rstb.royalsocietypublishing.org Research Cite this article: Smith K, Perfors A, Fehe ´r O, Samara A, Swoboda K, Wonnacott E. 2017 Language learning, language use and the evolution of linguistic variation. Phil. Trans. R. Soc. B 372: 20160051. http://dx.doi.org/10.1098/rstb.2016.0051 Accepted: 1 August 2016 One contribution of 13 to a theme issue ‘New frontiers for statistical learning in the cognitive sciences’. Subject Areas: evolution Keywords: learning, iterated learning, language Author for correspondence: Kenny Smith e-mail: [email protected]Language learning, language use and the evolution of linguistic variation Kenny Smith 1 , Amy Perfors 2 , Olga Fehe ´r 1 , Anna Samara 3 , Kate Swoboda and Elizabeth Wonnacott 3 1 University of Edinburgh, Edinburgh, UK 2 University of Adelaide, Adelaide, Australia 3 University College London, London, UK KS, 0000-0002-4530-6914 Linguistic universals arise from the interaction between the processes of language learning and language use. A test case for the relationship between these factors is linguistic variation, which tends to be conditioned on linguistic or sociolinguistic criteria. How can we explain the scarcity of unpredictable variation in natural language, and to what extent is this property of language a straightforward reflection of biases in statistical learning? We review three strands of experimental work exploring these questions, and introduce a Bayesian model of the learning and transmission of linguistic variation along with a closely matched artificial language learning experiment with adult participants. Our results show that while the biases of language learners can potentially play a role in shaping linguistic systems, the relationship between biases of learners and the structure of languages is not straightfor- ward. Weak biases can have strong effects on language structure as they accumulate over repeated transmission. But the opposite can also be true: strong biases can have weak or no effects. Furthermore, the use of language during interaction can reshape linguistic systems. Combining data and insights from studies of learning, transmission and use is therefore essential if we are to understand how biases in statistical learning interact with language transmission and language use to shape the structural properties of language. This article is part of the themed issue ‘New frontiers for statistical learning in the cognitive sciences’. 1. Introduction Natural languages do not differ arbitrarily, but are constrained so that certain properties recur across languages. These linguistic universals range from fundamental design features shared by all human languages to probabilistic typo- logical tendencies. Why do we see these commonalities? One widespread intuition (see e.g. [1]) is that linguistic features which are easier to learn or which offer advantages in processing and/or communicative utility should spread at the expense of less learnable or functional alternatives. They should therefore be over-represented cross-linguistically, suggesting that linguistic universals arise from the interaction between the processes of language learning and language use. In this paper, we take linguistic variation as a test case for exploring this relationship between language universals and language learning and use. Variation is ubiquitous in languages: phonetic, morphological, syntactic, semantic and lexical variation are all common. However, this variation tends to be predict- able: usage of alternate forms is conditioned (deterministically or probabilistically) in accordance with phonological, semantic, pragmatic or sociolinguistic criteria. For instance, in many varieties of English, the last sound in words like ‘cat’, ‘bat’ and ‘hat’ has two possible realizations: either [t], an alveolar stop, or [ ], a glottal stop. However, whether [t] or [ ] is used is not random, but conditioned on linguistic and social factors. For instance, Stuart-Smith [2] showed that T-glottaling & 2016 The Authors. Published by the Royal Society under the terms of the Creative Commons Attribution License http://creativecommons.org/licenses/by/4.0/, which permits unrestricted use, provided the original author and source are credited. on June 13, 2018 http://rstb.royalsocietypublishing.org/ Downloaded from
13
Embed
Language learning, language use and the evolution of ...rstb.royalsocietypublishing.org/content/royptb/372/1711/20160051... · Language learning, language use and the ... language
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
on June 13, 2018http://rstb.royalsocietypublishing.org/Downloaded from
rstb.royalsocietypublishing.org
ResearchCite this article: Smith K, Perfors A, Feher O,
& 2016 The Authors. Published by the Royal Society under the terms of the Creative Commons AttributionLicense http://creativecommons.org/licenses/by/4.0/, which permits unrestricted use, provided the originalauthor and source are credited.
