Académique Documents
Professionnel Documents
Culture Documents
1. Introduction
At the time when the first Emergence of Language volume was published (MacWhinney,
1999), many researchers in language comprehension argued that comprehension behavior was emergent from processes that weighed probabilistic information from many
sources to arrive at the most likely interpretation of linguistic input (see MacDonald
and Seidenberg, 2006, for review). For example, English speakers have no trouble understanding sentence (1) below, even though most of the words in the sentence bank, to,
cash, and check have multiple meanings and parts of speech, and the sentence even contains two different meanings of to. Given all this ambiguity, why is comprehension so
easy?
(1) I went to the bank to cash a check.
The answer from constraint-based accounts of comprehension is that ambiguity might
be overwhelming in isolation, but in the context of a broader sentence and discourse,
comprehenders can rapidly settle on what is the most likely interpretation of nominally
ambiguous input. They do so in part by favoring interpretations that are more frequent
overall (the monetary sense of bank is more frequent than its other meanings in English
as a whole), but the real power in the system comes from context-dependent processing:
check in the context of bank and cash likely refers to a bank check rather than other meanings. This view has sparked extensive research investigating the nature of the constraint
integration and the time course of weighing probabilistic information during sentence
processing. It also raises two important questions: (1) How do people learn to weigh all
the probabilities so rapidly? And (2) where do these probabilities come from?
Answers to (1) take several different forms but all involve claims in which representations (such as word meanings) vary in speed of access as a function of past use, and that
people implicitly learn the statistics of their environment, including their linguistic environment, from a very young age (e.g., Lany and Saffran, 2011). Note that environment
here may encode simple co-occurrences independent of structure or word order (such
The Handbook of Language Emergence, First Edition. Edited by Brian MacWhinney and William OGrady.
2015 John Wiley & Sons, Inc. Published 2015 by John Wiley & Sons, Inc.
82
Maryellen C. MacDonald
as the common co-occurrence of boy and girl), and these co-occurrences may be helpful
in comprehension, but many constraints are highly sensitive to knowledge about exact
word order or sentence structure, so that, for example, to is interpreted differently in
the context of went to the and to cash. Similarly, its the exact sentence environment, not
simple co-occurrences of cash and check, that guides interpretations in Ill cash a check vs.
Ill check if I have cash (see MacDonald and Seidenberg, 2006; Tanenhaus and Trueswell,
1995, for discussion of frequency- and context-sensitive comprehension mechanisms).
These issues concerning knowledge of sentence structures lead us to question (2):
Why does the language have the statistical properties it has, and not others? More specifically for sentence structures, why are some kinds of sentences much more common than
others? I began to address this question in the first Emergence of Language volume, where
I sketched three puzzles about how some aspects of language seemed to emerge from
other language domains (MacDonald, 1999). Those ideas knocked around in my head
for an embarrassingly long time, gradually acquired a broader empirical base with new
studies that my colleagues and I conducted, and eventually emerged as a more specific
account of interactions between language production, language comprehension, and
language form that I have called the ProductionDistributionComprehension (PDC)
Account (MacDonald, 2013a). The PDC claim is that language has many of the statistics
it has, and therefore constraint-based comprehension processes yield the comprehension
patterns that they do, in large measure because of the way language production works.
At some level this has to be true: language production is a necessary step in the creation
of utterances and therefore of their statistical patterns over time and over producers. But
the PDC is more than the observation that speaking produces language. It is that biases
inherent in the production process actively create important distributional regularities
in languages, which in turn drive constraint-based comprehension processes.
Before we investigate these claims more fully, it is important to define their limits.
