Content uploaded by Erin Isbilen
Author content
All content in this area was uploaded by Erin Isbilen on Jul 28, 2018
Content may be subject to copyright.
Bridging artificial and natural language learning:
Comparing processing- and reflection-based measures of learning
Erin S. Isbilen (esi6@cornell.edu)
Cornell University, Department of Psychology, Ithaca, NY 14850 USA
Rebecca L.A. Frost (rebecca.frost@mpi.nl)
Max Planck Institute for Psycholinguistics, Language Development Department, Nijmegen, 6525 XD, Netherlands
Padraic Monaghan (p.monaghan@lancaster.ac.uk)
Lancaster University, Department of Psychology, Lancaster, LA1 4YF, UK
Morten H. Christiansen (christiansen@cornell.edu)
Cornell University, Department of Psychology, Ithaca, NY 14850 USA
Abstract
A common assumption in the cognitive sciences is that
artificial and natural language learning rely on shared
mechanisms. However, attempts to bridge the two have
yielded ambiguous results. We suggest that an empirical
disconnect between the computations employed during
learning and the methods employed at test may explain these
mixed results. Further, we propose statistically-based
chunking as a potential computational link between artificial
and natural language learning. We compare the acquisition of
non-adjacent dependencies to that of natural language
structure using two types of tasks: reflection-based 2AFC
measures, and processing-based recall measures, the latter
being more computationally analogous to the processes used
during language acquisition. Our results demonstrate that
task-type significantly influences the correlations observed
between artificial and natural language acquisition, with
reflection-based and processing-based measures correlating
within – but not across – task-type. These findings have
fundamental implications for artificial-to-natural language
comparisons, both methodologically and theoretically.
Keywords: statistical learning; chunking; language; artificial
language learning; cross-situational learning; non-adjacent
dependencies; learning; memory; serial recall; methodology
Introduction
Connecting individual differences in artificial and natural
language learning is an ongoing endeavor in the cognitive
sciences. These studies operate on the assumption that
artificial language learning tasks designed for use in the
laboratory draw on the same cognitive processes that
underpin language acquisition in the real world (e.g.,
Saffran, Aslin, & Newport, 1996). Yet, attempts to bridge
artificial and natural language learning have yielded mixed
results, often finding weaker correlations between language
measures that should in theory rely on shared computations
(Siegelman, Bogaerts, Christiansen & Frost, 2017). Part of
the problem may lie in the nature of the tests used to
evaluate learning; although artificial and natural language
learning may rely on the same computational processes,
different tests may tap into separate subcomponents of these
skills, making the relationship difficult to unpack.
Artificial language learning tasks are assumed to capture
key aspects of how learners acquire language in the real
world: by drawing on the distributional information
contained in speech. Through exposure to statistical
regularities in the input, the cognitive system picks up on
linguistic units without awareness by the learner (Saffran et
al., 1996). Yet, in adults, statistical learning is typically
tested using measures that require participants to reflect on
their knowledge and provide an overt judgment, such as in
the two-alternative forced-choice task (2AFC); a test that,
while potentially informative, only provides a meta-
cognitive measure of learning. Indeed, language learning
measures can be broadly divided into two categories:
reflection-based measures (e.g., 2AFC), which translate the
primary effects of learning into a secondary response, and
processing-based measures, which rely on the same
computations as the learning itself (Christiansen, 2018). In
psycholinguistic research, it is often the case that the
learning measures employed at test do not align with the
processes employed during learning. We propose that this
disconnect may have constrained prior observations of the
relationship between artificial and natural language skills.
We seek to resolve some of this ambiguity in the study at
hand.
In the current study, we assess the degree to which
statistical learning abilities map onto natural language
acquisition, and evaluate correlations within and between
reflection- and processing-based measures. For the purpose
of this paper, we characterize artificial language learning as
statistical learning in a highly constrained, simplified
context, using the Saffran et al. (1996) familiarization
method. We simulate natural language acquisition by
presenting participants with a more complex cross-
situational learning task that utilizes natural vocabulary and
grammar with corresponding referents. For each part of the
experiment, we included two types of tests: reflection-based
tasks (2AFC), and processing-based tasks (recall), to allow
for a comparison of learning between and within task types.