Language learning, language use and theevolution of linguistic variation
Kenny Smith1, Amy Perfors2, Olga Feher1, Anna Samara3, Kate Swobodaand Elizabeth Wonnacott3
1University of Edinburgh, Edinburgh, UK2University of Adelaide, Adelaide, Australia3University College London, London, UK
KS, 0000-0002-4530-6914
Linguistic universals arise from the interaction between the processes of
language learning and language use. A test case for the relationship between
these factors is linguistic variation, which tends to be conditioned on linguistic
or sociolinguistic criteria. How can we explain the scarcity of unpredictable
variation in natural language, and to what extent is this property of language
a straightforward reflection of biases in statistical learning? We review three
strands of experimental work exploring these questions, and introduce a
Bayesian model of the learning and transmission of linguistic variation
along with a closely matched artificial language learning experiment with
adult participants. Our results show that while the biases of language learners
can potentially play a role in shaping linguistic systems, the relationship
between biases of learners and the structure of languages is not straightfor-
ward. Weak biases can have strong effects on language structure as they
accumulate over repeated transmission. But the opposite can also be true:
strong biases can have weak or no effects. Furthermore, the use of language
during interaction can reshape linguistic systems. Combining data and
insights from studies of learning, transmission and use is therefore essential
if we are to understand how biases in statistical learning interact with language
transmission and language use to shape the structural properties of language.
This article is part of the themed issue ‘New frontiers for statistical learning
in the cognitive sciences’.
1. IntroductionNatural languages do not differ arbitrarily, but are constrained so that certain
properties recur across languages. These linguistic universals range from
fundamental design features shared by all human languages to probabilistic typo-
logical tendencies. Why do we see these commonalities? One widespread intuition
(see e.g. [1]) is that linguistic features which are easier to learn or which offer
advantages in processing and/or communicative utility should spread at the
expense of less learnable or functional alternatives. They should therefore be
over-represented cross-linguistically, suggesting that linguistic universals arise
from the interaction between the processes of language learning and language use.
In this paper, we take linguistic variation as a test case for exploring this
relationship between language universals and language learning and use.
Variation is ubiquitous in languages: phonetic, morphological, syntactic, semantic
and lexical variation are all common. However, this variation tends to be predict-
able: usage of alternate forms is conditioned (deterministically or probabilistically)
in accordance with phonological, semantic, pragmatic or sociolinguistic criteria.
For instance, in many varieties of English, the last sound in words like ‘cat’, ‘bat’
and ‘hat’ has two possible realizations: either [t], an alveolar stop, or [ ], a glottal
stop. However, whether [t] or [ ] is used is not random, but conditioned on
linguistic and social factors. For instance, Stuart-Smith [2] showed that T-glottaling
input to G1 input to G2 ...G1 G5 G5 languageG1 language
cow fipsi
ngle
mul
tiple
cow tay
pig fip
pig fip
pig tay
cow fip
cow fip
cow fip
pig tay
pig tay
pig tay
...
cow fip
cow fip
cow tay
pig fip
pig fip
pig tay
...
cow fip
cow fip
cow tay
pig fip
pig fip
pig tay
...
cow fip
cow tay
cow tay
pig fip
pig fip
pig tay
...
cow fip
cow fip
cow tay
pig fip
pig fip
pig tay
cow fip
cow fip
cow tay
pig fip
pig fip
pig tay
...
cow fip
cow fip
cow tay
pig fip
pig tay
pig tay
cow fip
cow tay
cow tay
pig fip
pig fip
pig tay
...
cow fip
cow fip
cow tay
pig fip
pig tay
pig tay
...
cow fip
cow fip
cow tay
pig fip
pig fip
pig tay
cow fip
cow fip
cow tay
pig fip
pig fip
pig tay
...
cow fip
cow fip
cow fip
pig fip
pig tay
pig tay
...
cow fip
cow fip
cow fip
pig fip
pig tay
cow fip
cow fip
cow fip
pig fip
pig tay
pig tay pig tay
... ...
cow fip
cow tay
cow tay
pig fip
pig fip
pig fip
...
cow fip
cow tay
cow tay
pig fip
pig fip
pig tay
...