In saying that language form and language comprehension processes owe a great deal
to language production, I do not mean that language production processes are the only
source of language form and language comprehension. The language producers aim is
to communicate a message, and so of course the producers utterance must reflect that
message. If production difficulty were the only constraint on utterance form, then every
utterance would be some easily produced grunt. Instead, the claim here is that during the
process of converting an intended message to a linguistic utterance developed to convey
that message, many implicit choices must be made for the form of the utterance, and the
production system gravitates toward those message-appropriate forms that are easier
than other forms. Thus while the message is clearly central in dictating the utterance
(the whole point of the utterance is to convey the message), language production processes themselves also shape the utterance form. Sometimes this approach is contrasted
with communicative efficiency (Jaeger, 2013; Ramscar and Baayen, 2013), but the PDC
isnt anti-efficiency; indeed it makes a prediction about how the balance between ease
of production and good communication may shake out: Because production is harder
than comprehension (Boiteau, Malone, Peters, and Almor, 2014), an efficient system is
one in which tuning utterance forms to aid production fluency is overall more efficient
than one that is tuned to the needs of the comprehender (see MacDonald, 2013a, 2013b,
for discussion).
MacDonald (2013a) summarized evidence for three biases in production, each of
which promotes the use of easier utterance forms over more difficult ones. The basic
83
logical chain is the following: (1) Language production (and all motor action) is controlled by a plan that is at least partially developed before it is executed. In the case of
language production, we will call this plan an utterance plan, independently of whether
the utterance plan is ultimately spoken, signed, or written.1 (2) The utterance plan must
be held in memory until it is ready to execute. (3) Language producers must monitor
the state of their utterance plan and its execution to make sure that each part of the plan
is executed at the right time, and that upcoming parts are ready to be executed. (4) An
utterance plan is essential for fluent production, but these memory and monitoring
burdens can lead to disfluency and other errors. There is therefore pressure to limit
the amount and complexity of advance planning, and so language producers learn to
plan incrementally, meaning that they plan some portion of the utterance and begin
to execute this plan (e.g. to speak) while they are simultaneously planning upcoming
portions. (5) This incremental interleaving of planning and execution is accomplished
more easily with certain utterance forms than others, and MacDonald argued for three
types of biases that tend to result in more easily planned forms in the utterance plan.
These biases reflect the essential memory demands of constructing, maintaining, and
monitoring an utterance plan. The three biases are first generally described below,
followed by examples of how they influence production patterns and comprehension
in English and other languages.
84
Maryellen C. MacDonald
to other memory-based effects that are not tied to noun animacy or other features of
message salience.
85
These three biases are only sketched here, but each of them has quite extensive empirical support within language production research.2 Though this chapter is supposed to
address the emergence of comprehension processes, its worth noting that these biases are
themselves at least in part emergent from other non-linguistic systems. These include
the nature of motor and action planning, where a plan precedes execution and must be
maintained in memory, the nature of recall from long-term memory, which is such that
some content is inherently easier to recall than other knowledge, the nature of increased
speed and fluency of action with practice, and the nature of ordering of actions in action
or motor planning, which is such that attention or queuing mechanisms order easier
actions before more difficult ones. Given all these commonalities between language production and production of other kinds of complex plans and motor sequences, it is not
surprising that non-linguistic action research also yields evidence biases that resemble
Easy First, Plan Reuse, and Reduce Interference (see MacDonald, 2013a for discussion).
Production processes require a winner-take-all system, meaning that producers must
settle on only one alternative form: we must settle on either Give Mona the book, or Give the
book to Mona, and we cannot utter some blend of the two. This winner-take-all characteristic, together with adherence to these three production biases, means that a producers
utterances will tend to favor certain forms that mitigate production difficulty over forms
that are more difficult to plan and execute. Aggregating these effects across many, many
language producers, we can see that the language as a whole will tend to have a higher
proportion of easier forms than of more difficult ones. Again, ease of production is not
the only influence on utterance form, but the argument here is that it has substantial
effects on the distributional regularities in the language. These distributional regularities
in turn are the fodder for constraint-satisfaction processes in language comprehension,
to which we turn next.
