon to be
exact. Then you have to attach meaning to each word, and, if the
word is part of a sentence, your brain makes use of the other
types of stored information connected with each word mentioned
earlier: the grammatical category of each word; the assignment
of roles to participants in an event; the structures that different
verbs can appear in; the grammatical attachments to words, especially
verbs. Apart from needing to know what a sentence is about, it
is these other types of information which are the focus of this
chapter, and not vocabulary (lexical semantics). These other types
of information figure prominently in the sentence-comprehension
research to do with aphasia, which we come to now.
[Back to top]
6. Brain damage and aphasia
is the name given to a whole host of language impairments that
may follow brain disease and traumatic brain damage. This group
of impairments may arise from: bleeding into the brain ( hemorrhage)
or blockage of the blood supply to the brain (thrombosis and embolism);
bruising or tearing of brain tissue (contusion); pressure on brain
tissue (from a tumor, depressed skull fracture, pooling of tissue
fluid or blood); and atrophy or other generalized changes in brain
tissue (as in the dementias, although aphasia used to be applied
to focal brain damage only). We do not know how many different
types of aphasia there are, because they are still being described,
but the knowledge we have so far is quite extensive. Here, we
will limit ourselves to what is topical in sentence-comprehension
impairments. Aphasia can be studied in its own right, in order
better to understand what has become dysfunctional for the patient
and how best to remediate it. It can also be studied from the
point of view of what light it sheds on intact processes. In our
approach, we will consider just comprehension impairments from
the point of view of what it tells us about how the brain should
work if unimpaired. In considering comprehension deficits, we
will confine ourselves to those that seem to affect parsing and
[Back to top]