Figure 1. Illustration of single-person and multiple-person chains, here with S ¼ 2. In single-person chains (a), each individual learns from the (duplicated, as S ¼ 2)data produced by the single individual at the previous generation. In multiple-person chains (b), each individual learns from the pooled language produced by theindividuals at the previous generation, with all individuals at a given generation being exposed to the same pooled input. (Online version in colour.)
rstb.royalsocietypublishing.orgPhil.Trans.R.Soc.B
372:20160051
3
on June 13, 2018http://rstb.royalsocietypublishing.org/Downloaded from
3. Transmission(a) Iterated learning and regularizationAs well as being restructured by the biases of individual
language learners, languages are shaped by processes of
transmission. Modelling work exploring how socially learned
systems change as they are transmitted from person to
person has established that weak biases in learning can be
amplified as a result of transmission (e.g. [17]): their effects
accumulate generation after generation. The same insight has
been applied to the regularization of unpredictable variation.
Reali & Griffiths [18] and Smith & Wonnacott [19] use an
experimental iterated learning paradigm where an artificial
language is transmitted from participant to participant, each
learner learning from data produced by the previous parti-
cipant in a chain of transmission. Both studies show that
unpredictable variation, present in the language presented to
the first participant in each chain of transmission, is gradually
eliminated, resulting in the emergence of languages entirely
lacking unpredictable variation. This happens even though
each learner has only weak biases against variability—both
studies used adult participants and relatively simple learning
tasks, providing ideal circumstances for probability matching.
Let us consider one of these studies in more detail. Smith &
Wonnacott [19] trained participants on a miniature language
for describing simple scenes involving moving animals,
where every scene consisted of one or two animals (pig, cow,
giraffe or rabbit) performing an action (a movement), and the
accompanying description consisted of a nonsense verb, a
noun, and (for scenes featuring two animals) a post-nominal
marker indicating plurality. This plural marker varied unpre-
dictably: sometimes plurality was marked with the marker
fip, sometimes with the marker tay. After training on this minia-
ture language, participants labelled the same scenes repeatedly,
generating a new miniature language. The language produced
by one participant was then used as the training language for
the next participant in a chain of transmission, passing the
language from person to person (figure 1a).
When trained on an unpredictably variable input
language, most participants in this experiment reproduced
that variability fairly faithfully: their use of the plural marker
was statistically indistinguishable from probability matching.
However, when the language was passed from person to
person, plural marking became increasingly predictable.
While some chains of transmission gradually converged on a
system where only one plural marker was used (e.g. plurality
was always marked with fip, with tay dying out), the most
common outcome after five ‘generations’ of transmission was
a conditioned system of variation: some nouns marked the
plural with fip, other nouns used tay and the choice of
marker was entirely predicted by the noun being marked.
The language as a whole therefore retained variation, but (as
in natural languages) that variability was conditioned, in this
case, on the linguistic (lexical) context.
This shows that transmission of language from person to
person via iterated learning, intended as an experimental
Table 1. Measures of variability for illustrative scenarios in a population consisting of a single speaker (s1: first three rows) or two speakers (s1 and s2,remaining rows), each speaker producing two labels for each of two nouns (cow and pig) using two plural markers (fip and tay). Note that, for a languageproduced by a single speaker, H(MarkerjNoun) and H(MarkerjNoun, Speaker) are necessarily identical.