86
Maryellen C. MacDonald
and Saffran, 2011) and continuing through adulthood (Amato and MacDonald, 2010;
Wells, Christiansen, Race, Acheson, and MacDonald, 2009), which allow comprehenders
to learn over the input that they have encountered. What the PDC adds to this work is a
link to the origin of the statistics used in constraint-based comprehension, in that important distributions can be traced to producers attempts to reduce production difficulty.
In the following sections, we review two classic examples in sentence comprehension,
making these links between production choices, distributions, learning, and sentence
comprehension more explicit.
Verb modification ambiguity: John said that his cousins left yesterday.
Local modification: John will say that his cousins left yesterday.
Distant modification: John will say that his cousins left tomorrow.
Equivalent message to (2c): Tomorrow John will say that his cousins left.
English comprehenders very strongly prefer to interpret ambiguous sentences like (2a)
to have the local modification interpretation (as in 2b) rather than distant modification (2c). This pattern is often thought to arise directly from innate syntactic parsing
or memory biases to favor local phrasal relationships over long-distance ones, variously formulated as Right Association (Kimball, 1973), Late Closure (Frazier, 1987), and
Recency (Gibson, Pearlmutter, Canseco-Gonzalez, and Hickok, 1996). A key assumption has been that these parsing principles operate on purely syntactic representations
without lexical content (e.g., Frazier, 1987). This approach accorded well with the fact
that, with few exceptions (Altmann, van Nice, Garnham, and Henstra, 1998; Fodor and
Inoue, 1994), the lexical content of sentences like (2) has minimal effect on English speakers strong bias in favor of local modification, making verb modification ambiguities the
best available evidence for lexically independent innate parsing algorithms that operate
over abstract syntactic structures.
As Table 3.1 summarizes, the PDC approach accounts for the local interpretation
biases without innate parsing algorithms. Instead the effects emerge from comprehenders learning over the distributional regularities in the language, which in turn stem
from the biases of producers to favor certain sentence forms that minimize production
difficulty.
In Step 1 in the table, the Easy First production bias discourages production of distant
modification sentences like (2c) because more easily planned alternatives exist. In (2c),
87
Table 3.1. ProductionDistributionComprehension (PDC) account of greater comprehension difficulty for ambiguities resolved with distant modification (2c) than with
local modification (2b). Modified from MacDonald, 2013a
PDC steps
1
Production
Distribution
Comprehension
a relatively long phrase (that his cousins left) precedes a short one (yesterday), but Easy
First promotes a short-before-long phrase order, as in (2d) or John said yesterday that
his cousins left. Step 2 identifies the distributional consequences of speakers avoiding
utterances like (2c): in comprehenders previous experience, ambiguous sentences
like (2a) overwhelmingly are associated with a local modification interpretation like
(2b). Comprehenders learn these statistics and are guided by them in interpretation
of new input (Step 3). They have difficulty comprehending largely unattested forms
like (2c), but they readily comprehend the special type of distant modification sentences that dont violate Easy First and that do exist in the language. These sentences,
such as the examples in (3) (with brackets to indicate the local vs. distant modification), are ones in which the modifier (very slowly) is longer than the embedded
verb phrase (swimming), so that the Easy First short-before-long bias promotes a
verb verb modifier structure independent of whether a local or distant interpretation
is intended.
(3) (a) Local modification
(b) Distant modification
88
Maryellen C. MacDonald
For these special cases in which distant modification is common in past experience,
comprehenders can readily interpret ambiguities with either the local or the distant modification interpretation, as dictated by the lexical and discourse context. Its even possible
to find ambiguous sentences of this type in which comprehenders initially strongly prefer the distant modification interpretation, something that should never happen if there
is an innate parsing bias toward local modification. An example is in (4), which is a
quote from US Supreme Court Chief Justice John Robertss majority opinion (original
available at http://www.supremecourt.gov/, opinion date June 26, 2013). As the bracketing shows, the embedded verb phrase to do so is shorter than the modifier for the first
time here, and so the word order in (4a) is the preferred one, independent of the writers
intended interpretation. Upon reading the ambiguous version in (4a), a tempting interpretation is (4b), in which the court had declined for the first time to do something.