1856
For our processing-based measure, we employed a
chunking-based recall task, building on the suggestion that
chunking plays a key role in statistical learning and
language acquisition (see Christiansen, 2018, for a review).
In 2AFC tasks, participants are required to indicate their
preference for one stimulus over another, which is taken to
indicate learning. In recall – a task which is thought to rely
on chunking – participants repeat syllable strings that are
either congruent or incongruent with the statistics of the
input, with recall errors acting as a window into learning.
That is, learning is indexed by better recall of consistent
items when controlling for baseline phonological working
memory (Conway, Bauernschmidt, Huang & Pisoni, 2010;
Isbilen, McCauley, Kidd & Christiansen, 2017). If chunking
occurs during language acquisition, chunking-based tasks
may yield a better measure of learning than reflection-based
tasks such as 2AFC.
In the first part of the experiment, participants engaged in
a statistical learning task adapted from Frost and Monaghan
(2016), to test segmentation and generalization of non-
adjacent dependencies (the artificial language task). In the
second part, participants learned a fragment of Japanese,
comprising a small vocabulary and simple grammar using a
cross-situational learning task adapted from Rebuschat,
Ferrimand, and Monaghan (2017) and Walker,
Schoetensack, Monaghan and Rebuschat (2017; the natural
language task). We hypothesized that the correlations
observed between artificial and natural language learning
would show a strong effect of task type: reflection-based
measures would be more likely to correlate with other
reflection-based measures, whereas processing-based
measures would be more likely to correlate with other
processing-based measures. Such a pattern would have
important implications for individual differences work, and
about the deductions that can be applied to natural language
acquisition from artificial language learning tasks.
Part 1: Non-adjacent dependency learning in
an artificial language
In Part 1, we tested adults’ learning of an artificial language
composed of non-adjacent dependencies, which are
relationships between linguistic units that occur across one
or more variable intervening units (e.g., in an AXC structure
where units A and C reliably co-occur, but X varies
independently). These dependencies are found at multiple
levels of linguistic abstraction, including morphology within
words and syntactic dependencies between words, thereby
providing a tightly-controlled artificial structure that shares
structural similarity with natural language.
We examined learners’ ability to segment these non-
adjacent dependency sequences from speech, and generalize
them to new instances - skills which are integral to natural
language learning. We tested both segmentation and
generalization with a reflection-based task (2AFC), and a
processing-based task, the statistically-induced chunking
recall task (SICR; Isbilen et al., 2017). In the SICR task,
participants are presented with 6-syllable-long strings, that
are either composed of two words from the input, or the
same syllables presented in a random order. If participants
have successfully chunked the items in the artificial
language during training, we expect that they should
perform significantly better on recalling the strings derived
from the statistics of the input language. While 2AFC is
scored as a correct-incorrect binary, SICR is scored syllable-
by-syllable, which we suggest may provide more in-depth
insights into segmentation and generalization skills.
Building on the results of Frost and Monaghan (2016), we
hypothesized that both tasks would yield evidence of
simultaneous segmentation and generalization. However,
due to the differences in task demands between reflection-
and processing-based tests, we expected to see limited
correlations between measurement types.
Method
Participants 49 Cornell University undergraduates (30
females; age: M=19.43, SD=1.30) participated for course
credit. All participants were native English speakers, with
no experience learning Japanese.
Materials The same language and stimuli as Frost and
Monaghan (2016) were used, derived from Peña, Bonatti,
Nespor and Mehler (2002). The language was composed of
9 consonant-vowel syllables (be, du, fo, ga, li, ki, pu, ra, ta),
arranged into three tri-syllabic non-adjacent dependencies
containing three varying middle syllables (A1X1–3C1,
A2X1–3C2, and A3X1–3C3; 9 words in total). Four different
versions of the language were created to control for
potential preferences for certain phoneme combinations.
Syllables used for the A and C items contained plosives (be,
du, ga, ki, pu, ta), while the X syllables contained
continuants (fo, li, ra). The resulting 9 items are referred to
as segmentation words, sequences that were presented
during training. Nine generalization words were also
created, and were only presented at test. The generalization
words contained trained non-adjacent dependencies, but
with novel intervening syllables (thi, ve, zo, e.g., A1Y1-3C1).
The generalization words measure participants’ ability to
extrapolate the knowledge of the non-adjacent dependencies
gained during training to novel, unheard items.
