Person of Interest: Experimental
Investigations into the
Learnability of Person Systems
Mora Maldonado
Jennifer Culbertson
Person systems convey the roles entities play in the context of speech
(例如, speaker, addressee). As with other linguistic category systems,
not all ways of partitioning the person space are equally likely crosslin-
guistically. Different theories have been proposed to constrain the set
of possible person partitions that humans can represent, explaining
their typological distribution. This article introduces an artificial lan-
guage learning methodology to investigate the existence of universal
constraints on person systems. We report the results of three experi-
ments that inform these theoretical approaches by generating behav-
ioral evidence for the impact of constraints on the learnability of differ-
ent person partitions. Our findings constitute the first experimental
evidence for learnability differences in this domain.
关键词: person systems, pronouns, artificial language learning, 林-
guistic universals, 语义学
1 介绍
Person systems—exemplified by pronominal paradigms (例如, 我, 你, 她)—categorize entities
as a function of their role in the context of speech: the speaker, the addressee, 和别的, WHO
play no active role in the conversation. As in other semantic domains, it has long been observed
that person systems exhibit what appears to be constrained variation across languages: 一些
person systems are frequent, while others are rare or do not occur at all (Cysouw 2003, Baerman,
棕色的, and Corbett 2005).1
This article has profited from the comments and suggestions of Daniel Harbour and Peter Ackema, as well as of
two anonymous LI reviewers. We also wish to thank the audiences of The Alphabet of Universal Grammar at the British
Academy for interesting discussion and Estudio Chirrikenstein for the beautiful illustrations. This project has received
funding from the European Research Council (ERC) under the European Union’s Horizon 2020 research and innovation
程序 (grant agreement no. 757643).
1 This kind of typological tendencies has been seen in other semantic domains, involving both content and logical
字. 例如, crosslinguistic regularities have been argued to provide evidence for a universal basis for color
分类, reflecting properties of the human perceptual system (Kay and Regier 2007, Gibson et al. 2017, Zaslavsky
等人. 2019). Similar arguments have been made to explain the distribution of kinship systems across languages (肯普
and Regier 2012). 相关地, the study of logical words has also revealed that connectives and quantifiers found in natural
languages only cover a very small subset of all possible meanings, indicating the existence of semantic universals (Barwise
and Cooper 1981, 匹安多糖, Tenenbaum, and Goodman 2016, Chemla, Buccola, and Dautriche 2019, Steinert-Threlkeld
and Szymanik to appear).
Linguistic Inquiry, 体积 53, 数字 2, 春天 2022
295–336
(西德:2) 2020 by the Massachusetts Institute of Technology. Published under
a Creative Commons Attributions 4.0 国际的 (抄送 4.0) 执照.
https://doi.org/10.1162/ling_a_00406
295
我
D
哦
w
n
哦
A
d
e
d
F
r
哦
米
H
t
t
p
:
/
/
d
我
r
e
C
t
.
米
我
t
.
e
d
你
/
我
我
/
n
G
A
r
t
我
C
e
–
p
d
我
F
/
/
/
/
5
3
2
2
9
5
2
0
1
4
2
4
4
/
我
我
n
G
_
A
_
0
0
4
0
6
p
d
.
F
乙
y
G
你
e
s
t
t
哦
n
0
7
S
e
p
e
米
乙
e
r
2
0
2
3
296
M O R A M A L D O N A D O A N D J E N N I F E R C U L B E R T S O N
Typological regularities of this sort have led linguists to propose universal constraints on
possible person systems (Silverstein 1976, Ingram 1978, Noyer 1992, Harley and Ritter 2002,
Bobaljik 2008, Harbour 2016, Ackema and Neeleman 2018). Such constraints are often conceived
的 (either implicitly or explicitly) as reflecting characteristics of human linguistic capacity that
have consequences for learning: 具体来说, they are assumed to delimit the space of hypotheses
entertained by the learner (Chomsky 1965; but see, 例如, Perfors, Tenenbaum, and Regier 2011).
然而, while person has been extensively investigated from formal and typological perspec-
特维斯, the link between hypothesized universal constraints on person systems and learnability re-
mains largely unexplored (though see Nevins, Rodrigues, and Tang 2015 for an artificial learning
方法, 和布朗 1997, Hanson 2000, Hanson, Harley, and Ritter 2000, Moyer et al. 2015
on acquisition). 在本文中, we introduce an artificial language learning methodology to investi-
gate the existence of universal constraints on person systems. 第一的, we summarize some additional
theoretical background from which we will derive a number of specific predictions regarding the
learnability of these systems.
1.1 The Person Space
As mentioned above, there are three conversational roles typically delimited in the person space:
the speaker, the addressee, 和别的, who are not active participants in the conversation. Follow-
ing standard assumptions, we represent them as i, 你, and o, 分别 (例如, Harbour 2016).
From this ontology, we obtain seven logically possible person categories or “metapersons”: 我, io,
iu, iuo, 你, uo, 哦 (Sokolovskaja 1980, Bobaljik 2008, Sonnaert 2018). Research on the typological
distribution of person systems, 然而, has found evidence that only four of these are grammati-
calized as person categories: first exclusive (我), first inclusive (iu), 第二 (你), 第三个 (哦).
This asymmetry can be directly captured by assuming that the speaker and addressee are
unique—there are no forms that express uniquely multiple speakers, or uniquely multiple addres-
sees—but there can be an undefined number of others (following Harbour 2016).2 The meanings
expressed by the unattested combinations (io, iuo, uo) can be captured as the interaction between
person and number; the four core person categories can each be pluralized by adding extra others
(see footnote 2; Boas 1911). 桌子 1 illustrates these person and number categories (数字
expressed by the presence or absence of the subscript o). To account for the above-mentioned
restrictions on person categories, theories of the person space have traditionally posited two
primitive binary features: [(西德:3)speaker] 和 [(西德:3)addressee] (or other equivalent notations; Silverstein
2 This assumption is not trivial. As soon as multiple speakers and addressees are allowed in the ontology, each
logically available combination of the three entities—i, io, iu, iuo, 你, uo, o—should count as a possible person category
independent of number. 例如, one form would refer to the speaker alone (我) and another form to the speaker plus
someone else (io), each with a plural alternative (ii vs. ioo). 然而, this contrast is never grammaticalized: no language
distinguishes between plural expressions referring to multiple speakers/addressees and expressions referring to the speaker/
addressee plus others. 的确, this has been formulated as a typological universal: pluralities containing participants
(speakers or addressees) are never formally distinguished from pluralities containing others. This universal has been
extensively discussed in the literature on person systems (Greenberg 1988, Cysouw 2003, Bobaljik 2008, Wechsler 2010),
but is not directly investigated here.
我
D
哦
w
n
哦
A
d
e
d
F
r
哦
米
H
t
t
p
:
/
/
d
我
r
e
C
t
.
米
我
t
.
e
d
你
/
我
我
/
n
G
A
r
t
我
C
e
–
p
d
我
F
/
/
/
/
5
3
2
2
9
5
2
0
1
4
2
4
4
/
我
我
n
G
_
A
_
0
0
4
0
6
p
d
.
F
乙
y
G
你
e
s
t
t
哦
n
0
7
S
e
p
e
米
乙
e
r
2
0
2
3
P E R S O N O F I N T E R E S T
297
桌子 1
Four-person system
Attested
categories Binary-features account
1 EXCL
1 INCL
2
3
io
iuo
uo
oo
[(西德:5)speaker (西德:6)addressee]
[(西德:5)speaker (西德:5)addressee]
[(西德:6)speaker (西德:5)addressee]
[(西德:6)speaker (西德:6)addressee]
桌子 2
Example personal pronoun systems
Ilocano
MIN
AUG
Mandarin
PL
SG
英语
SG
1 EXCL
1 INCL
2
3
共
ta
mo
已经
mi
tayo
yo
和
wo
女性
— zamen
nimen
ni
tamen
ta
我
—
你
she/he/they
PL
我们
我们
你
他们
1976, Ingram 1978, Noyer 1992, Bobaljik 2008). As shown in table 1, the interaction between
these two binary features predicts all and only the four attested categories.
An example of a language with a four-way person distinction in its pronominal system is
Mandarin. As shown in table 2, each person category is expressed by a different pronoun, 和
additional morphology marking whether the referent is singular or plural. This system has seven
forms total, since the inclusive is inherently nonsingular (it necessarily refers to both the speaker
and the addressee) and thus always features plural morphology in Mandarin. 另一个例子
is Ilocano, which differs from Mandarin in that it makes a minimal/augmented number distinction
rather than singular/plural. This system has eight forms, including two distinct inclusive forms,
one minimal (ta (西德:4) speaker and addressee only) and the other augmented (tayo (西德:4) speaker,
addressee, 和别的).
In many other languages, the meaning space is partitioned such that not all possible person
and/or number categories are expressed by distinct forms. Such languages exhibit homophony.
例如, inclusive languages differ from noninclusive languages (terminology from Daniel
2005) like English, where there is homophony between the first and inclusive persons (此外
to homophony between second person singular and plural; 见表 2). Feature-based accounts
of person often derive a restricted set of partitions of the person and number space as defined by
the presence or absence of homophony among cells in the space. Such theories only derive
homophony patterns by contrast neutralization or underspecification: a distinction that is made
available by the grammar might not be active in a specific language (Halle and Marantz 1994,
Harbour 2008, Harley 2008, Pertsova 2011). 具体来说, the set of features in table 1 straightfor-
wardly derives three person homophony patterns on the basis of which contrasts are left underspec-
我
D
哦
w
n
哦
A
d
e
d
F
r
哦
米
H
t
t
p
:
/
/
d
我
r
e
C
t
.
米
我
t
.
e
d
你
/
我
我
/
n
G
A
r
t
我
C
e
–
p
d
我
F
/
/
/
/
5
3
2
2
9
5
2
0
1
4
2
4
4
/
我
我
n
G
_
A
_
0
0
4
0
6
p
d
.
F
乙
y
G
你
e
s
t
t
哦
n
0
7
S
e
p
e
米
乙
e
r
2
0
2
3
298
M O R A M A L D O N A D O A N D J E N N I F E R C U L B E R T S O N
ified ([(西德:3)speaker], [(西德:3)addressee], 或两者). 例如, neutralizing the [(西德:3)addressee] feature would
generate syncretism between 1 EXCL and 1 INCL categories (grouped as [(西德:5)speaker]), on the one
手, 和之间 2 和 3 (grouped as [(西德:6)speaker]), 在另一. Other feature-based homophony
patterns can also be derived from this system by restricting underspecification to specific natural
类. 例如, the aforementioned clusivity distinction is lost when two meanings that
share the feature [(西德:5)speaker] (i and iu) become indistinguishable (例如, 英语). 那是, 这
grouping of 1 EXCL and 1 INCL categories relies on them belonging to the same natural class
[(西德:5)speaker]. These kinds of feature-based patterns are often referred to as systematic homophony.