The broader context makes it clear that the correct interpretation is instead the local
modification in (4c), where the doing something is for the first time. I havent made
a study of what lexical statistics, such as past co-occurrences of decline and temporal
expressions like for the first time, might promote the incorrect interpretation here, but
the point is clear that the existence of sentences like this, together with the empirical
data from MacDonald and Thornton (2009), argue against an innate comprehension
bias for local modification. Instead, comprehenders have a learned bias toward what
has happened in the past, and that this prior linguistic experience owes to aspects of
production planning.
(4) (a) Original ambiguous sentence: We decline to do so for the first time here.
(b) Distant modification:
We [decline [to do so] for the first time here].
(c) Local (intended) modification: We [decline [to do so for the first time here]].
This claim for the role of past experience in subsequent comprehension processes is at the
heart of constraint-based accounts of language comprehension, which have been applied
to many other syntactic ambiguities (MacDonald and Seidenberg, 2006; Tanenhaus and
Trueswell, 1995). The added value of the PDC is, first, a greater emphasis on the role of
learning probabilistic constraints (e.g., Amato and MacDonald, 2010; Wells et al., 2009),
and, second, an account of the production basis for many of the language distributions
that people learn and use to guide comprehension. Extending the PDC to other syntactic
ambiguities is ongoing; the approach holds promise because (1) these ambiguities turn
on the relative frequency of alternative uses of language, which can be readily learned
from input (Wells et al., 2009), and (2) certain production choices affect syntactic ambiguity. For example, variation in availability of genitive forms (the professors daughter vs.
the daughter of the professor) in English vs. other European languages affects the distribution of noun modification ambiguities and their interpretation in these languages
(see Mitchell and Brysbaert, 1998, for review and Thornton, MacDonald, and Gil, 1999,
for constraint-based studies of cross-linguistic similarities and differences). Similarly,
producers manage production demands through the use of optional words (e.g. V. Ferreira and Dell, 2000), which have substantial effects on ambiguity, the distribution of
formmeaning pairings, and consequent experience-driven ambiguity resolution processes. Thus the PDC prediction is that all syntactic ambiguities can ultimately be traced
to producers implicit utterance choices (at least some of which are in the service of
reducing utterance planning difficulty), the consequent distributions in the language,
and comprehenders learning over those distributions.
89
90
Maryellen C. MacDonald
In the years since those initial observations, relative clauses have made a mark on virtually every facet of language comprehension research. One reason is Miller and Chomskys (1963) argument for a competenceperformance distinction of infinite capacity for
recursion, but constrained in practice by working memory limitations. Several other factors also promoted the prominence of relative clauses in comprehension work. First,
relative clauses are widely held to be syntactically unambiguous (Babyonyshev and Gibson, 1999), so that comprehension difficulty cant be attributed to ambiguity resolution
processes. Second, subject and object relatives can be made to differ by only the order
of two phrases, as in the order of the senator and attacked in (6ab), so that researchers
can contrast comprehension of sentences for which the lexical content seems perfectly
matched. The vast majority of a very large number of studies in English and many other
languages, across children, adults, and individuals with brain injury, disease, or developmental atypicality, show that object relatives are more difficult than their matched
subject relatives (see OGrady, 2011, for review). The logic here seems perfectly clear:
Because the difference in difficulty cant be ascribed to lexical factors or ambiguity resolution, it must reflect purely syntactic operations and the memory capacity required to
complete them (Grodner and Gibson, 2005).
(6) (a) Object relative: The reporter [that the senator attacked] admitted the error.
(b) Subject relative: The reporter [that attacked the senator] admitted the error.