For the 2AFC test, 18 additional foil words were created,
which were paired with segmentation and generalization
words. Foils for the segmentation test comprised part-word
sequences that spanned word boundaries (e.g., CAX,
XCA). Foils for the generalization test were part-words but
with one syllable switched out and replaced with a novel
syllable, to prevent participants from responding based on
novelty alone (e.g., NCA, XNA, CAN, see Frost &
Monaghan, 2016). For the SICR test, 27 six-syllable strings
were created: 9 composed of two concatenated segmentation
words (e.g., A1X1C1A2X2C2), 9 composed of two
generalization words (e.g., A1Y1C1A2Y2C2), and 9 foils.
The foils used the same syllables as the experimental items
in a pseudorandomized order that avoided using any
1857
transitional probabilities or non-adjacent dependencies from
the experimental items.
All stimuli were created using the Festival speech
synthesizer (Black et al., 1990). Each AXC string lasted
~700 ms, and was presented using E-Prime 2.0.
Procedure For training, the 9 segmentation words were
concatenated into a continuous stream that participants
heard for 10.5 minutes. Participants were instructed to listen
carefully to the language and pay attention to the words it
might contain.
To test learning, two different tasks were used: the 2AFC
task and the SICR task (Isbilen et al., 2017). The order of
the two tests was counterbalanced to account for potential
order effects. In the 2AFC task, participants were presented
with 18 pairs of words: 9 segmentation pairs and 9
generalization pairs, with each pair featuring a target word
and corresponding foil. Segmentation and generalization
trials were randomized within the same block of testing.
Participants were instructed to carefully listen to each word
pair and indicate which of the two best matched the
language they heard during training. In the SICR task, 27
strings were randomly presented for recall: 9 segmentation
items, 9 generalization items, and 9 foils that served as a
baseline working memory measure. Participants were asked
to listen to each string and say the entire string out loud to
the best of their ability. Participants were not informed of
any underlying structure of the strings in either task.
Results and Discussion
First, we examined the data for task order effects (2AFC
first/SICR second versus SICR first/2AFC second), and
language effects (which of the four randomized languages
participants heard). A one-way ANOVA revealed a
significant effect of order on both SICR measures
(Segmentation: F(3,45)=-2.30, p=.026; Generalization:
F(3,45)=-3.30, p=.002), with participants who received
2AFC prior to SICR scoring significantly higher on these
two measures. Similarly, language significantly impacted
SICR generalization performance, F(3,45)=6.94, p=.0006,
suggesting that different syllable combinations may vary in
difficulty when being spoken aloud. All subsequent analyses
involving SICR in the remainder of the paper control for
order, and for SICR generalization, for both order and
language.
2AFC Performance Replicating the findings of Frost
and Monaghan (2016), participants showed simultaneous
segmentation and generalization of non-adjacent
dependencies, with performance on both tasks being
significantly above chance (Segmentation: M=.84, SD=.13;
t(48)= 18.44, p<.0001; Generalization: M=.70, SD=.21;
t(48)= 6.61, p<.0001). Performance was significantly more
accurate on segmentation than generalization trials,
t(48)=5.77, p<.0001, and segmentation and generalization
scores were highly correlated: r(47)=.53, p<.0001.
SICR Performance Participants’ verbal responses from
the SICR task were transcribed by two coders blind to the
study design. The transcriptions were subsequently scored
against the target test items to obtain measures of overall
accuracy (the total number of syllables recalled in the
correct serial position), and non-adjacent dependency
accuracy (the number of A-C pairings recalled from each
item, out of the two possible pairings: e.g., A1xC1A2xC2).