然而, many different homophony patterns have been documented both within and across
语言 (Zwicky 1977, Corbett, 棕色的, and Baerman 2002, Cysouw 2003, Baerman, 棕色的,
and Corbett 2005, Baerman and Brown 2013), not all of which can be derived by feature neutraliza-
的. 在某些情况下, two or more meanings that do not share any feature are nevertheless expressed
by the same form in a given language. This so-called accidental or random homophony is therefore
not described in terms of contrast neutralization, as the targeted meanings do not belong to the
same natural class (例如, defined by the features in table 1). Partitions that are not derivable by
a theory are often assumed to arise through historical accident, target mainly individual paradigms
in a given language, and may be marginal typologically (Halle and Marantz 1994, Pertsova 2011,
Sauerland and Bobaljik 2013; but see Cysouw 2003). But whether the typological evidence accords
with this prediction is not always clear.3
While estimating the frequency distribution over partitions of the person space is complex,
a number of theories have been developed to make more fine-grained predictions about possible
homophony patterns. Harley and Ritter (2002) put forward a universal feature geometry for person
based on three privative features, as illustrated in figure 1. This derives the same set of four
person categories as the binary-features account in table 1 but also establishes hierarchical relations
它们之间, in order to make more accurate predictions about the typological frequency of
homophony patterns (see also Be´jar 2003, 麦金尼斯 2005, Cowper and Hall 2009 for similar
方法). 例如, this system derives homophony between 1 EXCL, 1 INCL, 和 2 categor-
ies and therefore predicts this to arise systematically.
与此相类似, Harbour (2016) posits a theory specifically designed to capture the robust
typological generalization in (1), also known as Zwicky’s observation (see Zwicky 1977:
717–719).
3 As it turns out, determining the crosslinguistic frequency of different partitions of the person (and number) 空间
is not straightforward. It crucially depends on the specific assumptions made about how to count different systems. 一些
authors (Cysouw 2003, Sauerland and Bobaljik 2013) include individual person-marking paradigms (例如, verbal agreement
and pronominal systems) within a single language. Using this metric for counting, the frequency distribution across
languages is extremely skewed: in Cysouw’s dataset, 的 4,140 possible partitions of an eight-cell person/number
空间, 仅有的 61 are attested (calculated over 265 paradigms). A slightly different approach that also counts paradigms is
given by Baerman, 棕色的, and Corbett (2005) and Baerman and Brown (2013). Still another possibility is to count on
the basis of an abstract notion of person partition, across all the paradigms in a given language. 例如, Harbour
(2016) proposes a superposition technique: a distinction is neutralized in a language if and only if it is inactive across
all paradigms. Under this superposition approach, 有 15 possible partitions of a four-cell person space, 但仅 5
are attested typologically (with one marginal exception; see appendix A2.1 in Harbour 2016).
我
D
哦
w
n
哦
A
d
e
d
F
r
哦
米
H
t
t
p
:
/
/
d
我
r
e
C
t
.
米
我
t
.
e
d
你
/
我
我
/
n
G
A
r
t
我
C
e
–
p
d
我
F
/
/
/
/
5
3
2
2
9
5
2
0
1
4
2
4
4
/
我
我
n
G
_
A
_
0
0
4
0
6
p
d
.
F
乙
y
G
你
e
s
t
t
哦
n
0
7
S
e
p
e
米
乙
e
r
2
0
2
3
P E R S O N O F I N T E R E S T
299
Referring Expression
Participant
Addressee
Speaker
数字 1
Feature geometry account in Harley and Ritter 2002. The underlined daughter node Speaker represents the
default interpretation of the bare Participant node. Languages may or may not allow both dependent nodes
to be specified together. If they do, the inclusive category is obtained when Speaker and Addressee are
present simultaneously.
(1) Languages that do not have a dedicated phonological form for an inclusive person (‘you
and us’) always assimilate it into the first plural person (‘us’) and never into the second
(‘you’) or third (‘them’).
Harbour makes use of two binary features, [(西德:3)作者] 和 [(西德:3)参与者]. While the features them-
selves denote lattices containing referential entities (我, 你, iu, 哦), the values of the features are
modeled as complementary operations on lattices. Because features are similar to functions, lan-
guages can differ not only in which features are active, but also in the order of feature composition.
A similar approach is taken by Ackema and Neeleman (2018; see also Ackema and Neeleman
2013), with the main goal of accounting for the typological tendency described in Baerman,
棕色的, and Corbett 2005 and summarized in (2).4
(2) Languages that feature homophony between first (inclusive and exclusive) 第二个
plural pronouns (‘us’ (西德:4) ‘you’) and between second and third (‘you’ (西德:4) ‘them’) are far
more frequent than those instantiating first-third homophony.
Each of these approaches (which we will discuss in more detail below) introduces different
theoretical apparatus to capture these typological observations. 有, 然而, a number of
obvious limitations that make basing theories exclusively on typological evidence problematic.
1.2 From Typology to Learning
There is extensive literature now documenting (and in some cases proposing solutions to) 这
problems posed by typological data samples (for an excellent overview, see Cysouw 2005). 为了
4 As an anonymous reviewer points out, the approach taken by Baerman, 棕色的, and Corbett (2005) is a conservative
一: their typological counts are restricted to cases where whole forms are identical, disregarding morpheme syncretism.
This might result in the exclusion of morphemes that are syncretic but occur in different combinations. Given that person
features are often instantiated in morphemes rather than in whole words, the typological tendency stated in (2) 应该
be considered with some caution.
我
D
哦
w
n
哦
A
d
e
d
F
r
哦
米
H
t
t
p
:
/
/
d
我
r
e
C
t
.
米
我
t
.
e
d
你
/
我
我
/
n
G
A
r
t
我
C
e
–
p
d
我
F
/
/
/
/
5
3
2
2
9
5
2
0
1
4
2
4
4
/
我
我
n
G
_
A
_
0
0
4
0
6
p
d
.
F
乙
y
G
你
e
s
t
t
哦
n
0
7
S
e
p
e
米
乙
e
r
2
0
2
3
300
M O R A M A L D O N A D O A N D J E N N I F E R C U L B E R T S O N
一, such data are generally sparse, and in many cases the number of languages behind a given
typological generalization is quite small. 例如, the largest sample of person/number para-
digms, from Cysouw 2003, includes only around 200 语言. Sparse data lead to unreliable
estimates of relative frequency, particularly in the tail of the distribution. 例如, 它不是
possible to confidently conclude on the basis of small samples that a given partition is impossible
(例如, see Piantadosi and Gibson 2014, and also footnote 3).
而且, typological data are massively confounded: there are many factors that shape
typological distributions (例如, historical accidents, genetic relations between languages, facts
about diachrony; 看, 例如, Pagel, Atkinson, and Meade 2007, Bickel 2008, Cysouw 2010, Dunn
等人. 2011), only a subset of which are relevant for building theories of the generative capacity
of the linguistic system. The immediate consequence is that these data sources typically cannot
be used to argue for a causal link between the cognitive or linguistic system and particular features
语言的 (例如, see discussion in Culbertson 2012, Piantadosi and Gibson 2014, Ladd, 罗伯茨,
and Dediu 2015).
As a response to these general issues—which are relevant for typological data in any do-
main—there has been an increasing attempt to bring behavioral data on learning to bear on
linguistic theories. 具体来说, artificial language learning experiments have now been used to
link typological universals to human learning and inference in a number of domains including
phonology (例如, Wilson 2006, Moreton 2008, 白色的 2017, Martin and Peperkamp 2020), syntax
(例如, Culbertson, 斯摩棱斯基, and Legendre 2012, Tabullo et al. 2012, Culbertson and Adger
2014, Martin et al. 2019), morphology (例如, Fedzechkina, Jaeger, and Newport 2012, Saldana,
Oseki, and Culbertson 2019), and lexical categorization (Carstensen et al. 2015, Chemla, Buccola,
and Dautriche 2019); for reviews, see Culbertson 2012, to appear, Moreton and Pater 2012.
The present study uses artificial language learning experiments to test a set of predictions
derived from the feature-based theories of person described above. By incorporating this new
source of data, we can corroborate—or not—the universal constraints on person partitions hypoth-
esized on the basis of typological data. The article proceeds as follows. In Experiment 1, 我们
establish an experimental setup to test some basic assumptions of feature-based systems, 包括
whether systematic and random homophony are treated differently by learners acquiring a new
person system. In Experiment 2, we investigate whether the universal typological tendency known
as Zwicky’s observation is supported by a learnability advantage, as predicted by Harbour (2016).
最后, in Experiment 3 we explore potential asymmetries in the learnability of different partitions
of first, 第二, and third person categories, as predicted by different theories (例如, Harley and
里特尔 2002, Ackema and Neeleman 2018).
2 实验 1: Something about Us
2.1 介绍
While different theories of person have hypothesized different inventories of features to constrain
the person space, they all assume the features to be universal, 那是, part of the human linguistic
容量 (Harley and Ritter 2002, Bobaljik 2008, Harbour 2016, 除其他外). This as-
我
D
哦
w
n
哦
A
d
e
d
F
r
哦
米
H
t
t
p
:
/
/
d
我
r
e
C
t
.
米
我
t
.
e
d
你
/
我
我
/
n
G
A
r
t
我
C
e
–
p
d
我
F
/
/
/
/
5
3
2
2
9
5
2
0
1
4
2
4
4
/
我
我
n
G
_
A
_
0
0
4
0
6
p
d
.
F
乙
y
G
你
e
s
t
t
哦
n
0
7
S
e
p
e
米
乙
e
r
2
0
2
3
P E R S O N O F I N T E R E S T
301
sumption predicts that all things being equal, humans should have access to, and therefore be
able to learn, feature distinctions that are not at play in their native language (first prediction).
Feature-based theories also predict that learners should be sensitive to natural classes as defined
by feature structure: categories that share a feature should be more readily mapped onto the same
phonological form. These systematic homophony patterns are predicted to be (easily) learnable
(second prediction) in contrast to random homophony, where there is no featural basis for mean-
ings to share a form.
In Experiment 1, we targeted these two predictions by focusing on person categories that
involve the speaker (first exclusive and inclusive persons), and their interaction with number
特征. We investigated the contrasts that arise by the interaction of two binary features, 一
for person ([(西德:3)addressee]) and one for number ([(西德:3)最小的]) (see Noyer 1992, Bobaljik 2008,
Harbour 2014, for more developed accounts).5 这 [(西德:3)addressee] feature ensures a clusivity con-
特拉斯特: it distinguishes between groups of individuals that include both speaker and addressee (IE。,
iu, iuo) and those that exclude the addressee (IE。, 我, io). 这 [(西德:3)最小的] feature distinguishes the
minimal elements that satisfy each person category (IE。, 我, iu) from nonminimal (augmented)
pluralities, where the reference may include other(s) 还有 (IE。, io, iuo). The resulting four-cell
partition of this person-number space is given in table 3.
There are multiple partitions of this space as defined by homophony. We focus here on
bipartitions, where two pronominal forms cover the four-cell space. One possible bipartition uses
only the [(西德:3)addressee] distinction, thus contrasting exclusive and inclusive, but neutralizing number
(“Person-contrast” in table 4a). A second possible bipartition uses the [(西德:3)最小的] distinction,
resulting in one form for both minimal categories and another for both nonminimal (augmented)
类别 (“Number-contrast” in table 4b). 有趣的是, to make this particular distinction, 这
paradigm also relies on (or presupposes) 这 [(西德:3)addressee] person contrast: the individuals i
(speaker alone) and iu (speaker and addressee alone) are only “minimal” if the speaker-addressee
dyad (iu) is treated as a smallest element that satisfies the [(西德:5)addressee] feature. 那是, 这
[(西德:5)最小的] feature can only pick up the smallest element of the inclusive category if there is an
inclusive person category to begin with. 因此, we must assume that the [(西德:3)addressee] contrast is
active in this system, even if it is not encoded by different phonological forms. A schematic
桌子 3
Reduced person space
MIN
(西德:5)min
AUG
(西德:6)min
1 EXCL [((西德:5)sp)(西德:6)add]
1 INCL [((西德:5)sp)(西德:5)add]
我
iu
io
iuo
5 We are not committed to this specific inventory; our predictions would hold for any theory that posits the contrasts
他们自己, regardless of the structure of the feature space.