This competenceperformance account of working memory overflow in relative clause
comprehension continues as the dominant perspective in linguistics, language acquisition, adult psycholinguistics, and communicative disorders, despite criticisms of each
of the components of this argument. These criticisms include evidence that multiply
center-embedded sentences need not be incomprehensible (Hudson, 1996), that comprehension difficulty is strongly influenced by the words in the sentence and therefore
cannot reflect purely syntactic processes (Reali and Christiansen, 2007; Traxler, Morris
and Seely, 2002), that object relatives do contain a non-trivial amount of ambiguity
directly related to comprehension difficulty, again refuting the assumption that relative
clauses provide a pure measure of syntactic difficulty (Gennari and MacDonald, 2008;
Hsiao and MacDonald, 2013), that the degree of prior experience with object relatives
predicts comprehension success in children and adults, a result not captured by memory
overload approaches (Roth, 1984; Wells et al., 2009), that peoples comprehension capacity for recursive structures is more accurately described by a system in which working
memory is inseparable from linguistic knowledge than by one with separate competence and performance (Christiansen and Chater, 2001), and that, cross-linguistically,
relative clause complexity does not always predict comprehension difficulty (Carreiras,
Duabeitia, Vergara, de la Cruz-Pava, and Laka, 2010; Lin, 2008). The resilience of memory overflow accounts in the face of these myriad challenges in part reflects the essential
usefulness of the constructs of working memory capacity and competenceperformance
distinctions in cognitive science. However, a second factor is that there has been no
really compelling alternative account that captures both the subject-object relative
asymmetry as well as these other phenomena. The PDC approach aims to provide
exactly this.
91
92
Maryellen C. MacDonald
100
Describing animates
Describing inanimates
80
60
40
20
0
English
Spanish
Serbian
Japanese
Language
Korean
Mandarin
Figure 3.1. The frequency with which object relative clauses are produced to describe animate
and inanimate entities in a picture description task, calculated as a percentage of all relative
clauses produced. The English, Spanish, and Serbian data are from Experiments 1a, 2, and 3
respectively of Gennari et al. (2012). The Japanese data are from Montag and MacDonald (2009),
Korean from Montag et al. (in preparation), and Mandarin from Hsiao and MacDonald (in
preparation)
our pictures. These percentages are calculated over all relative clause responses, so the
percentage of passive relatives (as in 7b) is the inverse of the object relatives shown in
the graph.
Figure 3.1 shows that in six diverse languages, when people are describing something inanimate (e.g., toy), they readily produce object relatives like (7a), but they almost
never do so in describing something animate (man). Instead, they utter passive relatives
like (7b). This result holds across head direction: English, Spanish, and Serbian have a
head-first relative clause structure in which the noun being described precedes the relative clause, as in the toy [that the girls hugging], while Japanese, Korean, and Mandarin
have a head-final [relative clause] head-noun structure. The effect also holds over wide
variation in case marking: Serbian, Japanese, and Korean have extensive case marking
on the nouns, while the other three languages have little or none. Perhaps most interesting for our purposes here, active object relatives and passive relatives have identical
word order in Japanese, Korean, and Mandarin (the passive is indicated by a passive
morpheme on the verb or at the start of a relative clause), and its clear from Figure 3.1
that the animacy effects hold in these cases as well as they do in the three Indo-European
languages, for which word order does differ in the two relative clause types.
These results raise two important questions within the PDC: why does animacy have
these strong effects, and what are the consequences for comprehension? Both of these
questions are addressed in Table 3.2. Step 1 in this table describes how producers use
of object relatives vs. passive relatives is shaped by the joint action of Easy First, Plan
Reuse, and Reduce Interference biases in production planning. On this view, animate
nouns are more likely to be in passive constructions not simply because they are more
conceptually salient and more quickly retrieved from memory (as in Easy First) but
93
Table 3.2. PDC account of greater comprehension difficulty for object than subject relative clauses (citations refer to English results). Modified from MacDonald 2013a
1
Object relatives (7a) are common when the noun being described is inanimate
(toy) but are avoided when the relative clause describes something animate (boy),
for which passive relatives (7b) are produced instead (Gennari et al., 2012;
Montag and MacDonald, 2009). These patterns are owed to at least three
production biases:
a. Easy First: animate nouns are conceptually prominent and easily retrieved from
memory, leading to their position in early or prominent sentence positions. The
passive relative (7b) allows the described noun to be in the prominent subject
position of the relative clause.