Replicating the results of Isbilen et al. (2017), participants
accurately recalled significantly more syllables in the
correct order for the experimental items than the random
items. These results held for both the segmentation items
(Experimental: M=34.84, SD=10.16; Random: M=13.55,
SD=6.04; t(48)= 23.11, p<.0001), and for the generalization
items (Experimental: M=27.71, SD=10.56; Random:
M=8.35, SD=4.37; t(48)= 15.98, p<.0001). Similarly, the
number of non-adjacent dependencies (syllables in the 1st &
3rd and/or the 4th & 6th serial positions) recalled for
experimental items was significantly higher than those
recalled for the random, both for the segmentation items
(Experimental: M=8.53, SD=4.52; Random: M=2.57,
SD=2.34; t(48)= 13.34, p<.0001), and for the generalization
items (Experimental: M=6.63, SD=4.42; Random: M=1.50,
SD=1.45; t(48)= 9.51, p<.0001). Unlike 2AFC, the SICR
results revealed no significant difference in performance
between the segmentation and generalization difference
scores (experimental minus random), although
generalization scores were slightly lower due to the
inclusion of unfamiliar syllables. These results held for both
overall recall, and for the total number of non-adjacent
dependencies recalled (p=.10 in both cases). This difference
between 2AFC and SICR may in part stem from differences
in task demands: differences in familiarity between
segmentation and generalization items may influence ratings
in 2AFC more, due to its meta-cognitive nature. SICR
generalization and segmentation performance was
significantly correlated: r(47)=.34, p=.02.
Correlations between 2AFC and SICR To evaluate the
relationship between reflection- and processing-based
measures, correlations were run between 2AFC and SICR
scores. The SICR values used for the correlations were the
total difference scores, to maximize the measures’
comparability to 2AFC (which is akin to a difference score),
while also controlling for baseline phonological working
memory (the subtraction of the random items from the
experimental items). The only significant correlation was
between 2AFC segmentation and SICR segmentation,
r(47)=.31, p=.04 (not correcting for multiple comparisons).
1
No other SICR and 2AFC measures were significantly
correlated (all p>.08). In line with our hypothesis, these
findings suggest that reflection- and processing-based
measures appear to capture largely different aspects of
statistical learning skills (Christiansen, 2018; Siegelman et
al., 2017).
The results of Part 1 replicate the results of Frost &
Monaghan (2016) of simultaneous segmentation and
generalization of non-adjacent dependencies using both
2AFC and SICR. Taken together, these findings suggest that
1
In a pilot version of this study, (N=61) no such correlation was
observed, potentially suggesting a type II error: r(59)=-.04, p=.76.
1858
statistical-chunking processes may be able to account for the
segmentation and generalization of non-adjacent
dependencies, as well as that of sequential dependencies.
Furthermore, we found that although reflection- and
processing-based measures showed evidence of learning,
performance across the two tasks was largely uncorrelated.
To test whether this pattern extends to natural language
acquisition, Part 2 of the experiment evaluated grammar and
vocabulary acquisition using patterns from natural language,
with a comparison of 2AFC and recall task types.
Part 2: Cross-situational language learning of
Japanese
Natural language acquisition involves a host of different
factors, including word segmentation, word-referent
mapping, discovery of sequential structure, and
generalization to novel instances. In the second part of this
experiment, we increased the complexity of the task to
explore the degree to which the learning taking place during
segmentation and the discovery of non-adjacent dependency
structure maps onto more naturalistic language acquisition.
A cross-situational language learning task based on Walker
et al. (2017) was administered, exposing participants to
Japanese sentences co-occurring with complex scenes.
Cross-situational learning simulates naturalistic language
learning in the lab by analogy to infants’ acquisition of
word-referent mappings non-ostensively, through hearing
instances of a word occurring with the same referent across
different contexts. Similar to Part 1, both reflection-based
and processing-based measures were used to evaluate
learning. 2AFC tests of noun, marker, and verb learning
were performed. Additionally, a combined forced-choice
and recall task was also administered to test syntax
acquisition: participants repeated whole sentences they
heard out loud, after which they rated the grammaticality of
each sentence. We hypothesized that vocabulary and
grammatical regularities would be acquired simultaneously,
similar to the concurrent segmentation and generalization in
the non-adjacent dependency task. Furthermore, we
anticipated that while all tests would show some evidence of
learning, only within task-type correlations would be
significantly related.
Method
Participants The same 49 participants from Part 1
partook in Part 2 immediately following the first task.
