Chunking Ability Shapes Sentence Processing at Multiple Levels of Abstraction
Department of Psychological Sciences, University of Liverpool
Morten H. Christiansen (
[email protected])
Department of Psychology, Cornell University
Abstract As an intuitive demonstration of the necessity of
Several recent empirical findings have reinforced the notion
chunking, imagine being tasked with recalling a string of
that a basic learning and memory skill—chunking—plays a letters, presented auditorily: u o p f m r e e p o a e c s g n p l
fundamental role in language processing. Here, we provide i r. After a single presentation of the string, very few
evidence that chunking shapes sentence processing at multiple listeners would be able to recall a sequence consisting of
levels of linguistic abstraction, consistent with a recent even half of the letters (cf. Cowan, 2001). However, if
theoretical proposal by Christiansen and Chater (2016). exposed to the exact same set of letters but re-ordered
Individual differences in chunking ability at two different slightly, virtually any listener would able to recall the entire
levels is shown to predict on-line sentence processing in
separate ways: i) phonological chunking ability, as assessed sequence with ease: f r o g m o u s e p a p e r p e n c i l.
by a variation on the non-word repetition task, predicts Clearly, such a feat is possible by virtue of the ability to
processing of complex sentences featuring phonological rapidly chunk the sequence into familiar sub-sequences
overlap; ii) multiword chunking ability, as assessed by a (frog, mouse, paper, pencil).
variation on the serial recall task, is shown to predict reading According to the proposal of Christiansen and Chater
times for sentences featuring long-distance number agreement (2016), the Now-or-Never Bottleneck requires language
with locally distracting number-marked nouns. Together, our
users to perform similar chunking operations on speech and
findings suggest that individual differences in chunking
ability shape language processing at multiple levels of text in order to process and learn from the input. This is
abstraction, consistent with the notion of language acquisition necessary both due to the fleeting nature of sensory memory
as learning to process. and the speed at which information is encountered during
processing. Specifically, language users must perform
Keywords: sentence processing; chunking; learning;
memory; usage-based approach; language Chunk-and-Pass processing, whereby input is chunked as
rapidly as possible and passed to a higher, more abstract
Introduction level of representation. Information at higher levels must
also be chunked before being passed to still higher,
Language takes place in real time; a fairly uncontroversial increasingly abstract levels of representation.
observation, yet one with far-reaching consequences that are Thus, in order to communicate in real-time, language
rarely considered. For instance, a typical English speaker users must chunk at multiple levels of abstraction, ranging
produces between 10 and 15 phonemes per second from the level of the acoustic signal to the level of
(Studdert-Kennedy, 1986), yet the ability of the auditory phonemes or syllables, to words, to multiword units, and
system to process discrete sounds is limited to around 10 per beyond. Indeed, mounting empirical evidence supports the
second, beyond which the signal is perceived as a single notion of chunking at levels higher than that of the
buzz (Miller & Taylor, 1948). Moreover, the auditory trace individual word: children and adults appear to store and
is limited to about 100ms (Remez et al., 2010). utilize chunks consisting of multiple words in
Compounding matters even further, human memory for comprehension and production (e.g., Arnon & Snider, 2010;
sequences is limited to between 4 and 7 items (e.g., Cowan, Bannard & Matthews, 2008). Moreover, usage-based (e.g.,
2001; Miller, 1956). Simply put, the sensory signal is so Tomasello, 2003) and generative (e.g., Culicover &
incredibly short-lived, and our memory for it so very Jackendoff, 2005) theoretical approaches have highlighted
limited, that language would seem to stretch the human the importance of such units in grammatical development
capacity for information processing beyond its breaking and sentence processing alike.
point. We refer to this as the Now-or-Never bottleneck Chunking has been considered a key learning and
(Christiansen & Chater, 2016). memory mechanism in mainstream psychology for over half
How is language learning and processing possible in the a century (e.g., Miller, 1956), and has been used to
face of this real-time constraint? A key piece of the puzzle, understand specific aspects of language acquisition (e.g.,
we suggest, lies in chunking: through experience with Jones, 2012; Jones, Gobet, Freudenthal, & Pine, 2014).