我
D
哦
w
n
哦
A
d
e
d
F
r
哦
米
H
t
t
p
:
/
/
d
我
r
e
C
t
.
米
我
t
.
e
d
你
/
我
我
/
n
G
A
r
t
我
C
e
–
p
d
我
F
/
/
/
/
5
3
2
2
9
5
2
0
1
4
2
4
4
/
我
我
n
G
_
A
_
0
0
4
0
6
p
d
.
F
乙
y
G
你
e
s
t
t
哦
n
0
7
S
e
p
e
米
乙
e
r
2
0
2
3
302
M O R A M A L D O N A D O A N D J E N N I F E R C U L B E R T S O N
桌子 4
Different partitions of a reduced person space
MIN
AUG
A. Person-contrast
(Number homophony)
乙. Number-contrast
(Person homophony)
C. Random
d. English-like
1 EXCL
1 INCL
1 EXCL
1 INCL
1 EXCL
1 INCL
1 EXCL
1 INCL
derivation of the bipartitions in table 4a and table 4b using [(西德:3)addressee] 和 [(西德:3)最小的] 特征
is given in the online appendix (figure A.1) (https://doi.org/10.1162/ling_a_00406).
Partitions with random, non-feature-based homophony are also possible. 例如, exclu-
sive minimal (我) and inclusive augmented (iuo) may share one form, and inclusive minimal (iu)
and exclusive augmented (io) 其他. This meaning-to-form mapping cannot be expressed in
terms of neutralizing a single semantic distinction (“Random-contrast” in table 4c). Put differently,
there is no natural class that groups only exclusive minimal and inclusive augmented. 注意
there are three other random partitions of this reduced person space.
Our experiment targeted English-speaking learners. In order to make clear learnability predic-
tions for this population, it is important to understand how noninclusive systems of first person
pronouns (as in English singular I vs. plural we) are typically derived in feature-based theories.
As observed above, inclusive systems may differ in whether they make a number distinction
within the inclusive: languages like Mandarin do not, whereas languages like Ilocano distinguish
pronouns referring to speaker and addressee alone from pronouns referring to speaker, addressee,
和别的 (见表 2). To account for this variation, theories of number morphology often
distinguish the classical singular/plural contrast, based on the [(西德:3)atomic] feature, from a minimal/
augmented one, based on the [(西德:3)最小的] distinction (Noyer 1992, Harbour 2011, 2014).6 Infor-
mally, 这 [(西德:3)atomic] feature picks up elements from the domain as a function of whether or not
they have proper parts, whereas the [(西德:3)最小的] feature picks out the smallest element(s) 在里面
domain. On their own, these two contrasts only yield different partitions for languages that distin-
6 Another argument that has been used to support the existence of both [(西德:3)atomic] 和 [(西德:3)最小的] contrasts comes
from languages that make a singular/dual/plural distinction, but this is not relevant for our purposes (Harbour 2014,
Martı´ 2020).
我
D
哦
w
n
哦
A
d
e
d
F
r
哦
米
H
t
t
p
:
/
/
d
我
r
e
C
t
.
米
我
t
.
e
d
你
/
我
我
/
n
G
A
r
t
我
C
e
–
p
d
我
F
/
/
/
/
5
3
2
2
9
5
2
0
1
4
2
4
4
/
我
我
n
G
_
A
_
0
0
4
0
6
p
d
.
F
乙
y
G
你
e
s
t
t
哦
n
0
7
S
e
p
e
米
乙
e
r
2
0
2
3
P E R S O N O F I N T E R E S T
303
guish an inclusive person category—namely, for systems where the smallest pronominal referent
might be nonatomic (IE。, iu). In noninclusive languages, 相比之下, the sets of minimal and
atomic pronominal referents are identical; 两个都 [(西德:3)atomic] 和 [(西德:3)最小的] contrasts will always
lead to the same results (see figure A.1 in the online appendix). 这里, we follow much previous
literature in assuming that minimal/augmented pronominal systems are necessarily inclusive, 和
we take languages like English, which make only a singular/plural distinction, to be based on
the simpler [(西德:3)atomic] 对比 (例如, Noyer 1992, Harbour 2014, 2016). Under these assumptions,
然后, 两者都 [(西德:3)最小的] 和 [(西德:3)addressee] features are nonnative for English speakers in
the context of first person partitions. 笔记, 然而, that even if languages like English were
described as making use of the [(西德:3)最小的] feature, this contrast alone would still be insufficient
to derive the bipartitions in table 4a–c, 作为 [(西德:3)addressee] feature is also required (see also the
discussion in section 2.4).7
In this experiment, we therefore test the two predictions laid out above: 第一的, that English
speakers should be able to learn contrasts (桌子 4) that are not directly instantiated in their
language but are broadly attested typologically—namely, the inclusive/exclusive distinction and
the minimal/augmented distinction (例如, in Ilocano; 托马斯 1955, Cysouw 2003); 第二, 那
these unfamiliar feature-based paradigms should be easier for English speakers to learn than
random homophony paradigms (table 4c). If these predictions are borne out, then we can conclude
that the person-number space is indeed based on a set of universal features, such as those posited
by the theories described above. We can then take the next step of testing different predictions
made by particular feature-based theories. As a sanity check, we also test whether English-speaking
learners are biased in favor of a person system that resembles their own, as in table 4d. If partici-
pants perceive the similarity between the new system they are learning and the English person-
number system, then this is a good indication that our experimental methodology is successfully
engaging the linguistic space we intend.
To test these predictions, we used two complementary artificial language learning paradigms.
In Experiment 1A, participants were taught a pronominal system that matched one of the four
paradigms in table 4, and were then tested on how accurately they were able to learn each pattern
(“ease of learning” paradigm; Culbertson, Gagliardi, 和史密斯 2017, Tabullo et al. 2012). 在
原则, two paradigms might be equally learnable (after some amount of exposure), 然而
one of them might still be preferred. In Experiment 1B, we therefore investigated differences in
the likelihood of inferring a given paradigm in the absence of explicit data (“poverty of the
stimulus” paradigm; Wilson 2006, Culbertson and Adger 2014). Participants were trained on only
two cells of the paradigm in table 3 and then had to use the forms they had learned to express
all the cells in the paradigm. 换句话说, they had to extrapolate the taught forms to the
remaining two categories. How they should extrapolate was ambiguous on the basis of their
训练, and how they did so indicated which underlying paradigm they had inferred. 例如,
7 Crucially for us, 这 [(西德:3)addressee] distinction is nonnative in the presence of the [(西德:5)speaker] feature (IE。, (西德:2)[(西德:5)speaker],
[(西德:3)addressee](西德:3) is nonnative). 当然, English speakers do have experience with this feature: it is used to distinguish
first person (collapsing inclusive and exclusive) from second person.
我
D
哦
w
n
哦
A
d
e
d
F
r
哦
米
H
t
t
p
:
/
/
d
我
r
e
C
t
.
米
我
t
.
e
d
你
/
我
我
/
n
G
A
r
t
我
C
e
–
p
d
我
F
/
/
/
/
5
3
2
2
9
5
2
0
1
4
2
4
4
/
我
我
n
G
_
A
_
0
0
4
0
6
p
d
.
F
乙
y
G
你
e
s
t
t
哦
n
0
7
S
e
p
e
米
乙
e
r
2
0
2
3
304
M O R A M A L D O N A D O A N D J E N N I F E R C U L B E R T S O N
a participant might be trained on two distinct forms for exclusive minimal (speaker only) 和
exclusive augmented (speaker plus others), and then tested on how they mapped these forms to
the two remaining categories including the addressee. If they used the augmented form for both
new categories, then they had inferred an English-like paradigm.
2.2 方法
Both experiments, including all hypotheses, 预测, and analyses, were preregistered: Experi-
ment 1A (https://osf.io/w8n25) and Experiment 1B (https://osf.io/j2rcn). Materials, 数据, 和
scripts are provided at https://osf.io/ca8yp/?view_only(西德:4)24c66919e3ff41cf8a01f7c328dead6e.
All analyses are as per the preregistration unless we say otherwise. These experiments were
implemented using the JavaScript library jsPsych (De Leeuw 2015) and presented to participants
in a web browser.
2.2.1 Design Participants in Experiment 1A were randomly assigned to one of four conditions:
English-like, Person-contrast, Number-contrast, and Random-contrast (见表 4). 参加者
in all conditions were taught two pronominal forms mapped into four person categories (exclusive
最小的, inclusive minimal, exclusive augmented, and inclusive augmented). All conditions in-
stantiated bipartitions of the person space with two-to-one mappings, but differed on which con-
trast was directly reflected in the forms (and which one was neutralized): an English-like contrast
([(西德:3)atomic]), a person contrast ([(西德:3)addressee]), a number contrast ([(西德:3)最小的]), or a random ho-
mophony pattern.
In Experiment 1B, participants were randomly assigned to one of three conditions, illustrated
在表中 5.8 Conditions differed in which subset of two first person categories was trained (批判的
training set) and held out (critical held-out set). This determined which alternative full paradigms
were consistent with the two categories participants learned. Condition 1 was consistent with an
English-like pattern (or with feature-based homophony). Conditions 2 和 3 were consistent with
either a person or a number contrast system (IE。, feature-based homophony), or with a random
contrast system (IE。, random homophony).
All participants in both Experiment 1A and Experiment 1B were additionally exposed to
another four pronominal forms that mapped into the second and third person singular and plural
类别. These forms were used as fillers and were not analyzed.
2.2.2 Materials The same materials were used in Experiments 1A and 1B. In both cases, 这
language consisted of six different pronoun forms, used for the filler categories (2SG/ PL, 3SG/
PL), plus the critical first person forms. For each participant, these six lexical items were randomly
drawn from a list of eight CVC words created following English phonotactics: kip, dool, heg,
rib, bub, veek, tosh, lom. Items were presented orthographically.
To express the pronoun meanings, we commissioned a cartoonist to draw scenarios involving
a family of three sisters and their parents. Each family member had a clearly defined role in the
8 Two additional conditions were also run, but are not reported here. Since these are orthogonal to the main aim
of the experiment, we simply refer to Maldonado and Culbertson 2019 for more details.
我
D
哦
w
n
哦
A
d
e
d
F
r
哦
米
H
t
t
p
:
/
/
d
我
r
e
C
t
.
米
我
t
.
e
d
你
/
我
我
/
n
G
A
r
t
我
C
e
–
p
d
我
F
/
/
/
/
5
3
2
2
9
5
2
0
1
4
2
4
4
/
我
我
n
G
_
A
_
0
0
4
0
6
p
d
.
F
乙
y
G
你
e
s
t
t
哦
n
0
7
S
e
p
e
米
乙
e
r
2
0
2
3
P E R S O N O F I N T E R E S T
305
桌子 5
Conditions in Experiment 1B. There were two training and two held-out categories per condition. 每个
training category was mapped to a different pronominal form (here called A or B), schematically
represented with light and dark gray. Participants had to use the training forms they learned (A or B) 到
express the held-out meanings (cells with A or B?). There are four different paradigms compatible with
the training per condition, as specified in the rightmost column.