Materials A small lexicon of Japanese words was used
for this experiment, taken from Rebuschat et al. (2017). The
language consisted of 6 nouns (fukuoru, owl; kame, turtle;
niwatori, chicken; shimauma, zebra; ushi, cow; zou,
elephant), four verbs (kakusu, hide; mochiageru, lift; taosu,
push; tobikueru, jump), and two case markers (-ga and -o),
which were appended to the end of each noun to indicate
whether the noun was the subject (-ga) or the object (-o) of
the sentence. For instance, the sentence “kamega
shimaumao taosu” would indicate that the turtle (subject)
pushes the zebra (object). The language also used Japanese
syntax, with sentences having two possible grammatical
orders: subject-object-verb (SOV), and object-subject-verb
(OSV). For training, 192 sentences were generated. For test,
96 additional sentences were presented: 24 for each of the
marker, noun, and verb tests, and 24 for the combined
syntax and recall task. Of the syntax stimuli, 12 were
ungrammatical items that used word orderings that are
invalid according to Japanese syntax (OVS, VOS, VSO,
SVO). The frequency, order, and object-subject assignment
of each word were all balanced. All auditory training and
test stimuli were created by a native Japanese speaker.
With each sentence, complex scenes depicting cartoon
animals as the referents for the nouns were also presented,
engaging in the action indicated by the verb of each
sentence (hiding, lifting, pushing, or jumping). During
training, two such scenes were presented, one the target
scene and the other a distractor, to allow for the accrual of
word-referent mappings through the use of cross situational
statistics. During the syntax test, only the target scene was
presented. All stimuli were presented in E-Prime 2.0.
Procedure The experiment consisted of two training
blocks, and two test blocks. During training participants
heard a Japanese sentence while watching two scenes play
on the computer screen: one displaying the target, and the
other the foil. The foil scene varied from the target both in
terms of the nouns and actions depicted. Participants were
asked to judge to the best of their ability which scene the
sentence referred to. Unknown to the participant, the last
trials of training tested their knowledge of the nouns, verbs,
and markers of the language, using the same method by
varying the two scenes by just one property (e.g., only one
object was different, or only the action was different, or
only the subject/object roles were different). Following
training, 12 syntax test trials were administered, which
presented a sentence paired with a single scene. Participants
were told that the speaker of these sentences were learning
Japanese, and that their task was to repeat the speaker’s
sentence out loud (the recall measure), and then indicate
whether the speaker’s sentence sounded “good” or “funny”
(the forced-choice measure). While this task is slightly
different from the other 2AFC measures in this experiment,
in which participants choose between a target and a foil,
they both require participants to engage in reflection about
learned material.
After the conclusion of the first training block and syntax
test, the same procedure was completed once more, starting
with training and ending with another syntax test. Each
training block contained 4 marker test trials, 4 verb test
trials, and 6 noun test trials. Each syntax test block
contained 12 test trials: 6 grammatical, and 6 ungrammatical
sentences. No feedback about participants’ performance was
provided at any time during training or test.
Results and Discussion
2AFC results Following the methods employed by Walker
et al. (2017), data from both testing blocks were pooled for
the analyses. With the exception of noun learning (M=.54,
1859
SD=.20; t(48)=1.58, p=.12), scores on all 2AFC measures
were significantly above chance (Marker test: M=.57,
SD=.19; t(48)=2.80, p=.0075; Verb test: M=.60, SD=.14;
t(48)=5.52, p=.0075; Syntax test: M=.61, SD=.14;
t(48)=5.52, p=.0075). Thus, our results showed that
vocabulary and grammatical acquisition of natural language
structure can occur simultaneously. The lack of significant
noun learning, while inconsistent with the findings of
Walker et al. (2017), may be explained by the fact that the
nouns in this task were longer (containing more syllables),
which may have contributed to reduced learning. The only
2AFC tests that were significantly correlated with each
other were performance on the syntax and verb test,
r(47)=.39, p=.0065, which is consistent with the findings of
Walker et al. (2017).
Syntax recall results Participants’ verbal responses were
transcribed by a coder blind to the study design, and were
scored against the targets, with a point given for each
syllable recalled in the correct serial position. Overall, no
effect of grammaticality was found on recall performance,
with participants recalling approximately equal numbers of
syllables in the correct order for both the grammatical
(M=73.08, SD=19.46) and ungrammatical items (M=72.14,
SD=18.57; t(48)= .60, p=.55). However, unlike the artificial
language stimuli, the natural language recall items vary in
the total number of syllables, ranging from 8 to 14 syllables.
A linear mixed effects model of the raw recall data revealed
that while grammaticality and string length had no effect
independently on recall performance, the interaction of
grammaticality and string length was significant,
t(1147)=2.78, p=.0055, with length relating to significant
detriments in recall scores for ungrammatical items, but not
grammatical ones. This suggests that learning the statistical
structure of the language stabilized recall of the grammatical
items independent of length, whereas the ungrammatical
items were more severely impacted by memory limitations.