language P(fip) H(Marker) H(MarkerjNoun) H(MarkerjNoun, Speaker)
s1: cow fip, cow fip, pig fip, pig fip 1 0 0 0
s1: cow fip, cow fip, pig tay, pig tay 0.5 1 0 0
s1: cow fip, cow tay, pig fip, pig tay 0.5 1 1 1
s1: cow fip, cow fip, pig fip, pig fip 1 0 0 0
s2: cow fip, cow fip, pig fip, pig fip
s1: cow fip, cow fip, pig tay, pig tay 0.5 1 0 0
s2: cow fip, cow fip, pig tay, pig tay
s1: cow fip, cow fip, pig tay, pig tay 0.5 1 1 0
s2: cow tay, cow tay, pig fip, pig fip
s1: cow fip, cow tay, pig fip, pig tay 0.5 1 1 1
s2: cow fip, cow tay, pig fip, pig tay
S = 2 S = 5 S = 10
0
1/3
2/3
1
(a)
(b)
(c)
0
1/3
2/3
1
sing
lem
ultip
le
0 1 2 3 4 5 0 1 2 3 4 5 0 1 2 3 4 5
0 1 2 3 4 5 0 1 2 3 4 5 0 1 2 3 4 5
0 1 2 3 4 5 0 1 2 3 4 5 0 1 2 3 4 5
0
1/2
1
H(M
arke
r)
single
multiple
0
1/2
1
H(M
arke
r|N
oun)
p(m
1)
generation generationgeneration
S = 2 S = 5 S = 10
S = 2 S = 5 S = 10
Figure 2. Simulation results. (a) Proportion of plurals marked using m1. Each line shows an individual chain, for 20 simulation runs. In single-person chains, we seegreater divergence between chains, with some chains converging on always or seldom using m1; by contrast, in multiple-person chains, particularly for larger S, theinitial level of variability is retained longer. (b) Entropy of plural marking, averaged over 100 runs, error bars indicate 95% CIs. This overall measure of variabilityshows the trend visible in (a) more clearly: while variability is gradually lost over generations in all conditions (as indicated by reducing H(Marker)), this loss ofvariability is slower in multiple-person chains, particularly with larger S. (c) Conditional entropy of plural marking given the noun being marked, averaged over thesame 100 runs. While we reliably see the emergence of conditioned variation in single-person chains, as indicated by reducing H(MarkerjNoun), this process isslowed in multiple-person chains, particularly for larger S.
rstb.royalsocietypublishing.orgPhil.Trans.R.Soc.B
372:20160051
6
on June 13, 2018http://rstb.royalsocietypublishing.org/Downloaded from
Figure 3. Simulation results for multiple-person chains, where we manipulate whether learners can access and exploit speaker identity during learning. Results forsingle-person chains are shown for reference, all results are averaged over 100 simulation runs. The pair of plots on the left shows H(MarkerjNoun), i.e. conditionalentropy of plural marking given the noun being marked, averaged across the population—this captures the extent to which there is a conditioned system ofvariability that is shared and therefore independent of speaker. There is no effect of the speaker identity manipulation here, and learning from multiplepeople slows the development of a population-wide system of conditioned variation. The right pair of plots shows H(MarkerjNoun, Speaker), i.e. conditional entropyof plural marking, given the noun being marked and the speaker—this measure captures the development of speaker-specific systems of conditioned variation,where each speaker conditions their marker use on the noun being marked, but allows that different speakers might use different systems of conditioning. Here, wesee an effect of the speaker identity manipulation: in the no speaker identity models, these results essentially mirror those for H(MarkerjNoun), i.e. no conditioningdevelops due to mixing of data from multiple speakers; however, in the model where learners can track speaker identity, speaker-specific systems of conditioningemerge, as indicated by smoothly reducing H(MarkerjNoun, Speaker).
rstb.royalsocietypublishing.orgPhil.Trans.R.Soc.B
372:20160051
8
on June 13, 2018http://rstb.royalsocietypublishing.org/Downloaded from
H(MarkerjNoun, Speaker) serves as a diagnostic of whether or
not learners attend to speaker identity when learning and
reproducing variable systems. More generally, this difference
in the behaviour of the model based on speaker identity
shows that the cumulative regularization effect is not only
modulated by the number of individuals a learner learns
from, but also how they handle that input (do they track var-
iant use by individuals, or simply for the whole population?)
and how they derive their own output behaviour from their
input (do they model a single individual from their input, or
produce output which reflects the usage across all their input?).
The model therefore shows that regularization of variation
does not automatically ensue from learning and transmission,
even in situations where learners have biases in favour of regu-
larity. Smith & Wonnacott [19] showed that weak biases at the
individual level can accumulate and therefore be unmasked by
iterated learning; this model shows that the reverse is also poss-
ible, and that biases in learning can be masked by the dynamics
of transmission in populations. Note that this is true even if
learners have much stronger biases for regularity than that
we used here (see endnote 1). In other words, the slowing
effects of learning from multiple speakers apply even if indi-
vidual learners make large reductions to the variability of
their input (as child learners might: [4,5]). We return to the
implications of this point in the general discussion.
(d) Learning from multiple people: experimentThe model outlined in the previous section makes two pre-
dictions. First, learning from multiple individuals will slow
the cumulative conditioning seen in Smith & Wonnacott
[19]. Second, the degree of slowing will be modulated by
the extent to which learners are able to attend to (or choose
to attend to) speaker identity when tracking variability. We
test these predictions with human learners, using a paradigm
based closely on Smith & Wonnacott [19].