language, we learn to rapidly recode incoming information Nevertheless, few have sought to understand how it may
into chunks which can then be passed to higher levels of shape more complex linguistic skills, such as sentence
representation. processing. McCauley and Christiansen (2015) took an
initial step in this direction, showing that individual
2681
differences in low-level chunking abilities were predictive controlling for such factors as attention, motivation, and—to
of reading times for sentences involving relative clauses, the extent that it is separable—working memory.
demonstrating the far-reaching impact of basic chunking
skills in shaping complex linguistic behaviors. Method
The present study seeks to evaluate the predictions of the Participants 42 native English speakers from the Cornell
Chunk-and-Pass framework more closely, by examining undergraduate population (17 females; age: M=19.8,
individual variation in chunking at two different levels of SD=1.2) participated for course credit. Of the original 45
abstraction. Specifically, whereas chunking has previously subjects, one was excluded due to audio recording errors,
been treated as a uniform memory ability, we test the novel while two subjects failed to complete all three tasks.
theoretical prediction that chunking abilities may be Materials Experimental stimuli consisted of word trigrams
relatively independent at different levels of linguistic spanning a range of frequencies, extracted from the
abstraction. Participants were first asked to take part in a American National Corpus (Reppen, Ide & Suderman,
multiword-based serial recall task (Part 1) designed to yield 2005) and the Fisher corpus (Cieri, Graff, Kimball, Miller &
a measure of chunking at the word level. This was followed Walker, 2004). The combined corpus contained a total of 39
by a variation on the non-word repetition task (Part 2), million words of American English. Each item was
designed to yield a measure of phonological chunking compositional (non-idiomatic). Item frequencies, per million
ability. Importantly, due to the memory limitations words, ranged from 40 to .08, averaging at .73.
discussed above, participants must utilize chunking in order Each word was synthesized independently using the
to recall more than a few discrete words or phonemes in Festival speech synthesizer (Black, Clark, Richmond, King
these tasks (e.g., Cowan, 2001; Miller, 1956). Finally, & Zen, 2004) and concatenated into larger strings consisting
participants took part in an online self-paced reading task of 12 words (4 trigrams). Each trigram was matched as
(Part 3). The results show that chunking ability at each level closely as possible for frequency with the others occurring
predicts different aspects of sentence processing ability: in a sequence.
chunking at the phonological level predicts the extent to To provide a non-chunk-based control condition, each
which low-level phonological information interferes with or item was matched to a sequence of words which contained
facilitates complex sentence processing, while chunking at identical functors but random frequency-matched content
the multiword level predicts the role of local information in words (in order to avoid semantic overlap effects on recall,
processing sentences with long-distance dependencies. content words were not re-used). The ordering of the words
was then randomized. An example of a matched set of
Part 1: Measuring Individual Differences in sequences is shown below:
Word Chunking Ability
The first task sought to gain a measure of individual 1) have to eat good to know don’t like them is really nice
participants’ ability to chunk words into multiword units. To 2) years got don’t to game have she mean to them far is
this end, we specifically isolate chunking as a mechanism
by employing a classic psychological paradigm: the serial The final item set consisted of 20 sequences (10
recall task. Serial recall has a long history of use in studies experimental, 10 control).
of chunking, dating back to some of the earliest relevant Procedure Each trial featured a 12-word sequence
work (e.g., Miller, 1956), as well being used to extensively presented auditorily. Each word was followed by a 250ms
study individuals’ chunking abilities (e.g., Ericsson, Chase, pause. Immediately upon completion of the string, the
& Faloon, 1980). participant was prompted to verbally recall as much of the
Participants were tasked with recalling strings of 12 sequence as possible. Responses were recorded digitally and
individual words, with each string consisting of 4 separate later transcribed by a researcher blind to the conditions as
word trigrams extracted from a large corpus of English. well as the purpose of the study.