b. Plan Reuse: the rate of passive relatives varies with the viability of passives in
the language more generally, reflecting the reuse of passive forms from other
sentence types (Montag and MacDonald, 2009).
c. Reduce Interference: there is more interference between conceptually similar entities (e.g. two animate nouns as in man/girl in the hugging picture in
Figure 3.1) than there is when an animate entity (girl) acts on an inanimate one
(toy). This interference can be reduced by omitting the agent in the utterance
plan, which is possible in passive forms (7b), but not in object relatives (7a). The
higher the conceptual similarity between sentence participants in the event to be
described, the more speakers produce passive agent-omission relative clauses
(Gennari et al., 2012).
2
3
People readily learn these correlations between animacy and relative clause type
(Wells et al., 2009).
Comprehenders who encounter the start of a relative clause have very different
expectations of how it will end, depending on whether something animate or
inanimate is being described, with consequences for comprehension:
a. When relative clauses describe something inanimate like toy, English speakers
rapidly anticipate an object relative (7a); for animates (boy), object relatives are
vanishingly rare and are not expected by comprehenders (Gennari and MacDonald, 2008).
b. The less producers are willing to say an object relative to convey a particular
message, the less comprehenders expect one, and the more difficult the comprehension is when a sentence in fact turns out to contain an object relative
clause (Gennari and MacDonald, 2009).
also because the passive construction allows producers to omit mention of the agent
of the action, in that they can omit the by-phrase (e.g., by the girl) in the passive (7b),
but this omission is not an option in object relatives like (7a). Of course speakers can
choose agentless constructions for rhetorical reasons, perhaps the best-known of which
is exemplified by the responsibility-ducking Mistakes were made. My colleagues and
I have argued that in addition to these situations in which agentless structures are chosen
94
Maryellen C. MacDonald
to convey a particular message, agent omission also has a difficulty-reduction function within language production, specifically allowing the producer to reduce the memory interference that arises when two semantically similar entities, such as man/girl, are
part of the message to be conveyed. Gennari, Mirkovic, and MacDonald (2012) found
these agent-omission effects in English and Spanish (where both the passive and a second structure permit agent omission). They showed that these agent-omission effects
stemmed not from animacy per se but from the semantic overlap of entities that arises
when both an agent and a patient are animate. In both languages, they manipulated the
similarity of animate entities interacting with one another in pictures and found that the
more similar the agent and a patient are, the more likely speakers were to omit the agent
in their utterance.
These results suggest that utterance planning difficulty affects utterance form. That
is, the passive bias for animate-headed relative clauses is not simply an effect of how the
producer frames the message to be conveyed (though that also has an effect on utterance form). Instead, the passive usage shown in Figure 3.1 also reflects the operations
of the language production system, so that ease of retrieval from long-term memory,
and the maintenance processes within working memory, shape the utterance choices
that speakers make. These results may help to tease apart alternative hypotheses concerning more egocentric production (in which production choices aid the producer) and
audience design, in which utterance forms are chosen to aid the comprehender. Communication clearly requires elements of both (see Jaeger, 2013; MacDonald, 2013a, 2013b),
but the agent-omission data appear to be evidence for a production-based motivation for
utterance form. That is, semantic similarity is known to impair recall in memory tasks
and to increase speech errors in production studies (see MacDonald, 2013a), suggesting that there are real production costs to planning an utterance containing semantically
similar items. Our data suggest that producers mitigate this cost by choosing a structure
in which they can omit one of the semantically overlapping entities. Although more
research is needed, its less obvious how these choices could help the comprehender;
semantic relationships in a sentence are often thought to have a facilitative effect, as in
priming, and so eliminating these semantic associations wouldnt seem beneficial.