(i) MethodsParticipants. 150 native English speakers (112 female, 38 male,
mean age 21 years) were recruited from the University of
Edinburgh’s Student and Graduate Employment Service
and via emails to undergraduate students. Participants were
paid £3 for their participation, which took approximately
20 min.
Procedure. Participants worked through a computer pro-
gram, which presented and tested them on a semi-artificial
language. The language was text-based: participants
observed objects and text displayed on the monitor and
entered their responses using the keyboard. Participants
progressed through a three-stage training and testing regime:
1) Noun familiarization. Participants viewed pictures of three
cartoon animals (cow, pig and dog) along with English
nouns (e.g. ‘cow’). Each presentation lasted 2 s, after which
the text disappeared and participants were instructed to
retype that text. Participants then viewed each picture a
second time, without accompanying text, and were asked
to provide the appropriate label.
2) Sentence training. Participants were exposed to sentences
paired with pictures. Pictures showed either single animals
Figure 4. Experimental data. (a,b) Proportion of plurals marked using the marker that was initially in the majority in each chain. Each line shows an individualchain. For two-person chains, filled shapes indicate speaker ID provided, hollow shapes indicate no speaker ID. (c) Total entropy of the languages, as indicated byH(Marker), averaged over all chains (error bars indicate 95% CIs). Overall variability declines only slowly in both conditions: while some chains converge on always ornever marking plurals with the majority initial marker, most retain variability in plural marking. (d ) H(MarkerjNoun), i.e. conditional entropy of marker choice givennoun. In one-person chains, conditioned systems of variability rapidly emerge, as indicated by reducing H(MarkerjNoun); as predicted by the model, this devel-opment of conditioned variation is slowed in two-person chains. (e) H(MarkerjNoun, Speaker), i.e. conditional entropy of marker choice given the noun beingmarked and the speaker, for two-person chains only, split according to whether participants were provided with information on speaker identity. Contrary tothe predictions of the model, providing speaker identity makes no difference to the development of speaker-specific conditioning of variation.
rstb.royalsocietypublishing.orgPhil.Trans.R.Soc.B
372:20160051
10
on June 13, 2018http://rstb.royalsocietypublishing.org/Downloaded from
0.106, p ¼ 0.392) and a marginal interaction between con-
dition and generation (b ¼ 20.199, SE ¼ 0.099, p ¼ 0.044)
reflecting a tendency for two-person chains to underuse the
on June 13, 2018http://rstb.royalsocietypublishing.org/Downloaded from
review of relevant modelling work). This work finds that reci-
procal priming during interaction leads to convergence within
pairs of participants, typically on a system lacking unpredict-
able variation. Intriguingly, there is also increased regularity
in participants who thought they were interacting with another
human participant, but were in fact interacting with a computer
that used the variants in their trained proportion and was not
primed by the participants’ productions [34]. Regularization
here cannot be due to priming (as priming of the participant
by the computer interlocutor should keep them highly vari-
able), nor can it be due to reciprocal priming (as the computer
is not primed by the participant); it must therefore reflect an
(intentional or unintentional) strategic reduction in unpredict-
able variation promoted by the communicative context,
consistent with the data from Perfors [16]. There are, however,
subtle differences between the kind of regularization we
see in this pseudo-interaction and the regularization we see in
genuine interaction. Reciprocal priming in genuine interaction
leads to some pairs converging on a system that entirely
lacks variation, whereas pseudo-interacting participants never
became entirely regular.
Finally, there are inherent asymmetries in priming that may
serve to ‘lock in’ conditioned or regular systems at the expense
of unpredictable variability [33,35]. While variable users can
accommodate to a categorical partner by increasing their fre-
quency of usage, categorical users tend not to accommodate
to their variable partners by becoming variable: consequently,
once a grammatical marker reaches a critical threshold in a
population such that at least some individuals are categorical
users, alignment during interaction should drive the population
towards uniform categorical marker use, as variable users align
to a growing group of categorical users. This, in combination
with the regularizing effects of communicative task framing
[16,34] and reciprocal priming [34] suggests that interaction
may be a powerful mechanism for reducing unpredictable vari-
ation, which might play a role in explaining the scarcity of truly
unpredictable variation in natural language.