Importantly, in order to recall more than a few discrete The presentation order of the sequences was fully
items (as few as 4 in some accounts; e.g., Cowan, 2001), randomized. The entire task took approximately 15 minutes.
listeners must chunk the words of the input sequence into
larger, multiword units. In this case, we expect them to draw Results and Discussion
upon linguistic experience with the trigrams in the Participants recalled significantly more words from
experimental items. experimental strings than the frequency-matched control
In addition, we included a baseline performance measure: sequences. The overall recall rate for words occurring in
matched control strings, which featured identical functors to experimental items was 74.0% (SE=2.3%), while the recall
the experimental sequences, along with frequency-matched rate for control sequences was just 39.2% (SE=1.1%). The
content words (to avoid semantic overlap effects on recall), difference between conditions was significant (t(41)=18.8,
presented in random order. Thus, comparing recall for p<0.0001).
experimental and control trials provides a measure of word As the purpose of Part 1 was to gain an overall measure of
chunking ability that reflects language experience while chunk sensitivity, we calculated the difference between
conditions individually for each subject (M=34.8%,
2682
SE=1.8%), which afforded a measure of word-chunking For each of three different syllable lengths (4-, 5-, and 6-
ability that reflects language experience while controlling syllables), the algorithm extracted item pairs that differed
for factors such as working memory, attention, and maximally in sequence likelihood (based on phoneme
motivation. We refer to this difference measure as the Word trigram statistics) across two different sequential orderings
Chunk Sensitivity score, and it is used as a predictor of of the same set of syllables. In other words, pairs were
sentence processing ability in Part 3. selected in which one ordering of syllables was highly
In addition to bolstering previous empirical support for “chunk-like,” while the other ordering of the same syllables
compositional (non-idiomatic) multiword sequences as was less “chunk-like,” according to the phoneme statistics
linguistic units in their own right (e.g., Bannard & of the corpus. Four sets of non-words (the four in which the
Matthews, 2008), Part 1 revealed considerable individual pair differed most greatly in terms of sequence likelihood)
differences across participants in word chunking ability. were selected for each syllable length. An example of a
Recall rates for experimental items ranged from as high as highly “chunk-like” 4-syllable item is krew-ih-tie-zuh,
93.3% to just 30.4%, with difference scores across the which was matched to the less chunk-like tie-zuh-ih-krew.
conditions ranging from 50.8% as low as 3.0%. Thus, the final set of items included 24 non-words, eight
in each of three syllable-length conditions (4-, 5-, and 6-
Part 2: Measuring Individual Differences in syllable), with four being highly “chunk-like” and the other
Phonological Chunking Ability four consisting of alternate orderings of the same syllables
which were statistically less “chunk-like.”
While the first task sought to gain a measure of individual Procedure The task was split into two blocks, with all
participants’ chunking abilities at the level of words, Part 2 NWR item pairs counterbalanced between them. The
sought to gain a measure of chunking ability at the auditory presentation of each non-word was followed by a
phonological level. To this end, we re-purposed the standard 1500ms pause, after which the participant was prompted to
non-word repetition (NWR) task as a chunking task. NWR recall the item verbally. As with Part 1, responses were
has been used extensively to study various aspects of recorded digitally and scored offline. The task took
language development. Recent studies, however, have approximately 4 minutes to complete.
suggested that chunking may better account for NWR Correct responses received a score of 1. Responses
performance than more nebulous psychological constructs, involving alteration to a single phoneme (usually a vowel
such as working-memory (e.g., Jones, 2012; Jones et al., substitution, which could easily stem from differences in
2014). In one sense, the NWR task can be re-conceptualized regional dialect) received a score of 0.5. All other responses
as a serial recall task, as in Part 1. Following such work, and received scores of 0.