Mappings
Compatible paradigms
Condition 1
Condition 2
Condition 3
EXCL
INCL
EXCL
INCL
EXCL
INCL
MIN
A
AUG
乙
A or B? A or B?
English-like,
Number-contrast,
Random (西德:2) 2
MIN
A or B?
A or B?
AUG
A
乙
Person-contrast,
Random (西德:2) 3
MIN
AUG
A or B? A or B?
A
乙
Number-contrast,
Random (西德:2) 3
我
D
哦
w
n
哦
A
d
e
d
F
r
哦
米
H
t
t
p
:
/
/
d
我
r
e
C
t
.
米
我
t
.
e
d
你
/
我
我
桌子 6
Highlighted family members for each person category. To ensure that forms were not associated with
specific quantities, critical augmented categories randomly included one or two additional others. 第三
person singular meanings were always expressed with a female other.
类别
Highlighted set
1 EXCL.MIN
1 INCL.MIN
1 EXCL.AUG
1 INCL.AUG
speaker
speaker, addressee
speaker, 其他(s)
speaker, addressee, 其他(s)
2.MIN
2.AUG
3.MIN
3.AUG
addressee
addressee, 其他(s)
one other
multiple others
conversational context. The two older sisters were speech act participants (in all scenarios, 他们
were either speaker or addressee). The third (little) sister was spatially close, but never a speech
act participant. The parents were seated in the background (serving as additional others).
Pronouns were used as one-word answers to questions like Who will be rich? Meanings
were expressed by visually highlighting subsets of family members, as in table 6. 在某些情况下,
more than one pattern of visual highlighting could match the target meaning; options were then
selected randomly. An example illustrating the INCL.MIN trial is provided in figure 2. All questions
/
n
G
A
r
t
我
C
e
–
p
d
我
F
/
/
/
/
5
3
2
2
9
5
2
0
1
4
2
4
4
/
我
我
n
G
_
A
_
0
0
4
0
6
p
d
.
F
乙
y
G
你
e
s
t
t
哦
n
0
7
S
e
p
e
米
乙
e
r
2
0
2
3
306
M O R A M A L D O N A D O A N D J E N N I F E R C U L B E R T S O N
(A) Exposure (INCL.MIN)
(乙) What if . . . (EXCL.AUG)
(C) Testing (INCL.MIN)
数字 2
Illustration of exposure, what if . . . , and testing trials. Feedback was presented for 2000 ms after response
in exposure and what if . . . 试验.
我
D
哦
w
n
哦
A
d
e
d
F
r
哦
米
H
t
t
p
:
/
/
d
我
r
e
C
t
.
米
我
t
.
e
d
你
/
我
我
/
n
G
A
r
t
我
C
e
–
p
d
我
F
/
/
/
/
5
3
2
2
9
5
2
0
1
4
2
4
4
/
我
我
n
G
_
A
_
0
0
4
0
6
p
d
.
F
乙
y
G
你
e
s
t
t
哦
n
0
7
S
e
p
e
米
乙
e
r
2
0
2
3
P E R S O N O F I N T E R E S T
307
were English interrogative sentences of the form Who will . . . ?, which were randomly drawn
from a list of 60 different tokens.
2.2.3 Procedure The basic procedure was the same in Experiments 1A and 1B. 参加者
were first introduced to the family, including the names of the sisters, and were told they were
going to see the sisters playing with a hat that had two magical properties: whoever wore it could
see the future but would also talk in a mysterious ancestral language. Participants were instructed
to figure out the meanings of the words in this new language. They were given the hint that the
words were not names, and an example trial with an English pronoun (她). 此外, 这
speaker and addressee roles switched several times during the experiment to highlight that the
words were dependent on contextually determined speech act roles. This was induced by swapping
who had the magical hat.
Experiment 1A had two phases. In the training phase, participants were first exposed to six
pronominal forms in the new language, corresponding to the four filler and four critical person
类别. Each exposure trial had two parts: a scene where a question was asked, and a scene
where the question was answered with a pronoun form in the target language (例如, figure 2a).
To check that participants were paying attention, they were then asked to select the pronominal
form they had just seen from two alternatives. 有 12 training trials (2 repetitions per
形式). After this initial exposure, participants were tested on the trained forms in what we call
what if . . . 试验. What if . . . trials consisted of a question-and-answer scene, as in the exposure
phase, followed by a “What if?” scene in which a new set of individuals was highlighted. Partici-
pants were asked to pick the correct word for that meaning from two alternatives (例如, figure
2乙). 有 32 such trials (3 repetitions per control form, 6 per critical form). 参加者
were given feedback on their answers. Participants were then given a final, critical test. 试炼
consisted of a question scene, followed by a scene highlighting the referent(s), but no pronominal
形式. Participants had to pick the word corresponding to the meaning from two alternatives (例如,
figure 2c). This phase consisted of 24 试验 (3 repetitions per form). Participants received no
feedback during this phase.
Experiment 1B also had a training and a testing phase. 至关重要的是, during the training phase
participants were only trained on the pronouns in the filler and critical training sets (6 人
类别). 有 12 exposure trials (2 repetitions per form) 和 16 what if . . . 试验 (2
repetitions per filler form, 4 per critical training form). Participants were given feedback on their
答案. The critical testing phase included trials for the two remaining critical categories, 那
是, the held-out set. This phase consisted of 48 试验 (6 repetitions per form). Participants received
no feedback during this phase.
Both experiments included a pretraining phase where participants were exposed only to the
three singular person pronouns. This was done to familiarize participants with the setup by using
less complex stimuli (IE。, scenes where a single family member was highlighted as the pronominal
referent). At the end of both experiments, participants were given a debrief questionnaire, 哪个
included questions targeting how they interpreted the meanings they were taught. 重要的,
most participants reported having understood the words as pronouns. 例如, 参与者
in Experiment 1B (Condition 2) described the meaning of form A as ‘me or us not including
我
D
哦
w
n
哦
A
d
e
d
F
r
哦
米
H
t
t
p
:
/
/
d
我
r
e
C
t
.
米
我
t
.
e
d
你
/
我
我
/
n
G
A
r
t
我
C
e
–
p
d
我
F
/
/
/
/
5
3
2
2
9
5
2
0
1
4
2
4
4
/
我
我
n
G
_
A
_
0
0
4
0
6
p
d
.
F
乙
y
G
你
e
s
t
t
哦
n
0
7
S
e
p
e
米
乙
e
r
2
0
2
3
308
M O R A M A L D O N A D O A N D J E N N I F E R C U L B E R T S O N
you’ and the meaning of form B as ‘us including you’.9 More details about the procedure used
in these experiments can be found in table A.1 in the online appendix.
2.2.4 Participants A total of 197 English-speaking adults were recruited via Amazon Mechani-
cal Turk for Experiment 1A (English-like group: 48, Person-contrast: 49, Number-contrast: 50,
Random: 50). Two further participants were excluded for not being self-reported native speakers
of English. 的 197 参与者, 171 (English-like group: 44, Person-contrast: 41, 数字-
对比: 41, Random: 45) responded accurately on more than 80% of exposure trials during the
training phase and were considered for further analysis, according to our preregistered plan.
Participants were paid 2 USD for their participation, which lasted approximately 15 minutes.
A different group of 253 English speakers were tested in Experiment 1B (Condition 1: 87,
Condition 2: 87, Condition 3: 79). Per our preregistered plan, participants were excluded if
(A) their accuracy rates during exposure training were below 80%, 或者 (乙) they had not answered
correctly more than two-thirds of the training trials. Note that high accuracy rates on trained
critical items are important here because extrapolation of these forms is only interpretable if
participants have learned them. This resulted in analysis of 131 参与者 (Condition 1: 46,
Condition 2: 49, Condition 3: 36). Participants were paid 2.5 USD for their participation, 哪个
lasted approximately 20 minutes.
2.3 结果
2.3.1 Experiment 1A Figure 3 shows the proportion of correct responses in critical trials per
experimental condition (English-like, Person-contrast, Number-contrast, and Random-contrast)
during the testing phase. We ran two logistic mixed-effects models (using the lme4 software
包裹 (Bates et al. 2014) 在R中 (R核心团队 2018)) to evaluate the effect of the experimental
condition on accuracy rates (coded as 0 或者 1). Both models included by-participant random inter-
油炸. Here and throughout, the standard alpha level of 0.05 was used to determine significance,
and p-values were obtained using asymptotic Wald tests.
A first model assessed whether the English-like pattern was learned better than the alternative
paradigms. We used treatment contrast coding with the English-like paradigm as baseline; each
of the remaining conditions was compared with this fixed level. The model revealed that the
proportion of correct responses in the English-like group was significantly higher than chance ((西德:4)
(西德:4) 1.78, p (西德:5) .001). Accuracy in the Person-contrast and Number-contrast groups did not differ
significantly from that of the English-like baseline (Person-contrast: (西德:4)(西德:4) (西德:6)0.6, p (西德:4) .093; Num-
ber-contrast: (西德:4)(西德:4) (西德:6)0.38, p (西德:4) .28); 然而, accuracy in the Random group was significantly
lower than the baseline ((西德:4)(西德:4) (西德:6)1.24; p (西德:5) .001). This matches the visual pattern in figure 3.
9 Not all participants reported pronouns for these meanings. Interpreting participants’ responses in these cases is
not straightforward. 例如, a highly accurate participant reported the meaning of form A to be ‘sisters’. 这
suggests that questionnaire responses do not necessarily convey what participants have implicitly learned. We therefore
use questionnaire responses as a general sanity check but rely on accuracy rates to draw conclusions about participants’
performance in the experiment.
我
D
哦
w
n
哦
A
d
e
d
F
r
哦
米
H
t
t
p
:
/
/
d
我
r
e
C
t
.
米
我
t
.
e
d
你
/
我
我
/
n
G
A
r
t
我
C
e
–
p
d
我
F
/
/
/
/
5
3
2
2
9
5
2
0
1
4
2
4
4
/
我
我
n
G
_
A
_
0
0
4
0
6
p
d
.
F
乙
y
G
你
e
s
t
t
哦
n
0
7
S
e
p
e
米
乙
e
r
2
0
2
3
P E R S O N O F I N T E R E S T
309
数字 3
Accuracy rates in critical testing trials by condition in Experiment 1A. Error bars represent standard error
on by-participant means; gray dots represent individual participants’ means.
We ran a second model to explore the difference between the feature-based and random
图案. The analysis was restricted to Number-contrast, Person-contrast, and Random conditions.
We used treatment coding with the Random condition as baseline. The proportion of correct
responses in this baseline group was significantly higher than chance ((西德:4)(西德:4) 0.58, p (西德:5) .001), 但
significantly lower than in both feature-based conditions (Number-contrast: (西德:4)(西德:4) 0.8, p (西德:5) .01;
Person-contrast: (西德:4)(西德:4) 0.62, p (西德:4) .03). This suggests that participants trained to make a (nonnative)
person or number contrast were more accurate than those trained on a random contrast.
2.3.2 Experiment 1B Recall that participants in Experiment 1B were taught two pronominal
形式 (coded as forms A and B), which they had to use to describe both a critical trained set and
a held-out set of person meanings involving the speaker (级别: EXCL.MIN, EXCL.AUG, INCL.MIN,
INCL.AUG). 数字 4 shows the proportion of trials on which participants chose the form A (亲-
名词) for each critical trained and held-out meaning during the testing phase. Choice of the
same form across categories indicates homophony. A visual inspection of figure 4 suggests that
participants in Condition 1 consistently used one form for the EXCL.MIN category and the other
form for the remaining three categories; this indicates inference of an English-like paradigm.