5. ConclusionThe structure of languages should be influenced by biases in stat-
istical learning, because languages persist by being repeatedly
learnt, and linguistic universals may therefore reflect biases in
learning. But the mapping from learning biases to language
structure is not necessarily simple. Weak biases can have
strong effects on language structure as they accumulate over
repeated transmission. At least in some cases, the opposite can
also be true: strong biases can have weak or no effects.
Furthermore, learning biases are not the only pressure acting
on languages: language use can produce effects that can (but
need not) resemble the effects produced by learning biases,
but which might have subtly or radically different causes. Com-
bining data and insights from studies of learning, transmission
and use is therefore essential if we are to understand how
biases in statistical learning interact with language transmission
and language use to shape the structural properties of language.
We have used the learning of unpredictable variation as a test
case here, but the same arguments should apply to other linguis-
tic features: statistical learning papers frequently make
inferences about the relationship between biases in statistical
learning and features of language design based on studies of
learning in individuals, but in the absence of a detailed
understanding of how biases in learning interact with use and
transmission, these inferences should be treated with caution.
In our opinion, the literature on unpredictable variation provides
a useful exemplar for how we should combine data from statisti-
cal learning, transmission and use in attempting to explain the
universal properties of human languages.
Ethics. The experiment detailed in §3.4 was approved by the Linguis-tics and English Language Ethics Committee, School of Philosophy,Psychology and Language Sciences, University of Edinburgh. Allparticipants provided informed consent prior to participating.
Data accessibility. Raw data files for the experiment detailed in §3.4 areavailable at the Edinburgh DataShare repository of the Universityof Edinburgh: http://dx.doi.org/10.7488/ds/1462.
Authors’ contributions. K.Sm., A.P. and E.W. designed the model andexperimental study. K.Sm. implemented the model, and K.Sw. col-lected all experimental data. K.Sm., A.P., O.F., A.S. and E.Wcontributed to the drafting and revision of the article, and all authorsapproved the final version.
Competing interests. We have no competing interests.
Funding. This work was supported by the Economic and SocialResearch Council (grant number ES/K006339, held by K.Sm. andE.W), and by ARC Discovery Project DP150103280, held by A.P.
Endnotes1Similar results are obtained if we use a stronger bias in favour ofregularization, e.g. setting a to 0.1 or 0.001—in general, we expectour results to hold as long as learners have some bias for regularity,but that bias is not so strong that it completely overwhelms their data.2See Burkett & Griffiths [23] for further discussion and results forBayesian iterated learning in populations.3Because in the models and experiments we present here, all nounsoccur equally frequently, we can simply calculate entropy by nounand then average (yielding the term 1=O in the expressionabove)—if nouns differed in their frequency, then the by-noun entro-pies would be weighted proportional to their frequency, rather than asimple average. Similarly, in the expression for H(MarkerjNoun,Speaker), we exploit the fact that all speakers are represented equallyfrequently in the models and experiments presented here.4In all conditions, to convert test output from the generation g partici-pant into training input for generation g þ 1, for a given scene, wesimply inspected whether the generation g participant used fip, tay orno marker, and used this marking when training participant n þ 1.In situations where the marker was mistyped, we treated it as if the par-ticipant had produced the closest marker to the typed string, based onstring edit distance (e.g. ‘tip’ treated as ‘fip’). Errors in the verb or nounused were not passed on to the next participant.5A logit regression on the data from the two-person condition, withthe presence/absence of speaker ID, shows no significant effect onmajority marker use of speaker ID or the interaction between speakerID and generation, p . 0.14.6Again, an analysis of the effect of the speaker identity manipulationin the two-person data reveals no effect on H(Marker) of speaker IDand no interaction between speaker ID and generation, p . 0.835.7Again, including speaker identity as a predictor indicates no effecton H(MarkerjNoun) of speaker identity and no interaction withgeneration, p . 0.918.8There is weak evidence in our data that our learners are somewhat sen-sitive to speaker-based conditioning in their input, although clearly notsufficiently sensitive to trigger the effects predicted by the model formaximally identity-sensitive learners. For instance, there is a positivebut non-significant positive correlation (r¼0.225, p¼0.28) between thedegree of speaker-based conditioning in participants’ input and theiroutput in the two-person, speaker identity experimental data (wherewe measure this correlation on mutual information of marker use,noun and speaker, which is HðMarkerÞ �HðMarkerjNoun, SpeakerÞ,to control for spurious correlations arising for overall levels of variabil-ity). Samara et al. [12], covered in the discussion, provide better data onthis issue.