in keeping with the Now-or-Never perspective outlined
above, we propose that individual differences in chunking Results and Discussion
ability underlie differences in NWR performance. In turn,
NWR—with appropriately constructed stimuli—can serve Participants achieved a mean NWR accuracy rate of 54.1%
as an additional dimension along which to measure (SE=2.3%). While the overall differences between the high
chunking ability at the level of phonological processing. chunk-like (M=55.2%, SE=2.5%) and low chunk-like
Participants engaged in a standard NWR task, with each (M=53.1%, SE=2.5%) conditions were in the expected
non-word consisting of 4, 5, or 6 syllables. However, the direction, they were subtle, with a mean difference of 2.1%
stimuli were designed such that the same set of syllables (non-significant: t(41)=1.12, p>0.1). However, there was
occurred in two different non-words, but in different considerable individual variation in the size of this
orderings: one ordering yielded an item with high difference across participants (SE=1.9%), ranging from
“chunkability,” according to corpus statistics, while the 29.2% to less than 0%, at -16.6%. Therefore, in Part 3, we
other was estimated to be less “chunkable.” The two items assess both the overall NWR performance score as well as
were then counterbalanced across halves of the task. the difference between the conditions (which we refer to as
the Phonological Chunk Sensitivity score) as predictors of
Method sentence processing.
Importantly, neither the overall raw task performance
Participants The same 42 subjects from Part 1 participated (β=-0.03, p=0.9) nor the Chunk Sensitivity scores (β=-0.19,
directly afterwards in this task. p=0.22) from Parts 1 and 2 correlated with one another,
Materials Non-words were generated using an algorithm consistent with the notion that chunking at each level may
which took a large list1 of English syllables and randomly have different consequences for sentence processing.
generated syllable combinations that were evaluated
according to distributional statistics at the phoneme level. Part 3: Measuring Individual Differences in
For the purpose of supplying statistics, the combined corpus
used in Part 1 was automatically re-transcribed phonetically Sentence Processing and Chunking
using the Festival speech synthesizer (Black et al., 2004). In Part 1, we sought to gain a measure of individual
participants’ ability to chunk words together, while Part 2
aimed to provide a measure of phonological chunking
1 http://semarch.linguistics.fas.nyu.edu/barker/Syllables/ ability. In Part 3, the same subjects from the first two parts
2683
participated in a self-paced reading task designed to: i) Thus, participants’ resilience to phonological interference
assess on-line sentence processing across two different was hypothesized to be better predicted by Phonological
sentence types which were hypothesized to involve Chunk Sensitivity (Part 2), while participants’ susceptibility
chunking at the word and phonological levels, but to to local number mismatch was expected to be better
different extents; ii) determine the extent to which chunking predicted by Word Chunk Sensitivity (Part 1).
ability, as assessed in the first two tasks, predicted
processing difficulties for each sentence type. Method
The first sentence type featured long distance subject-verb Participants The same 42 subjects from Parts 1 and 2
number agreement with locally distracting number-marked participated in Part 3 immediately afterwards.
nouns, exemplified by (1): Materials There were two sentence lists—counterbalanced
across subjects—each consisting of 9 practice items, 20
1. The key to the cabinets was rusty from many years of experimental items, 20 matched control items, and 68 filler
disuse. items. There were two experimental conditions, each with
20 items; the first consisted of the OR sentences featuring
Previous work (Pearlmutter, Garnsey, & Bock, 1999) has phonological overlap (the first 20 items from Acheson &
shown that readers are slower to process the verb when the MacDonald, 2011). The second experimental condition
number of the local noun (cabinets) does not match that of consisted of grammatical sentences featuring long-distance
the head noun (key), resulting in the sequence (cabinets number agreement with locally distracting number-marked
was). Reading times are compared to sentences in which the nouns (the 16 items from Pearlmutter et al., 1999, plus four
number marking matches, as exemplified by (2): additional sentences with the same properties).