Participants in Conditions 2 和 3 appear somewhat noisier in their responses; 然而, distinct
patterns are evident. In Condition 2, one form is used for the two first person categories, 和, 在
least for some participants, the other form is used for the two inclusive categories (consistent
with maintenance of the person contrast, IE。, number homophony). In Condition 3, one form is
used for the minimal categories and the other for the plurals (consistent with maintenance of the
number contrast, IE。, person homophony). 笔记, 然而, that this figure shows by-participant
averages for each meaning category rather than which patterns individual participants produced.
我
D
哦
w
n
哦
A
d
e
d
F
r
哦
米
H
t
t
p
:
/
/
d
我
r
e
C
t
.
米
我
t
.
e
d
你
/
我
我
/
n
G
A
r
t
我
C
e
–
p
d
我
F
/
/
/
/
5
3
2
2
9
5
2
0
1
4
2
4
4
/
我
我
n
G
_
A
_
0
0
4
0
6
p
d
.
F
乙
y
G
你
e
s
t
t
哦
n
0
7
S
e
p
e
米
乙
e
r
2
0
2
3
310
M O R A M A L D O N A D O A N D J E N N I F E R C U L B E R T S O N
数字 4
Proportion of form A (as opposed to form B) choices for each first person category during the testing phase
in Experiment 1B. Choice of the same form (A or B) across categories indicates homophony. Error bars
represent standard error on by-participant means; dots represent individual participants’ means.
数字 4 suggests that there is relatively little variation across participants in Condition 1
compared with the other conditions; almost all participants chose the same form for each category,
and they tended to do so categorically. To confirm this statistically, we calculated the joint entropy
of the held-out set for each individual. This value indicates the degree of uncertainty or variability
in each participant’s mapping of the trained forms to the held-out categories. Participants who
are less consistent in their mapping will have higher joint entropy values. We then fit a linear
regression model predicting joint entropy by Condition (3 级别). We used treatment coding,
with Condition 1 as baseline. No random effects were included in the model, as each participant
had a single joint entropy value. As predicted, joint entropy rates were significantly higher for
Conditions 2 和 3 (intercept: (西德:4)(西德:4) 0.28; 与. 2: (西德:4)(西德:4) 0.44 (西德:3).13, p (西德:5) .001; 与. 3: (西德:4)(西德:4) 0.64
(西德:3).12, p (西德:5) .001).
A second analysis evaluated whether individual participants in Conditions 2 和 3 were more
likely to infer feature-based rather than random patterns (as suggested by figure 4).10 We calculated
the probability that participants were deriving a feature-based pattern (a person contrast in Condi-
的 2 or a number contrast in Condition 3) given their responses to held-out meanings (见图
A.2 in the online appendix). In Condition 2, we computed the probability of choosing form A
for the EXCL.MIN and form B for the INCL.MIN; in Condition 3, we computed the probability of
choosing form A for the EXCL.MIN and form B for the EXCL.AUG.
We then ran nonparametric Wilcoxon signed-rank tests per condition to determine whether
the probability of deriving a feature-based pattern was higher than chance. Given that there were
10 This analysis diverges from our preregistered plan, which was designed to test this same prediction with a different
analysis method. We believe the current analysis is both simpler and more technically sound.
我
D
哦
w
n
哦
A
d
e
d
F
r
哦
米
H
t
t
p
:
/
/
d
我
r
e
C
t
.
米
我
t
.
e
d
你
/
我
我
/
n
G
A
r
t
我
C
e
–
p
d
我
F
/
/
/
/
5
3
2
2
9
5
2
0
1
4
2
4
4
/
我
我
n
G
_
A
_
0
0
4
0
6
p
d
.
F
乙
y
G
你
e
s
t
t
哦
n
0
7
S
e
p
e
米
乙
e
r
2
0
2
3
P E R S O N O F I N T E R E S T
311
four paradigms compatible with the training in each condition, chance level was set at 25%. 这
results of these tests indicate that the probability of deriving a person contrast in Condition 2 曾是
not significantly different from chance ( p (西德:4) .406), but the probability of deriving a number
contrast in Condition 3 was above chance ( p (西德:5) .001). The same procedure was followed in
Condition 1 with respect to the probability of deriving an English-like pattern. 正如预期的那样, 这
probability was significantly higher than chance ( p (西德:5) .001).
2.4 讨论
The main aim of these first two experiments was to test whether learners are sensitive to feature
combinations, or contrasts, that are not present in their native language. We exposed English-
speaking learners to paradigms expressing four person-number categories in a new language. 我们
focused on systems instantiating either the inclusive/exclusive or the minimal/augmented contrast,
which have been argued to have a universal basis in two features, encoding person and number,
分别 (例如, [(西德:3)addressee], [(西德:3)最小的]). Our experiments also included an English-like para-
digm, as a sanity check. We predicted that participants would find the English-like paradigm
easiest to learn, followed by the two non-native-like feature-based homophony patterns, 和
random homophony being least learnable.
In Experiment 1A, we tested these predictions by training participants on one of four para-
digms and comparing how accurately they learned each one. Results confirmed a numeric advan-
tage for native-like pronoun systems: learners found it easier to learn a paradigm with the same
structure as English. 有趣的是, there was no statistically significant difference in learnability
between English-like systems and systems instantiating number or person homophony, suggesting
that participants can readily learn other feature-based partitions as well. Results also confirmed
that participants trained on a pronominal system with a (nonnative) person or number homophony
pattern outperformed those trained on a random homophony pattern. This supports the claim that
learners perceive this four-cell person space (IE。, EXCL, INCL, 2, 和 3) as the interaction of two
distinct features, rather than as a conjunction of four different categories, fully independent from
彼此 (in line with Sauerland and Bobaljik 2013). If learners divide the exclusive, 包括的,
最小的, and augmented categories into two natural semantic classes, one for person and one for
数字, then learning our systematic homophony paradigms simply involves one nonnative con-
trast each.
These results were for the most part confirmed in Experiment 1B, where participants were
trained on ambiguous data that required them to extrapolate trained forms to new meanings. 这里,
learners were significantly more likely to infer an English-like pattern. They were also more likely
to infer a paradigm with a feature-based number contrast (and person-based homophony) than a
random contrast, thus making productive use of the nonnative [(西德:3)最小的] distinction. As noted
多于, in order to do so—that is, to treat the i (speaker only) and the iu (speaker and addressee
仅有的) references as “minimal”–participants need to be sensitive not only to the minimal/aug-
mented contrast but also to the inclusive/exclusive one, as the dyad of speaker and addressee
can only be considered a minimal unit if there is an inclusive category.
我
D
哦
w
n
哦
A
d
e
d
F
r
哦
米
H
t
t
p
:
/
/
d
我
r
e
C
t
.
米
我
t
.
e
d
你
/
我
我
/
n
G
A
r
t
我
C
e
–
p
d
我
F
/
/
/
/
5
3
2
2
9
5
2
0
1
4
2
4
4
/
我
我
n
G
_
A
_
0
0
4
0
6
p
d
.
F
乙
y
G
你
e
s
t
t
哦
n
0
7
S
e
p
e
米
乙
e
r
2
0
2
3
312
M O R A M A L D O N A D O A N D J E N N I F E R C U L B E R T S O N
We did not find a significant difference in participants’ likelihood of inferring a feature-
based person contrast (and number homophony) over a random one. 换句话说, after being
trained on an inclusive/exclusive distinction in part of the paradigm (EXCL.AUG and INCL.AUG),
participants did not generalize this contrast to the held-out cells of the paradigm. 这表明
that although the nonnative clusivity distinction is indeed learnable (in Experiment 1A), 英语-
speaking learners do not necessarily make a productive use of it. The apparent difference between
number and person homophony is supported by a post hoc analysis showing that accuracy rates
on trained categories (before exclusion) are higher in Condition 3 than in Condition 2 ( p (西德:5) .001).
这 [(西德:3)addressee] person distinction in pronominal forms was thus harder to learn than the [(西德:3)mini-
马尔] distinction.
One possible explanation for this difference is that it reflects participants’ experience with
homophony in English: since English encodes a ([(西德:3)atomic]) number distinction, it is possible to
characterize it as a system with (仅有的) person homophony (IE。, a noninclusive language; Harbour
2016). 换句话说, English speakers have more experience with distinctions in number (在
一般的) than in clusivity.11
或者, the fact that English speakers do not generalize this person contrast could be
thought of as the result of applying a native constraint against such inclusive systems. 相比之下
with binary-features accounts, some approaches (例如, Harley and Ritter 2002) describe noninclu-
sive systems as making use of two privative features (Speaker and Addressee) together with a
constraint that prevents these features from cooccurring. 可以说, the apparent difficulty in
generalizing the inclusive/exclusive distinction might suggest that English speakers have a harder
time learning a distinction that violates a native constraint against the simultaneous specification
of the Speaker and Addressee features.12
总结一下, this study presents the first experimental evidence for differences in learnabil-
ity between alternative person paradigms. 不出所料, native-like paradigms are easiest to
learn and most likely to be inferred when input is ambiguous. More interestingly, paradigms
exhibiting homophony within a natural class are learned (and in some cases inferred) more readily
than paradigms with random homophony. 下文中, we build on these basic results to
investigate more specific constraints on the person space hypothesized by feature-based theories
of person.
11 重要的, our findings are not compatible with participants’ being only sensitive to the specific number distinc-
tions found in English. 尤其, both Conditions 2 和 3 involve collapsing categories that are expressed as distinct
pronouns—I and we—in English. While one could in principle argue that this explains why participants in Condition 2
are less likely to collapse the exclusive minimal and augmented categories, this would not explain why participants in
Condition 3 readily collapsed the exclusive and inclusive minimal.
12 Note that the advantage for a paradigm with a number contrast over a person contrast (IE。, for person over
number homophony) found in Experiment 1B contrasts with typological data, which suggests that languages with different
pronominal forms for exclusive and inclusive persons but no number distinction are more common than minimal/aug-
mented languages that do not make an inclusive contrast (Cysouw 2013). 然而, these counts are very sparse.
我
D
哦
w
n
哦
A
d
e
d
F
r
哦
米
H
t
t
p
:
/
/
d
我
r
e
C
t
.
米
我
t
.
e
d
你
/
我
我
/
n
G
A
r
t
我
C
e
–
p
d
我
F
/
/
/
/
5
3
2
2
9
5
2
0
1
4
2
4
4
/
我
我
n
G
_
A
_
0
0
4
0
6
p
d
.
F
乙
y
G
你
e
s
t
t
哦
n
0
7
S
e
p
e
米
乙
e
r
2
0
2
3
P E R S O N O F I N T E R E S T
313
3 实验 2: Within You, Without You
3.1 介绍
In a classic paper, Zwicky (1977) made the following observation regarding the crosslinguistic
distribution of person systems: in languages that do not distinguish clusivity (例如, 英语), 这
‘you and us’ inclusive meaning is always expressed as a form of ‘us’ and never as a form of
‘you’ (or ‘them’).13
At first glance, Zwicky’s generalization is quite surprising. Most feature-based approaches
to person systems (例如, Bobaljik 2008) assume that the inclusive person shares features with both
the first person exclusive (例如, [(西德:5)speaker]) and the second person exclusive (例如, [(西德:5)addressee]).