on June 13, 2018http://rstb.royalsocietypublishing.org/Downloaded from
References
rstb.royalsocietypublishing.orgPhil.Trans.R.Soc.B
372:20160051
1. Christiansen MH, Chater N. 2008 Language asshaped by the brain. Behav. Brain Sci. 31,489 – 509. (doi:10.1017/S0140525X08004998).
2. Stuart-Smith J. 1999 Glottals past and present: astudy of T-glottalling in Glaswegian. Leeds Stud.Engl. 30, 181 – 204.
3. Givon T. 1985 Function, structure, and languageacquisition. In The crosslinguistic study of languageacquisition, vol. 2 (ed. D Slobin), pp. 1005 – 1028.Hillsdale, NJ: Lawrence Erlbaum.
4. Hudson Kam CL, Newport EL. 2005 Regularizingunpredictable variation: the roles of adult and childlearners in language formation and change. Lang.Learn. Dev. 1, 151 – 95. (doi:10.1080/15475441.2005.9684215)
5. Hudson Kam CL, Newport EL. 2009 Getting it rightby getting it wrong: when learners changelanguages. Cognit. Psychol. 59, 30 – 66. (doi:10.1016/j.cogpsych.2009.01.001)
6. Hudson Kam CL, Chang A. 2009 Investigating thecause of language regularization in adults: memoryconstraints or learning effects? J. Exp. Psychol.Learn. Mem. Cogn. 35, 815 – 821. (doi:10.1037/a0015097)
7. West RF, Stanovich KE. 2003 Is probability matchingsmart? Associations between probabilistic choicesand cognitive ability. Mem. Cognit. 31, 243 – 251.(doi:10.3758/BF03194383)
8. Perfors A. 2012 When do memory limitations leadto regularization? An experimental andcomputational investigation. J. Memory Lang. 67,486 – 506. (doi:10.1016/j.jml.2012.07.009)
9. Ferdinand V, Thompson B, Kirby S, Smith K. 2013Regularization behavior in a non-linguistic domain.In Proceedings of the 35th Annual Conference of theCognitive Science Society (eds M Knauff, M Pauen,N Sebanz, I Wachsmuth), pp. 436 – 441. Austin, TX:Cognitive Science Society.
10. Ferdinand V, Kirby S, Smith K. In preparation. Thecognitive roots of regularization in language.
11. Hudson Kam CL. 2015 The impact of conditioningvariables on the acquisition of variation in adult andchild learners. Language 91, 906 – 937. (doi:10.1353/lan.2015.0051)
12. Samara A, Smith K, Brown H, Wonnacott E.Submitted. Acquiring variation in an artificiallanguage: children and adults are sensitive tosocially-conditioned linguistic variation.