Each list included, for each condition, 10 of the items in
2. The key to the cabinet was rusty from many years of their experimental form and 10 of the items in their control
disuse. form (without rhymes in the case of the OR sentences;
without locally distracting nouns in the case of the number
In other words, reading times are higher at the verb when agreement sentences). The lists were counterbalanced such
the local information is distracting. Following the finding half of the subjects saw the experimental versions of
that text-chunking ability predicts decreased difficulty with sentences the other half saw in their control form.
complex sentences involving long-distance dependencies Procedure Materials were presented in random order using
(McCauley & Christiansen, 2015), we hypothesized that a self-paced, word-by-word moving window display (Just,
participants with higher Word Chunk Sensitivity scores Carpenter, & Woolley, 1982). At the beginning of each trial,
(Part 1) would be less susceptible to interference from local a series of dashes appeared (one corresponding to each
information in sentences such as (1). Subjects that are better nonspace character in the sentence). The first press of a
able to rapidly chunk words together and pass them to marked button caused the first word to appear, while
higher levels of representation should not only experience subsequent button presses caused each following word to
decreased computational burden from long-distance appear. The previous word would return once more to
dependencies, but should be less affected by locally dashes. Reaction times were recorded for each button press.
distracting information. Following each sentence, subjects answered a yes/no
The second sentence type featured object-relative (OR) comprehension question using buttons marked “Y” and “N.”
clauses, which have been shown to be processed with The task took approximately 10 minutes.
greater ease by good text chunkers (McCauley &
Christiansen, 2015). However, in the present study we Results and Discussion
added an element of phonological interference: two pairs of
Only trials with correct answers to comprehension questions
words in each sentence exhibited phonological overlap.
were analyzed. Accuracy for the number agreement
Previous work has shown that low-level phonological
condition was 88.3%; for the object-relatives it was 80.0%.
overlap can interfere with the processing of sentences
Following Acheson & MacDonald (2011), raw reaction
featuring relative clauses (Acheson & MacDonald, 2011).
times over 3000ms were excluded. Prior to analysis, raw
An experimental item and its matched control are shown in
reaction times (RTs) were log-transformed.
(3) and (4):
Mean RTs for the main verb in the number agreement and
phonological overlap sentences were comparable to those in
3. The cook that the crook consoles controls the politician.
the corresponding original studies (respectively: Pearlmutter
4. The prince that the crook comforts controls the politician.
et al., 1999; Acheson & MacDonald, 2011), as was the size
of the mean difference between conditions. In the number
In line with the Chunk-and-Pass framework, we predicted
agreement condition, the verb in experimental items
that better phonological chunkers, as assessed in Part 2,
(M=361.1, SE=19.9) was processed more slowly than in
would be less susceptible to phonological interference, by
controls (M=316.7, SE=13.9), a mean difference of 44ms
virtue of their ability to more rapidly chunk and pass
(F1[1,41]=12.7, p<0.001; F2[1,18]=10.2, p<0.01). There
phonological information to a higher level of representation.
was a fair amount of individual variation in the difference
2684
Fig. 1: Correlation between Word Chunk Sensitivity (derived from Fig. 2: Correlation between Phonological Chunk Sensitivity
recall scores in Part 1) and the difference in main verb RTs (derived from repetition scores in Part 2) and the difference in
between sentences with locally distracting number information vs. main verb RTs for OR sentences with and without phonological
control sentences. overlap between words.
between conditions (SD=79.4). term did not reach significance. The model for the
The critical main verb in OR sentences featuring significant main effect had an R value of 0.36. A scatterplot
phonological overlap was processed more slowly (M=605.1, showing the correlation between Phonological Chunk
SE=70.6) than in matched controls (M=546.3, SE=42.2), a Sensitivity and the RT difference is shown in Figure 2:
mean difference of 58.8 which was non-significant better chunking ability resulted in less phonological
(F1[1,41]=1.21, p=0.28; F2[1,18]=0.04, p=0.8; see interference.
discussion). There was, however, considerable individual Thus, consistent with the predictions of the Chunk-and-
variation in the difference between conditions (SD=343.7), Pass framework, we find evidence for the notion that
especially relative to the size of group mean difference. chunking ability shapes sentence processing differently at
We were primarily interested in the extent to which two separate levels of abstraction: participants who were
differences in RTs between experimental and control more sensitive to word chunk information better processed
sentences could be predicted by the Chunk Sensitivity long-distance dependencies in the face of conflicting local
measures collected in Parts 1 and 2. Below, we analyze information, while those with higher phonological chunk
these relationships using multiple linear regression, with sensitivity better processed complex sentences with
Word Chunk Sensitivity and Phonological Chunk phonological overlap between words. That the two chunk
Sensitivity scores as predictors of RT differences between sensitivity measures did not correlate with one another
conditions (recall that the two metrics were not correlated).2 further underscores the notion of chunking taking place at
For the difference between sentences featuring locally multiple levels of abstraction.