的确, a number of languages have inclusive pronouns that can be morphologically decomposed
into first plus second person forms (例如, Bislama; Harbour 2011).14 This leads naturally to the
expectation that languages should be as likely to assimilate the inclusive with the second person
as they are to assimilate it with the first. 相比之下, no theory would predict the inclusive meaning
to be homophonous with the third person, as the inclusive and the third person do not have any
features in common (although see Rodrigues 1990 for a potential exception).
There have been two general approaches to Zwicky’s generalization in the literature. 这
first maintains the traditional set of features, but posits default feature specifications in order
to predict an asymmetry between first-inclusive and second-inclusive (Harley and Ritter 2002,
麦金尼斯 2005). Harley and Ritter’s (2002) feature geometry account maintains both Speaker
and Addressee features as dependent nodes of the feature Participant, but the Speaker feature is
considered to be less marked than the Addressee feature. 最后, in languages without an
inclusive distinction, a preference for assimilating the inclusive meaning into the first person is
预期的, as they share the default feature. Defaults can be overridden; 所以, 第二-
inclusive homophony pattern can still arise. 相比之下, a third-inclusive system is impossible.
The second approach is to use a different set of features. Harbour (2016) posits [(西德:3)作者]
和 [(西德:3)参与者], denoting the semilattices (西德:2)我(西德:3) 和 (西德:2)我, iu, 你(西德:3) 分别, with the values of the
features modeled as complementary operations on lattices. While in Harbour’s system the inclusive
and second person categories do share “ontological” primitives—both of them contain u—the
absence of a [(西德:3)addressee] feature—and of a lattice consisting only of (西德:2)你(西德:3)—creates an inherent
asymmetry in how speaker and addressee roles can be represented in a person partition.15 This
asymmetry derives Zwicky’s observation as a strong constraint on possible person systems. A
language that makes use of the [(西德:3)作者] feature will have a bipartition of the person space in
which i and iu are homophonous and u and o are homophonous. 相似地, a language with both
[(西德:3)作者] 和 [(西德:3)参与者] can have a tripartition (如果 [(西德:3)参与者] composes last; see Harbour
13 This is a generalization about languages and not about individual paradigms within a language, which might show
accidental homophony (see Harbour 2016 and examples therein).
14 Although note that there are also languages where the inclusive form patterns morphologically with the second
人 (例如, Ojibwa pronouns; Harley and Ritter 2002 from Schwartz and Dunnigan 1986).
15 Harbour’s (2016) proposed ontology consists of egocentrically nested subsets, such that the smallest subset in the
ontology contains the speaker alone—that is, 我 (西德:2) 我, iu, 你 (西德:2) 我, iu, 你, 哦. This ontology is shared by Ackema and Neeleman
(2018).
我
D
哦
w
n
哦
A
d
e
d
F
r
哦
米
H
t
t
p
:
/
/
d
我
r
e
C
t
.
米
我
t
.
e
d
你
/
我
我
/
n
G
A
r
t
我
C
e
–
p
d
我
F
/
/
/
/
5
3
2
2
9
5
2
0
1
4
2
4
4
/
我
我
n
G
_
A
_
0
0
4
0
6
p
d
.
F
乙
y
G
你
e
s
t
t
哦
n
0
7
S
e
p
e
米
乙
e
r
2
0
2
3
314
M O R A M A L D O N A D O A N D J E N N I F E R C U L B E R T S O N
2016 欲了解详情) in which i and iu are homophonous (with two additional forms for u and o), 或者
a quadripartition (如果 [(西德:3)参与者] composes first). Without a corresponding [(西德:3)addressee] feature,
尽管, there is no way to have a system that picks out the set including iu and u. 的确,
tripartitions involving homophony between inclusive and second person or inclusive and third
person are equally impossible.16
The theories outlined above differ critically in how second-inclusive and third-inclusive are
treated. For Harley and Ritter (2002), third-inclusive is singled out as underivable, while second-
inclusive is possible but more marked than first-inclusive. 相比之下, Harbour (2016) takes as
his starting point the idea that only first-inclusive tripartitions can be generated by the grammar.
On the basis of typology alone, it is impossible to adjudicate between these theories: both second-
and third-inclusive patterns are unattested. 而且, neither theory provides an explicit mecha-
nism for linking the feature-based representations (and operations) they posit to typology. 这
implicit link is learnability: only a subset of possible person partitions are learnable by humans;
alternatively, some are learned more readily than others. In Experiment 2, we investigated learners’
sensitivity to predicted asymmetries among noninclusive paradigms. 要做到这一点, we used an ease-
of-learning design: we trained English-speaking learners on a new language with an inclusive
that was a form of us (first-inclusive), a form of you all (second-inclusive), or a form of them
(third-inclusive), and compared how well the systems were learned. Given that English features
first-inclusive homophony and this is the only tripartition systematically attested in typology,
learners were predicted to prefer such paradigms over alternatives. Regarding second-inclusive
and third-inclusive homophony, if both patterns are directly ruled out by the grammar (如
Harbour 2016), learners were expected to be equally unlikely to learn either of them. 相比之下,
if learners are sensitive to the featural commonalities between inclusive and second person (例如,
[(西德:5)addressee]), a second-inclusive system was expected to be easier to learn than a third-inclusive
一 (Harley and Ritter 2002, 麦金尼斯 2005). This pattern of results would suggest that an
asymmetry between first-inclusive and second-inclusive languages should not be encoded as a
hard constraint on person systems (contra Harbour 2016).
3.2 方法
This experiment, including all hypotheses, 预测, and analyses, was preregistered at https://
osf.io/5h4m6. All materials, 数据, and scripts are provided at https://osf.io/p2c4r/?view_only(西德:4)
ba7a554345f84a2dbddfbcfea67691d3. This experiment was implemented using the JavaScript
library jsPsych (De Leeuw 2015) and presented in a web browser.
3.2.1 Design Participants were randomly assigned to one of three conditions, summarized in
桌子 7. Participants in all conditions were taught three pronominal forms mapped to four plural
person categories (first exclusive, 包括的, second exclusive, 第三个). Note that in this experi-
ment we rely on a singular/plural (not minimal/augmented) number contrast, with no number
16 An intermediate proposal is offered by Ackema and Neeleman (2013, 2018). In their proposed feature structure,
“there is no natural class . . . that comprises the first person inclusive and the second person, but not the first person
exclusive” (2013:910). 然而, second-inclusive patterns can still be obtained by incorporating an impoverishment rule
in the system (see footnote 28 欲了解详情). This is not possible for inclusive-third homophony.
我
D
哦
w
n
哦
A
d
e
d
F
r
哦
米
H
t
t
p
:
/
/
d
我
r
e
C
t
.
米
我
t
.
e
d
你
/
我
我
/
n
G
A
r
t
我
C
e
–
p
d
我
F
/
/
/
/
5
3
2
2
9
5
2
0
1
4
2
4
4
/
我
我
n
G
_
A
_
0
0
4
0
6
p
d
.
F
乙
y
G
你
e
s
t
t
哦
n
0
7
S
e
p
e
米
乙
e
r
2
0
2
3
P E R S O N O F I N T E R E S T
315
桌子 7
Conditions in Experiment 2. Grayed cells are mapping to a single pronominal form, white cells to
different and distinct forms.
A. First-inclusive
乙. Second-inclusive
C. Third-inclusive
1 EXCL
1 INCL
2
3
1 EXCL
1 INCL
2
3
1 EXCL
1 INCL
2
3
我
D
哦
w
n
哦
A
d
e
d
F
r
哦
米
H
t
t
p
:
/
/
d
我
r
e
C
t
.
米
我
t
.
e
d
你
/
我
我
distinction in the inclusive person (INCL). Conditions differed in whether the inclusive meaning
was assimilated into the first person plural (First-inclusive condition), the second person plural
(Second-inclusive condition), or the third person plural (Third-inclusive condition).
Participants in all three conditions were also exposed to three distinct pronominal forms
corresponding to the first, 第二, and third singular persons. Participants’ learning of these forms
was used as an exclusion criterion (见下文).
3.2.2 Materials The language consisted of six different pronominal forms: three forms were
used for the plural pronouns (critical categories), and three different forms were used for the
singular pronouns (filler categories). For each participant, these were randomly drawn from a list
of eight CVC words (see Experiment 1).
Pronouns were again used as one-word answers to English interrogative sentences of the
form Who will . . . ?, randomly drawn from a list of 60 different tokens. Meanings were expressed
by highlighting subsets of family members, as in table 8. Visual stimuli were the same as in
实验 1.
3.2.3 Procedure The general backstory was as in Experiment 1. Participants were instructed to
figure out the meanings of the words in the new language, and they were told that the words they
were learning were not names. As in Experiment 1, the speaker and addressee roles switched
several times during the experiment to highlight that the words were context-dependent.
The experiment had two phases, each composed of exposure and testing blocks (例如, figure
5). Trials in each of these blocks were analogous to those used in Experiment 1, except that
participants had to select the correct word for that meaning among three different options (不是
二).
/
n
G
A
r
t
我
C
e
–
p
d
我
F
/
/
/
/
5
3
2
2
9
5
2
0
1
4
2
4
4
/
我
我
n
G
_
A
_
0
0
4
0
6
p
d
.
F
乙
y
G
你
e
s
t
t
哦
n
0
7
S
e
p
e
米
乙
e
r
2
0
2
3
316
M O R A M A L D O N A D O A N D J E N N I F E R C U L B E R T S O N
桌子 8
Highlighted family members for each person category. 1 EXCL, 2, 和 3 plural categories randomly
included one or two additional others; the INCL category could refer to speaker and addressee alone or
include one or two others as well. In this experiment, there was no distinction between minimal and
augmented; 相当, there was a distinction between singular/atomic and plural.
类别
Highlighted set
1 EXCL.SG
2SG
3SG
1 EXCL.PL
1 INCL
2PL
3PL
speaker
addressee
one other
speaker, 其他(s)
speaker, addressee (其他(s))
addressee, 其他(s)
multiple others
During the first phase, participants were trained and tested on the three singular pronouns.
There were a total of 12 exposure and 12 testing trials (4 repetitions per form/meaning). Partici-
pants who responded accurately to at least two-thirds of the testing trials in this phase (8 正确的
responses) moved on to the second, critical phase. This phase was composed of two alternating
exposure and testing blocks targeting the mapping between three plural pronouns and four person
含义. There were a total of 24 exposure trials (6 repetitions per meaning) 和 48 testing
试验 (12 repetitions per meaning). More details about the experimental procedure are provided
in table A.2 in the online appendix.
The order of presentation of meanings was fully randomized within exposure and testing
blocks for each participant. As in Experiment 1, participants were given a debrief questionnaire
at the end of the experiment to check how they interpreted the forms they were trained on. 为了
例子, participants in the Second-inclusive condition described the meaning of the critical form
as ‘me and you or you all’ or as ‘group containing Ann or Mary’.
3.2.4 Participants A total of 320 English-speaking adults were recruited via Amazon Mechani-
cal Turk (First-inclusive group: 109, Second-inclusive: 101, Third-inclusive: 110). This does not
include workers who were excluded for not being self-reported native speakers of English (10)
and participants who failed to pass an attention check included at the beginning of the experiment
(35). This attention check was added because our exit questionnaires in Experiment 1 透露
that many participants had not read the instructions or were bots (Rouse 2015). While these
participants were usually excluded by our other criteria, the attention check allowed us to filter
them out in advance, distinguishing them from participants who just found the experiment hard.