We are just beginning to understand the factors behind the patterns in Figure 3.1,4
but it is clear that speakers very different choices for animate-describing and
inanimate-describing relative clauses have robust effects on the distributional regularities in these languages. Steps 23 of Table 3.2 show the cascade of consequences of
these choices. As described in Step 2, comprehenders who are exposed to the distributional regularities in their linguistic input implicitly learn the co-occurrences between
discourse environments, words, and sentence structure, so that, for example, they come
to expect object relative clauses modifying inanimate entities like toy but they do not
expect this structure modifying animate entities like man. Step 3 in the table reviews
how comprehenders rapidly bring this information to bear in comprehension, so that
people expect object relatives where theyre commonly produced but are surprised
by them in unexpected environments, leading to comprehension difficulty. The vast
majority of studies demonstrating the difficulty of object relatives have used materials
in which something animate is being described the very situation that producers
avoid and that comprehenders have learned not to expect. Gennari and MacDonald
(2008) showed that when readers encounter text that might be an animate-headed object
relative clause, such as The reporter that the senator , they expect the text to continue
95
with some other construction, such as The reporter that the senator had said was an idiot
didnt show up for work, and they have difficulty comprehending these sentences when
they turn out in fact to be object relative clauses, as in The reporter that the senator attacked
admitted the error. Gennari and MacDonald (2009) further tied these results to production
patterns: comprehenders dont expect animate-headed relative clauses to turn out to
be object relatives precisely because theyre almost never produced. Together, these
results suggest that object relative clause comprehension is simply another example
of ambiguity resolution comprehenders rely on past experience with relative clauses
to guide their interpretation of new ones, and it is this reliance that leads to incorrect
expectations for the unusual sentences that populate psycholinguists experiments.
The results do not reflect any pure effect of syntactic complexity on comprehension
(Gennari and MacDonald, 2008).5
On this view, relative clauses, which have been central to current conceptions of
memory and language use in virtually every subfield of psycholinguistics, turn out to
be wholly unsuited for that role, as they are not unambiguous, and their comprehension
reflects detailed knowledge of correlations between words and structures, not abstract
syntactic representations and putative burdens of holding abstract structures in memory. What then becomes of working memory limitations as a source of comprehension
difficulty, particularly within Miller and Chomskys (1963) competenceperformance
claims for infinite recursion limited by working memory? The short answer is that
researchers may further debate competenceperformance distinctions, but relative
clauses should no longer be offered as evidence of overflow of syntactic memory
representations that limit infinite recursive capacity. A more precise answer about
implications of the relative clause work requires closer attention to what working
memory is and isnt. In saying that the PDC account refutes claims for working
memory limitations in sentence comprehension, my colleagues and I do not mean
that working memory doesnt exist: to the contrary, a prime reason why language
users track the statistics of the language and use them to anticipate upcoming input
is precisely because language comprehension requires significant memory capacity,
and generating expectations for likely outcomes reduces these burdens. However, we
do reject the notion that peoples working memory capacity can be described as a
performance limitation independent of their linguistic knowledge/competence (Acheson
and MacDonald, 2009; MacDonald and Christiansen, 2002; Wells et al., 2009). Our
position reflects broader trends linking working memory and long-term knowledge
(Cowan, 2005), emergent from the temporary maintenance needs of other cognitive
processes (Postle, 2006). Specifically for relative clauses, comprehension capacity varies
with long-term knowledge of these structures, derived from experience. Language
producers provide some kinds of experiences (some kinds of relative clauses) more than
others, with consequences for language distributions, learning over those distributions,
and for the memory demands needed to comprehend these structures: the memory
capacity and experience cannot be separated. Of course computational limitations,
including memory limitations, are also at the heart of the PDC argument for why
producers prefer some utterance forms over others, but this does not mean that the
competenceperformance distinction can simply be shifted to production, because,
again, linguistic working memory, specifically the capacity to produce certain utterance
forms, is not separate from long-term linguistic knowledge or experience (Acheson and
MacDonald, 2009).