distracting number information and their control While we failed to find the same effect of phonological
counterparts, we found that Word Chunk Sensitivity was a overlap on processing as did Acheson and MacDonald
significant predictor of RT difference at the verb (β=-0.79, (2011), it is likely that our subjects (Cornell undergraduates)
t=-3.19, p<0.01), while Phonological Chunk Sensitivity and had more reading experience than subjects at UW-Madison,
the interaction term did not reach significance. The model and experienced less interference overall. Nonetheless, our
for the significant main effect had an R value of 0.42. The measure of phonological chunk sensitivity was sensitive
correlation between Word Chunk Sensitivity and the RT enough to pick up individual differences that predicted
difference is depicted in Figure 1. As can be seen, subjects sentence processing in the face of phonological interference.
with higher Word Chunk Sensitivity scores appear less Intriguingly, participants with very high Phonological
susceptible to interference from the locally distracting Chunk Sensitivity appeared to experience an advantage for
number information, as reflected by lower differences OR sentences featuring phonological overlap. This raises
between verb RTs for experimental vs. control sentences. the possibility that such subjects benefitted from
With regard to the difference between OR sentences with phonologically-based priming of subsequent rhyme words
and without phonological overlap, we found that in sentences such as (3). Further work will be necessary to
Phonological Chunk Sensitivity was a significant predictor evaluate this possibility.
of RT differences at the main verb (β=-3.49, t=-2.43,
p<0.05), while Word Chunk Sensitivity and the interaction General Discussion
In the present study, we show that individual differences in
2 We found that raw NWR performance scores resulted in
chunking ability predict on-line sentence processing at
weaker linear models and did not reach significance as a predictor. multiple levels of abstraction: chunking at the phonological
Therefore, we focus on the Phonological Chunk Sensitivity metric level is shown to predict the way phonological information
in the analyses (see Part 2).
2685
is used during complex sentence processing, while chunking during on-line sentence comprehension. Journal of Memory and
at the multiword level is shown to predict the ease with Language, 65, 193-207.
which long-distance dependencies are processed in the face Arnon, I. & Snider, N. (2010). More than words: Frequency effects
of conflicting local syntactic information. In Part 1, we for multi-word phrases. Journal of Memory and Language, 62,
67-82.
adapted the serial recall task—a paradigm used for over half Bannard, C. & Matthews, D. (2008). Stored word sequences in
a century to study memory, including chunking language learning. Psychological Science, 19, 241-248.
phenomena—in order to gain a measure of individual Christiansen, M.H. & Chater, N. (2016). The Now-or-Never
variation in subjects’ ability to chunk word sequences into bottleneck: A fundamental constraint on language. Behavioral &
multiword units. In Part 2, subjects participated in a NWR Brain Sciences, 39, e62.
task with non-words designed to vary according to the ease Cowan, N. (2001). The magical number 4 in short-term memory:
with which their phonemes could be chunked. The A reconsideration of mental storage capacity. Behavioral and
difference in correct repetition rates between highly chunk- Brain Sciences, 24, 87-114.
able and less chunk-able items provided a measure of Culicover, P.W. & Jackendoff, R. (2005). Simpler syntax. New
York: Oxford University Press.
individual variation in chunking ability at the phonological Ericsson, K.A., Chase, W.G., & Faloon, S. (1980). Acquisition of a
level. Finally, in Part 3 we showed that chunking at the memory skill. Science, 208, 1181-1182.
multiword level was predictive of processing for sentences Jones, G. (2012). Why chunking should be considered as an
with long-distance dependencies and distracting local explanation for developmental change before short-term
information, while chunking at the phonological level was memory capacity and processing speed. Frontiers in
predictive of complex sentence processing in the presence Psychology, 3 :167. DOI: 10.3389/fpsyg.2012.00167.
of phonological overlap between words. Jones, G., Gobet, F., Freudenthal, D., Watson, S.E. & Pine, J.M.