A total of 167 participants responded accurately on more than eight singular testing trials and
were allowed to continue with the critical plural pronoun phase, according to our preregistered
plan (First-inclusive group: 57, Second-inclusive: 55, Third-inclusive: 55). Participants who fin-
ished the two experimental phases (大约 20 minutes long) were paid 3.5 USD; partici-
pants who only completed the first part of the experiment (大约 5 minutes long) 是
paid 1 USD.
我
D
哦
w
n
哦
A
d
e
d
F
r
哦
米
H
t
t
p
:
/
/
d
我
r
e
C
t
.
米
我
t
.
e
d
你
/
我
我
/
n
G
A
r
t
我
C
e
–
p
d
我
F
/
/
/
/
5
3
2
2
9
5
2
0
1
4
2
4
4
/
我
我
n
G
_
A
_
0
0
4
0
6
p
d
.
F
乙
y
G
你
e
s
t
t
哦
n
0
7
S
e
p
e
米
乙
e
r
2
0
2
3
P E R S O N O F I N T E R E S T
317
(A) Exposure
(乙) Testing
数字 5
Illustration of exposure and testing trials for the EXCL.PL category in Experiment 2
3.3 结果
Mean accuracy rates on testing trials during the critical phase are given in figure 6. The effect
of condition and block on accuracy rates was analyzed using logistic mixed-effect models with
random by-participant intercepts and by-block slopes.17
17 There were two testing blocks in the critical phase, each preceded by an exposure block. Participants were generally
expected to improve with accumulated exposure, but this improvement could vary across conditions. Each model included
the effect of block on accuracy, as well as the interaction with condition. 然而, we report here only simple effects
regarding the second testing block. The complete model output can be found at https://osf.io/p2c4r/?view_only(西德:4)ba7a554
345f84a2dbddfbcfea67691d3.
我
D
哦
w
n
哦
A
d
e
d
F
r
哦
米
H
t
t
p
:
/
/
d
我
r
e
C
t
.
米
我
t
.
e
d
你
/
我
我
/
n
G
A
r
t
我
C
e
–
p
d
我
F
/
/
/
/
5
3
2
2
9
5
2
0
1
4
2
4
4
/
我
我
n
G
_
A
_
0
0
4
0
6
p
d
.
F
乙
y
G
你
e
s
t
t
哦
n
0
7
S
e
p
e
米
乙
e
r
2
0
2
3
318
M O R A M A L D O N A D O A N D J E N N I F E R C U L B E R T S O N
数字 6
Accuracy rates in critical testing trials by condition in Experiment 2. Error bars represent standard error on
by-participant means; gray dots represent individual participants’ means. Dashed black line indicates chance
价值 (for three-options forced-choice).
We first compared the First-inclusive with the Second-inclusive and Third-inclusive condi-
系统蒸发散 (contrasts were treatment-coded, with First-inclusive and Block 2 as baselines). The model
intercept was significant, indicating that accuracy in the First-inclusive (Block 2) was above
机会 ((西德:4)(西德:4) 1.59, p (西德:5) .001). 此外, compared with the First-inclusive condition, 准确性
was significantly lower in the Third-inclusive ((西德:4)(西德:4) (西德:6)1.83, p (西德:5) .001) and marginally lower in
the Second-inclusive ((西德:4)(西德:4) (西德:6)0.572, p (西德:4) .055).
We then compared the Second-inclusive and Third-inclusive conditions (contrasts were treat-
ment-coded, with Second-inclusive and Block 2 as baselines). The model intercept was significant,
indicating that accuracy in the Second-inclusive (Block 2) was above chance ((西德:4)(西德:4) 0.97, p (西德:5)
.001). Accuracy was significantly lower in the Third-inclusive ((西德:4)(西德:4) (西德:6)1.2, p (西德:5) .001).18
3.4 讨论
These findings confirm that participants are most successful at learning a new language that
features homophony between inclusive and first person meanings. This is as expected since this
pattern is systematically found in the typology and reflects our English-speaking participants’
native system. If Harbour (2016) is correct in deriving a hard constraint on tripartitions that
generates only first-inclusive homophony, then we might expect this preference to be very strong
的确. 然而, the difference in accuracy between the First-inclusive and Second-inclusive
conditions was only marginally significant. This result is consistent with the idea that learners
18 Following our preregistration, we ran a second version of each of these models, restricting the analysis to inclusive
含义 (since our hypotheses target the inclusive category specifically). The same pattern of results emerged.
我
D
哦
w
n
哦
A
d
e
d
F
r
哦
米
H
t
t
p
:
/
/
d
我
r
e
C
t
.
米
我
t
.
e
d
你
/
我
我
/
n
G
A
r
t
我
C
e
–
p
d
我
F
/
/
/
/
5
3
2
2
9
5
2
0
1
4
2
4
4
/
我
我
n
G
_
A
_
0
0
4
0
6
p
d
.
F
乙
y
G
你
e
s
t
t
哦
n
0
7
S
e
p
e
米
乙
e
r
2
0
2
3
P E R S O N O F I N T E R E S T
319
are sensitive to the featural overlap between inclusive and second person, as predicted by Harley
and Ritter (2002), 麦金尼斯 (2005), and Bobaljik (2008). Participants in the Second-inclusive
condition may be treating these as a natural class, relying on a shared feature to learn the partition.
This result supports a theory in which first- and second-inclusive tripartitions are both generated
by the grammar, but the latter is dispreferred (contra Harbour 2016, and possibly Ackema and
Neeleman 2018; but see footnote 16).19
重要的, we also found that learners have a bias against systems that assimilate the
inclusive into the third person. This result reveals that second- and third-inclusive systems, 尽管
存在 (一般来说) unattested in the typology, are not equal from a learnability perspective: 那里
is stronger pressure against third-inclusive than against second-inclusive homophony. 这是
again as predicted by theories like Harley and Ritter’s (2002), McGinnis’s (2005), and Bobaljik’s
(2008) (but arguably also by Ackema and Neeleman’s (2018)), which posit that, unlike first and
second person, third person does not form a natural class with the inclusive category (cf., 例如,
Rodrigues 1990). 因此, homophony between inclusive and third persons is not predicted
to occur systematically, though it might arise accidentally. The learnability cost for third-inclusive
systems can therefore be seen as another case of a bias against random homophony.
Returning to Zwicky’s observation, our findings suggest that the typological asymmetry
between alternative noninclusive systems may not have a simple correlate with learning. 在里面
typology, first-inclusive systems are attested systematically, while second- and third-inclusive
systems are not; in our experiment, third-inclusive systems were clearly dispreferred, 但是
advantage for first- over second-inclusive systems was much weaker. While this is consistent
with a theory positing weak learning biases (IE。, constraints that can be overridden given sufficient
证据) that penalize third-inclusive most, it still leaves the typological data partially unex-
平原.
One possibility is that there is an additional weak bias, not at play in our experiments, 那
further advantages first-inclusive systems. An obvious such candidate is a general egocentric
bias—that is, increased importance or salience of speakers to themselves (Charney 1980, Loveland
1984, Moyer et al. 2015). If individuals perceive the world as a function of their presence in it,
they may be more likely to adopt categorization systems that preserve this distinction. This would
lead to an asymmetry between first-inclusive and second-inclusive systems. This bias may be
weakened in the context of our experiment, where participants are passive learners and do not
themselves feature in the meanings they are learning. We return to this issue in section 5.
19 或者, as Daniel Harbour (pers. comm.) points out, one could argue that our participants are treating the
pronominal system we teach them as an instance of syncretism within an otherwise inclusive language. If this were the
案件, theories like Harbour’s (2016) could still account for our results. In Harbour’s system, inclusive languages feature/
make quadripartitions of the person space. Underlying quadripartitions might show feature-based syncretism of inclusive
and second person in some paradigms, as these categories do share a [(西德:5)参与者] feature. 然而, given that our
participants are speakers of a noninclusive language, it seems very unlikely that they would infer an underlying quadriparti-
tion from a three-form pronominal system. 或者, because Harbour (2016) posits a shared ontological primitive,
你, potentially linking inclusive and second person, one could reason that participants in the experiment are relying on
this ontological/semantic overlap to learn the system, without paying attention to the features themselves. 然而, 如果
that were the case, one would need to explain why this does not lead to second-inclusive systems in the typology.
我
D
哦
w
n
哦
A
d
e
d
F
r
哦
米
H
t
t
p
:
/
/
d
我
r
e
C
t
.
米
我
t
.
e
d
你
/
我
我
/
n
G
A
r
t
我
C
e
–
p
d
我
F
/
/
/
/
5
3
2
2
9
5
2
0
1
4
2
4
4
/
我
我
n
G
_
A
_
0
0
4
0
6
p
d
.
F
乙
y
G
你
e
s
t
t
哦
n
0
7
S
e
p
e
米
乙
e
r
2
0
2
3
320
M O R A M A L D O N A D O A N D J E N N I F E R C U L B E R T S O N
In the next section, we investigate a second typological generalization that appears to chal-
lenge the categorical distinction between participants and nonparticipants suggested by Zwicky’s
generalization.
4 实验 3: I Me Mine
4.1 介绍
It has long been observed that there is a fundamental difference between person categories involv-
ing speech act participants (first exclusive, 包括的, and second persons) and the third person,
which refers to other, nonparticipant individuals. Besides having a fixed reference, which does
not depend on discourse roles, in a number of languages the third person is treated as morphologi-
cally distinct from other person categories (Forchheimer 1953; see summary in Harley and Ritter
2002). These facts have often led researchers to propose that the third person is unmarked with
respect to the first and second person categories (Benveniste 1971), which are instead encompassed
by the same natural class: 参与者 (黑尔 1973, Silverstein 1976, Noyer 1992). Here we will
focus on one specific instantiation of this proposal, Harley and Ritter’s (2002) feature geometry
方法.
Harley and Ritter (H&右) capture the intuition that the third person is unmarked by treating
it as the default interpretation of the base node (called Referring Expression). First exclusive,
first inclusive, and second persons require the presence of the dependent node Participant. 为了
an illustration of H&R’s geometry, 见图 1.
Despite accounting for a number of interesting typological patterns, H&R’s system (和
others like it; see Be´jar 2003, 麦金尼斯 2005) is potentially challenged by the crosslinguistic
distribution of person homophony patterns. 具体来说, these approaches cannot account for the
following typological observation, simplified from (2).20
(3) Languages that feature homophony between second and third person categories and
between first and second are more frequent than those instantiating first-third ho-
mophony.
The numerically higher frequency of first with second (1(西德:5)2) and second with third (2(西德:5)3) 系统
relative to first with third (1(西德:5)3) systems is illustrated in figure 7. This tendency is found both in
free pronouns and in verbal agreement, but it is restricted to nonsingular or number-neutral contexts
(IE。, it does not hold for singular cases).21 注意 2(西德:5)3 homophony is also numerically more
common than 1(西德:5)2 图案. We return to this in section 4.5.
20 This generalization does not directly concern the inclusive category, as it mostly holds for noninclusive languages,
where the inclusive is collapsed with the first person exclusive. For the sake of simplicity, we therefore remove the
inclusive from the discussion and refer to both the first person inclusive and first person exclusive as the first (1) 人
类别.