13. Wonnacott E. 2011 Balancing generalization andlexical conservatism: an artificial language study
with child learners. J. Memory Lang. 65, 1 – 14.(doi:10.1016/j.jml.2011.03.001)
14. Culbertson J, Smolensky P, Legendre G. 2012Learning biases predict a word order universal.Cognition 122, 306 – 29. (doi:10.1016/j.cognition.2011.10.017)
15. Culbertson J, Newport EL. 2015 Harmonic biases inchild learners: in support of language universals.Cognition 139, 71 – 82. (doi:10.1016/j.cognition.2015.02.007)
16. Perfors A. 2016 Adult regularization of inconsistentinput depends on pragmatic factors. Lang. Learn.Dev. 12, 138 – 155. (doi:10.1080/15475441.2015.1052449)
17. Kirby S, Dowman M, Griffiths TL. 2007 Innatenessand culture in the evolution of language. Proc. NatlAcad. Sci. USA 104, 5241 – 5245. (doi:10.1073/pnas.0608222104)
18. Reali F, Griffiths TL. 2009 The evolution of frequencydistributions: relating regularization to inductivebiases through iterated learning. Cognition 111,317 – 328. (doi:10.1016/j.cognition.2009.02.012)
19. Smith K, Wonnacott E. 2010 Eliminatingunpredictable variation through iterated learning.Cognition 116, 444 – 449. (doi:10.1016/j.cognition.2010.06.004)
20. Ramscar M, Yarlett D. 2007 Linguistic self-correctionin the absence of feedback: a new approach to thelogical problem of language acquisition. Cogn. Sci.31, 927 – 60. (doi:10.1080/03640210701703576)
21. Ramscar M, Gitcho N. 2007 Developmental changeand the nature of learning in childhood. Trends Cogn.Sci. 11, 274 – 279. (doi:10.1016/j.tics.2007.05.007)
22. Rische JL, Komarova NL. 2016 Regularization oflanguages by adults and children: amathematicalframework. Cogn. Psychol. 84, 1 – 30. (doi:10.1016/j.cogpsych.2015.10.001)
23. Burkett D, Griffiths TL. 2010 Iterated learning ofmultiple languages from multiple teachers. In Theevolution of language: proceedings of the 8thinternational conference (EVOLANG 8) (eds ADMSmith, M Schouwstra, B de Boer, K Smith), pp. 58 –65. Singapore: Word Scientific.
24. Labov W. 1966 The social stratification of English inNew York City. Washington, DC: Center for AppliedLinguistics.
25. Labov W. 2001 Principles of linguistic change: socialfactors. Oxford, UK: Blackwell.
26. Neu H. 1980 Ranking of constraints on /t,d/deletion in american english: astatistical analysis. In
Locating language in time and space (ed. W Labov),pp. 37 – 54. New York, NY: Academic Press.
27. Trudgill P. 1974 The social differentiation of Englishin Norwich. Cambridge, UK: Cambridge UniversityPress.
28. Eckert P, McConnell-Ginet S. 1999 Newgeneralizations and explanations in language andgender research. Lang. Soc. 28, 185 – 201. (doi:10.1017/S0047404599002031)
29. Cameron D. 2005 Language, gender, and sexuality:current issues and new directions. Appl. Ling. 26,482 – 502. (doi:10.1093/applin/ami027)
30. Pickering MJ, Garrod S. 2004 Toward a mechanisticpsychology of dialogue. Behav. Brain Sci. 27,169 – 225. (doi:10.1017/S0140525X04000056)
31. Horn L. 1984 Toward a new taxonomy for pragmaticinference: Q-based and r-based implicature. InMeaning, form, and use in context: linguisticapplications (ed. D Schiffrin), pp. 11 – 42.Washington, DC: Georgetown University Press.
32. Clark E. 1988 On the logic of contrast. J. Child Lang.15, 317 – 335. (doi:10.1017/S0305000900012393)
33. Smith K, Feher O, Ritt N. 2014 Eliminatingunpredictable linguistic variation throughinteraction. In Proceedings of the 36th AnnualConference of the Cognitive Science Society (eds PBello, M Guarini, M McShane, B Scassellati), pp.1461 – 1466. Austin, TX: Cognitive Science Society.
34. Feher O, Wonnacott E, Smith K. Structural primingin artificial languages and the regularisation ofunpredictable variation. J. Memory Lang. 91, 158 –180.
35. Feher O, Ritt N, Smith K. In preparation. Eliminatingunpredictable linguistic variation throughinteraction.
36. Galantucci B. 2005 An experimental study of theemergence of human communication systems.Cogn. Sci. 29, 737 – 67. (doi:10.1207/s15516709cog0000_34)
37. Fay N, Garrod S, Roberts LL, Swoboda N. 2010 Theinteractive evolution of human communicationsystems. Cogn. Sci. 34, 351 – 386. (doi:10.1111/j.1551-6709.2009.01090.x)
38. Tamariz M, Ellison TM, Barr DJ, Fay N. 2014 Culturalselection drives theevolution of humancommunication systems. Proc. R. Soc. B 281,20140488. (doi:10.1098/rspb.2014.0488)
39. Steels L. 2011 Modeling the cultural evolution oflanguage. Phys. Life Rev. 8, 339 – 356. (doi:10.1016/j.plrev.2011.10.014)