96
Maryellen C. MacDonald
NOTES
1 The modality of the utterance (spoken, signed, texted, etc.) does have effects on the utterance plan
but we will ignore modality-specific influences here.
2 Examples of research on Easy First/Accessibility are: Bock, 1987; F. Ferreira, 1991; McDonald,
Bock and Kelly, 1993; Tanaka, Branigan, McLean, and Pickering, 2011. For discussion of Plan
Reuse/Syntactic Priming, see Bock, 1986, and Pickering and V. Ferreira, 2008. For discussions of
Reduce Interference see: Fukumura, van Gompel, Harley, and Pickering, 2011; Gennari, Mirkovic,
and MacDonald, 2012; and Smith and Wheeldon, 2004.
97
A few of these studies had written questions and written replies by participants rather than spoken
questions and responses. There have been no substantial differences in the proportions of structures
uttered in written and spoken versions of our studies to date.
One difference not discussed here is variation in the overall rate of object relatives in Figure 3.1, for
example why Serbian speakers prefer object relatives to passives to a much greater extent than do
speakers of the other languages in the figure. Although a definitive answer to this question awaits
additional research, Gennari et al. (2012) pointed to two properties of Serbian that may affect the
rate of passive usage. First, Serbian has obligatory case marking on the relative pronoun, which is
placed before the rest of the relative clause. The need to utter the case-marked form forces speakers to
commit to a relative clause form before beginning to utter it, and this early commitment may affect
passive usage. Second, Serbian has a greater freedom of word order within relative clauses than
the other languages in Figure 3.1. Gennari et al. observed some Serbian word order variations as a
function of noun animacy, and it may be that where speakers of other languages alternate between
structures, Serbian speakers alternate between word orders as a function of animacy or other factors.
Again, exactly why these patterns obtain is not yet clear.
A fuller treatment than is presented here would include the fact that object relatives with pronoun
embedded subjects (The boy/toy she splashed ) have different production biases, different rates of
production, and different comprehension patterns than the examples discussed here. We must also
consider whether Easy First, Plan Reuse, and Reduce Interference provide an adequate account of
why multiply embedded object relatives, like Miller and Chomskys (1963) The rat [that the cat [that the
dog chased] ate] died, are essentially never produced, and the extent to which comprehension difficulty
here can also be traced to ambiguity resolution gone awry rather than to hard limits on working
memory capacity.
REFERENCES
Acheson, D. J. and M. C. MacDonald. 2009.
Verbal working memory and language
production: Common approaches to the serial
ordering of verbal information. Psychological
Bulletin 135(1): 5068. doi:10.1037/a0014411.
Altmann, G. T. M., K. Y. van Nice, A. Garnham,
and J.-A. Henstra. 1998. Late closure in
context. Journal of Memory and Language 38(4):
459484. doi:10.1006/jmla.1997.2562.
Amato, M. S. and M. C. MacDonald. 2010.
Sentence processing in an artificial language:
Learning and using combinatorial constraints.
Cognition 116(1): 143148. doi:10.1016
/j.cognition.2010.04.001.
Babyonyshev, M. and E. Gibson. 1999. The
complexity of nested structures in Japanese.
Language 75(3): 423450.
Bock, J. K. 1986. Syntactic persistence in
language production. Cognitive Psychology
18(3): 355387. doi:10.1016
/0010-0285(86)90004-6.
Bock, K. 1987. An effect of the accessibility of
word forms on sentence structures. Journal of
Memory and Language 26(2): 119137.
doi:10.1016/0749-596X(87)90120-3.
Boiteau, T.W., P.K. Malone, S. A. Peters, and
A. Almor. 2014. Interference between
98
Maryellen C. MacDonald
99