Correlations between 2AFC and recall To investigate
the connection between the reflection- and processing-based
measures in the cross-situational learning task, correlations
were run between all 2AFC measures in Part 2 of the
experiment, and the recall difference scores. No significant
correlations were found between any of the 2AFC measures
and recall performance (all p>.14). These results may be
taken as further evidence that reflection- and processing-
based tests do not measure the same aspects of learning.
The relationship between artificial & natural
language learning
To determine the connection between non-adjacent
dependency learning in an artificial context to vocabulary
and grammar acquisition in a natural language context,
correlations were calculated between the data from Parts 1
and 2 of the experiment. These analyses were first
performed within task type, then between task types. We
predicted that the 2AFC measures from Part 1 would only
correlate with the 2AFC measures from Part 2, and that only
the recall measures from Parts 1 and 2 would be correlated.
Parts 1 & 2: Reflection-based measures
Correlations between all reflection-based (2AFC) measures
were performed. However, only two 2AFC measures were
correlated across the two parts of the experiment. First,
participants’ ability to segment words in the non-adjacent
dependency task positively correlated with their ability to
learn the nouns in the cross-situational learning task,
r(47)=.35, p=.0139. Second, generalization ability on the
non-adjacent dependency 2AFC task negatively correlated
with participants’ ability to pick up on the markers on the
cross-situational learning task, r(47)=-.28, p=.0496.
Parts 1 & 2: Processing-based measures
Partial correlations between the SICR and cross-situational
recall performance raw data, controlling for string length,
item type (experimental versus random for SICR, or
grammatical versus ungrammatical for the cross-situational
items) and repeated measures revealed that SICR and cross-
situational recall abilities significantly correlated with one
another. SICR segmentation (r(615)=.20, p<.0001)
demonstrated a stronger correlation to cross-situational
recall than did SICR generalization (r(561)=.14, p<.0009).
Parts 1 & 2: Between measure-types
Correlations between reflection- and processing-based tests
from Parts 1 and 2 of the experiment were performed. The
results revealed no significant correlations between task
types from the two parts of the experiment (Table 1).
Table 1: 2AFC and Recall Correlations
Marker
Test
Noun
Test
Syntax
Test
Verb
Test
2AFC
Seg.
2AFC
Gen.
SICR
Seg.
.02
.02
.12
.13
.31*
.26
SICR
Gen.
-.09
.08
-.07
.15
.12
.16
Cross-sit.
recall
-.22
.18
.09
-.02
.17
.15
General Discussion
Bridging artificial and natural language learning is an
important endeavor in the cognitive sciences. A first step to
stabilizing the link between in-lab observations and real-
world behavior may come from strengthening the
connection between the tasks used to test learning, and the
computations employed during learning. Here, we argue for
the role of statistically-based chunking as the computational
link between learning and testing.
Short-term memory recall is a robust indicator of long-
term learning (Jones & Macken, 2015; McCauley, Isbilen &
Christiansen, 2017), with the accrual of statistical
regularities over time aiding memory retention in the here-
and-now. The use of such recall tasks as a measure of in-lab
language learning is motivated by evidence supporting the
notion that chunking may play a key role in statistical
learning, and can account for language acquisition,
1860
processing, and production more broadly (Christiansen &
Chater, 2016). Our results strengthen this argument by
demonstrating that statistically-based chunking can also
account for the simultaneous learning and generalization of
non-adjacent dependencies: a complex, dynamic linguistic
structure. While our findings appear to support some degree
of separability between segmentation and generalization
skills (see also Frost & Monaghan, 2017), these abilities
also appear to inform one another, with segmentation
performance strongly predicting generalization ability.
While our study has important methodological and
theoretical implications for individual differences work, it
also has a number of limitations. First, although our cross-
situational paradigm simulates aspects of natural language
acquisition in the lab, it does not capture language
acquisition exactly as it occurs in the real world. Second,
while this language learning experiment implemented many
separate subcomponents of natural language by design, we
also acknowledge that these features changed the task
demands. Learning in the artificial language task involved
only segmentation and generalization of individual words,
whereas the cross-situational learning task incorporated
complex grammar, referents, and whole sentences. Stronger
correlations may have been observed if the structure of the
two different tasks were more similar (see Siegelman et al.,
2017, for discussion).