Expanding on the findings of a previous study that (2014). Why computational models are better than verbal
showed low-level chunking of sub-lexical letter sequences theories: The case of nonword repetition. Developmental
Science, 17, 298-310.
to predict sentence processing abilities (McCauley & Just, M. A., Carpenter, P. A., & Woolley, J. D. (1982). Paradigms
Christiansen, 2015), the present study supports the notion and processes in reading comprehension. Journal of
that chunking not only takes place at multiple levels of Experimental Psychology: General, 111, 228-238.
abstraction, but that individuals’ processing abilities may be McCauley, S.M. & Christiansen, M.H. (2011). Learning simple
differently shaped by chunking at each level. Moreover, statistics for language comprehension and production: The
chunking at lower levels (e.g., the phonological level) may CAPPUCCINO model. In L. Carlson, C. Hölscher, & T. Shipley
have serious consequences for processing at higher levels (Eds.), Proceedings of the 33rd Annual Conference of the
(e.g., sentence processing). Cognitive Science Society (pp. 1619-1624). Austin, TX:
This work is highly relevant to the study of language Cognitive Science Society.
McCauley, S.M. & Christiansen, M.H. (2014). Acquiring
acquisition. The Now-or-Never bottleneck imposes formulaic language: A computational model. Mental Lexicon, 9,
incremental, on-line processing constraints on language 419-436.
learning, suggesting a key role for chunking. Indeed, a McCauley, S.M. & Christiansen, M.H. (2015). Individual
number of recent computational modeling studies have differences in chunking ability predict on-line sentence
demonstrated that chunking can account for key empirical processing. In D.C. Noelle & R. Dale (Eds.), Proceedings of the
findings on children’s phonological development and word 37th Annual Conference of the Cognitive Science Society.
learning abilities (Jones, 2012; Jones et al., 2014), while Austin, TX: Cognitive Science Society.
other work has captured a role for chunking in learning to Miller, G.A. (1956). The magical number seven, plus or minus
comprehend and produce sentences (McCauley & two: Some limits on our capacity for processing information.
Psychological Review, 63, 81-97.
Christiansen, 2011, 2014). There exists a clear need for Miller, G.A. & Taylor, W.G. (1948). The perception of repeated
further developmental behavioral studies—including bursts of noise. Journal of the Acoustic Society of America, 20,
longitudinal studies—examining individual differences in 171-182.
chunking as they pertain to specific stages of language Pearlmutter, N.J., Garnsey, S.M. & Bock, K. (1999). Agreement
development as well as more general language learning processes in sentence comprehension. Journal of Memory and
outcomes. Language, 41, 427-456.
Remez, R.E., Ferro, D.F., Dubowski, K.R., Meer, J., Broder, R.S.
& Davids, M.L. (2010). Is desynchrony tolerance adaptable in
Acknowledgments the perceptual organization of speech? Attention, Perception, &
Thanks to Nick Chater and Gary Jones for helpful Psychophysics, 72, 2054-2058.
discussion, as well as S. Reig, K. Diamond, J. Kolenda, J. Studdert-Kennedy, M. (1986). Some developments in research on
Powell, S. Goldberg, and D. Dahabreh for assistance with language behavior. In N.J. Smelser & D.R. Gerstein (Eds.),
participant running and recruitment. Behavioral and social science: Fifty years of discovery (pp. 208-
248). Washington, DC: National Academy Press.
Tomasello, M. (2003). Constructing a language: A usage-based
References theory of language acquisition. Cambridge, MA: Harvard
Acheson, D.J. & MacDonald, M.C. (2011). The rhymes that the University Press.
reader perused confused the meaning: Phonological effects
2686