Methodologically, our results suggest that the empirical
disconnect between the learning targeted and the measures
used at test may influence the correlations observed between
artificial and natural language outcomes. While the
processes leveraged for both in-lab and real-world language
acquisition may be analogous, the similarity or dissimilarity
of the tasks used to measure learning – and the specific
computations each task relies on – may obscure the
connection between the targeted cognitive processes.
Moreover, there appears to be substantial individual
variation in performance on these two kinds of tasks, as
evidence of learning on one measure does not necessarily
translate to high performance on the other. While both
reflection- and processing-based measures test learning, our
results suggest that they may test slightly different kinds of
knowledge: meta-cognitive reflections over what was
learned, versus processing-based facilitation from accrued
statistics.
Acknowledgments
We would like to thank Phoebe Ilevebare, Farrah Mawani,
Eleni Kohilakis, Olivia Wang, Dante Dahabreh, and Jake
Kolenda for their help collecting and coding data. This work
was in part supported by NSF GRFP (#DGE-1650441)
awarded to ESI, a Cornell Department of Psychology
Graduate Research Grant awarded to ESI, and by the
International Centre for Language and Communicative
Development (LuCiD) at Lancaster University, funded by
the Economic and Social Research Council (UK)
[ES/L008955/1].
References
Christiansen, M.H. (2018). Implicit-statistical learning: A tale
of two literatures. Topics in Cognitive Science.
Christiansen, M.H. & Chater, N. (2016). The Now-or-Never
bottleneck: A fundamental constraint on language.
Behavioral and Brain Sciences, 39, e62.
Conway, C.M., Bauernschmidt, A., Huang, S.S. & Pisoni,
D.B. (2010). Implicit statistical learning in language
processing: Word predictability is the key. Cognition, 114,
356-371.
Frost, R.L.A., & Monaghan, P. (2016). Simultaneous
segmentation and generalisation of non-adjacent
dependencies from continuous speech. Cognition, 147, 70-
74.
Frost, R.L.A., & Monaghan, P. (2017). Sleep-driven
computations in speech processing. PloS one, 12(1),
e0169538.
Jones, G. & Macken, B. (2015). Questioning short-term
memory and its measurement: Why digit span measures
long-term associative learning. Cognition, 144, 1-13.
Isbilen, E.S., McCauley, S.M., Kidd, E. & Christiansen, M.H.
(2017). Testing statistical learning implicitly: A novel
chunk-based measure of statistical learning. Proceedings of
the 39th Annual Conference of the Cognitive Science
Society. Austin, TX: Cognitive Science Society.
McCauley, S.M. & Christiansen, M.H. (2011). Learning
simple statistics for language comprehension and
production: The CAPPUCCINO model. Proceedings of the
33rd Annual Conference of the Cognitive Science. Austin,
TX: Cognitive Science Society.
McCauley, S.M., Isbilen, E.S. & Christiansen, M.H. (2017).
Chunking ability shapes sentence processing at multiple
levels of abstraction. Proceedings of the 39th Annual
Conference of the Cognitive Science Society. Austin, TX:
Cognitive Science Society.
Peña, M., Bonatti, L., Nespor, M., & Mehler, J. (2002).
Signal-driven computations in speech processing. Science,
298, 604-607.
Rebuschat, P., Ferrimand, H., & Monaghan, P. (2017). Age
effects in statistical learning of Japanese: Evidence from the
cross-situational learning paradigm. Talk presented at the
International Association for the Study of Child Language.
Saffran, J. R., Newport, E. L., & Aslin, R. N. (1996). Word
segmentation: The role of distributional cues. Journal of
memory and language, 35(4), 606-621.
Siegelman, N., Bogaerts, L., Christiansen, M. H. & Frost, R.
(2017). Towards a theory of individual differences in
statistical learning. Phil. Trans. R. Soc. B, 372(1711),
20160059.
Walker, N., Schoetensack, C., Monaghan, P., & Rebuschat, P.
Simultaneous acquisition of vocabulary and grammar in an
artificial language learning task. Proceedings of the 39th
Annual Conference of the Cognitive Science Society. Austin,
TX: Cognitive Science Society.
1861