21 The counts in Baerman, 棕色的, and Corbett 2005 for personal pronouns and verbal agreement are drawn from
two different typological samples, making the counts not fully comparable to each other. 下文中, we specifically
focus on pronoun systems.
我
D
哦
w
n
哦
A
d
e
d
F
r
哦
米
H
t
t
p
:
/
/
d
我
r
e
C
t
.
米
我
t
.
e
d
你
/
我
我
/
n
G
A
r
t
我
C
e
–
p
d
我
F
/
/
/
/
5
3
2
2
9
5
2
0
1
4
2
4
4
/
我
我
n
G
_
A
_
0
0
4
0
6
p
d
.
F
乙
y
G
你
e
s
t
t
哦
n
0
7
S
e
p
e
米
乙
e
r
2
0
2
3
P E R S O N O F I N T E R E S T
321
1
2
3
2(西德:3)3 homophony
Pr: 14/200
Agr: 13/111
1
2
3
1(西德:3)2 homophony
Pr: 5/200
Agr: 11/111
1
2
3
1(西德:3)3 homophony
Pr: 1/200
Agr: 5/111
数字 7
Illustration of 1(西德:5)2, 2(西德:5)3, 和 1(西德:5)3 homophony patterns. Shades of gray indicate forms: cells with the same
shade use the same form. Typological counts for pronominal systems (Pr) are from Cysouw’s data set
(Baerman, 棕色的, and Corbett 2005, from Cysouw 2003), and counts for verbal agreement (Agr) are from
the Surrey Person Syncretism database (Corbett, 棕色的, and Baerman 2002, Baerman, 棕色的, and Corbett
2005).
在H&右 2002, 1(西德:5)2 homophony can arise through neutralization of the Addressee node, 和
1(西德:5)2(西德:5)3 homophony can arise through complete underspecification of the Participant distinction.
相比之下, there is no way for 1(西德:5)3 或者 2(西德:5)3 homophony patterns to be derived via feature
neutralization. This leads to the expectation that 1(西德:5)2 homophony will arise systematically and
will therefore be more common than both 1(西德:5)3 和 2(西德:5)3 homophony, which should in turn be
equally unlikely to arise (IE。, only through accidental homophony). This does not match up with
the typological counts in figure 7.
Ackema and Neeleman (A&氮) (2018) propose an alternative theory of person partly designed
to better account for the relative frequency of these homophony patterns. A&N redefine the person
space in terms of two privative person features, PROX (for proximate) and DIST (for distal). 在
line with Harbour (2016), they interpret these features as functions that operate on an input set
to deliver a subset as output. The crucial aspect of this account for our purposes is that the semantic
specification of these two features implies that one is shared by first and second person (PROX),
while the other is shared by second and third person (DIST).22 The immediate consequence is that
两个都 1(西德:5)2 和 2(西德:5)3 homophony can be generated, depending on which one of the two features
(PROX or DIST) is left underspecified. 相比之下, no feature is shared uniquely by first and third
人 (while excluding second person), ruling out the existence of a 1(西德:5)3 homophony pattern.23
A&N also predict an asymmetry between 1(西德:5)2 和 2(西德:5)3 homophony patterns (Peter Ackema,
pers. comm.). In their system, the second person category is the product of applying first PROX
22 In a nutshell, A&氮 (2018) assume that PROX and DIST features operate on a set containing all the possible sets
of person referents as nested subsets. 那是, the set containing all potential referents (Siuo) contains a subset containing
only speaker and addressee referents (Siu), which in turn contains a subset containing only speaker referents (和). 这
feature PROX operates on an input set and discards its outermost “layer,” whereas DIST selects this outermost layer. 这是
easy to see that in order to obtain, 例如, the first person referent one would need to apply the PROX feature twice,
as PROX(Siuo) (西德:4) (Siu) and PROX(Siu) (西德:4) 和.
23 A similar asymmetry can roughly be derived in Harbour’s (2016) 理论. Note that this is generally in line with
traditional accounts of person systems in which second and third person are grouped as a natural class under the feature
[(西德:6)speaker] (Forchheimer 1953, Noyer 1992, Pertsova 2011).
我
D
哦
w
n
哦
A
d
e
d
F
r
哦
米
H
t
t
p
:
/
/
d
我
r
e
C
t
.
米
我
t
.
e
d
你
/
我
我
/
n
G
A
r
t
我
C
e
–
p
d
我
F
/
/
/
/
5
3
2
2
9
5
2
0
1
4
2
4
4
/
我
我
n
G
_
A
_
0
0
4
0
6
p
d
.
F
乙
y
G
你
e
s
t
t
哦
n
0
7
S
e
p
e
米
乙
e
r
2
0
2
3
322
M O R A M A L D O N A D O A N D J E N N I F E R C U L B E R T S O N
and then DIST. If a language does not have a spell-out rule for this feature structure, it might in
principle collapse second person with first (based on the PROX feature) or with third (基于
the DIST feature). 然而, according to A&氮, there is a general principle (the “Russian Doll
Principle”) such that spell-out rules cannot apply to “inner” features without also mentioning
“outer” features, while the reverse is possible. Given that PROX is the innermost feature in the
second person structure, 2(西德:5)3 homophony patterns are predicted to be more common than 1(西德:5)2
那些.
总结一下, the crosslinguistic observation outlined above is well accounted for by
A&N’s approach (and Harbour’s (2016)), whereas it is problematic for H&R’s approach (和
similar ones). 然而, the typological data this generalization is built on are extremely sparse
and the magnitude of the differences is small: of approximately 200 languages sampled in Cysouw
2003, the most frequent 2(西德:5)3 homophony pattern is attested in the pronominal systems of only
14 语言, and the least common 1(西德:5)3 homophony pattern is attested in 1 (see Baerman, 棕色的,
and Corbett 2005:60). The limitations discussed above for typological data are therefore present
here in spades. But the issue is an important one: is the traditional and perhaps more intuitive
distinction between discourse participants and nonparticipants central to the organization of the
person space? Or is it one asymmetry among potentially many? 下文中, 我们使用
experimental paradigm developed above to bring learnability data to bear on this question.
Using the extrapolation paradigm (比照. Experiment 1B), we measured learners’ likelihood of
inferring each of the relevant patterns after training on an incomplete paradigm. Participants were
taught the meaning of two pronominal forms, which corresponded to a subset of person categories,
and they were then tested on how they extrapolated these forms to the remaining category. 为了
例子, some learners were taught two distinct forms for first and second person and then were
tested on which of those forms they used to express the held-out third person meaning. If they
used the first person form to express the new third person meaning, then they inferred a 1(西德:5)3
homophony pattern. A different pattern of extrapolation would indicate 2(西德:5)3 homophony, as de-
scribed in table 9. H&右 (2002) predict that learners will be more likely to infer 1(西德:5)2 homophony
relative to both 2(西德:5)3 和 1(西德:5)3 homophony. A&氮 (2018) predict that learners will be equally likely
to infer either 2(西德:5)3 或者 1(西德:5)2 homophony, but less likely to infer 1(西德:5)3 homophony.
4.2 方法
This experiment, including all hypotheses, 预测, and analyses, was preregistered at https://
osf.io/z9a35. All materials, 数据, and scripts can be found at https://osf.io/pf7q6/?view_only(西德:4)
96814a5352e54f91a3941b6895e02810. This experiment was implemented using the JavaScript
library jsPsych (De Leeuw 2015) and presented in a web browser.
4.2.1 Design Participants were randomly assigned to one of three conditions, summarized in
桌子 9. Conditions differed on which subset of two-person categories was used for training and
which category was held out. The training set determined which patterns of homophony partici-
pants could extrapolate to: Condition 1 was consistent with 1(西德:5)2 和 1(西德:5)3 图案, Condition 2
和 1(西德:5)2 和 2(西德:5)3, and Condition 3 和 2(西德:5)3 和 1(西德:5)3. The specific predictions of each account
我
D
哦
w
n
哦
A
d
e
d
F
r
哦
米
H
t
t
p
:
/
/
d
我
r
e
C
t
.
米
我
t
.
e
d
你
/
我
我
/
n
G
A
r
t
我
C
e
–
p
d
我
F
/
/
/
/
5
3
2
2
9
5
2
0
1
4
2
4
4
/
我
我
n
G
_
A
_
0
0
4
0
6
p
d
.
F
乙
y
G
你
e
s
t
t
哦
n
0
7
S
e
p
e
米
乙
e
r
2
0
2
3
P E R S O N O F I N T E R E S T
323
桌子 9
Conditions in Experiment 3. There were two training and one held-out (in boldface) categories per
状况. Each training category was mapped to a different pronominal form (A or B), schematically
represented with light and dark gray. Participants had to use the training forms they learned (A or B) 到
express the held-out meaning (cells with A or B?). The two rightmost columns state which of the
compatible paradigms participants were predicted to infer under Harley and Ritter’s (2002, H&右) 和
Ackema and Neeleman’s (2018, A&氮) 账户.
Mappings
Compatible
paradigms
H&右
A&氮
Condition 1
Condition 2
Condition 3
1
2
3
1
2
3
1
2
3
A or B?
A
乙
乙
A or B?
A
乙
A
A or B?
1(西德:3)2,
1(西德:3)3
1(西德:3)2,
2(西德:3)3
1(西德:3)3,
2(西德:3)3
1(西德:3)2 (西德:4) 1(西德:3)3
1(西德:3)2 (西德:4) 1(西德:3)3
1(西德:3)2 (西德:4) 2(西德:3)3
1(西德:3)2 (西德:2) 2(西德:3)3
2(西德:3)3 (西德:2) 1(西德:3)3
2(西德:3)3 (西德:4) 1(西德:3)3
given this design are summarized in the two rightmost columns of table 9. Note that both accounts
make the same predictions for Condition 1, but they differ in their predictions for Conditions 2
和 3.
All participants were additionally exposed to two pronominal forms that correspond to the
singular alternatives of the plural pronouns they were trained on. The person categories for singular
forms were always the same as the critical plural forms for a given participant, and were determined
按条件 (见表 9). 例如, participants in Condition 1 were additionally exposed
to second and third person singular pronouns.
4.2.2 Materials The language consisted of four different pronominal forms: two plural forms
(critical categories) and two singular forms (filler categories). These four lexical items were drawn
from the same list of eight CVC items used in Experiments 1 和 2. The reference of the pronouns
was expressed by highlighting a subset of family members, as in table 6, except that in this
experiment the inclusive category was never expressed. Visual stimuli were the same as in Experi-
评论 1 和 2.
4.2.3 Procedure After being introduced to the general backstory (see Experiment 1), 参与者
were instructed to figure out the meanings of the words in the new language. Participants were
given an example trial with an English pronoun (her or me depending on the condition) 那
would help them understand that the words they were learning were pronouns. As in the previous
实验, the speaker and addressee roles switched during the experiment to reinforce the
context-dependent meaning of the forms.
我
D
哦
w
n
哦
A
d
e
d
F
r
哦
米
H
t
t
p
:
/
/
d
我
r
e
C
t
.
米
我
t
.
e
d
你
/
我
我
/
n
G
A
r
t
我
C
e
–
p
d
我
F
/
/
/
/
5
3
2
2
9
5
2
0
1
4
2
4
4
/
我
我
n
G
_
A
_
0
0
4
0
6
p
d
.
F
乙
y
G
你
e
s
t
t
哦
n
0
7
S
e
p
e
米
乙
e
r
2
0
2
3
324
M O R A M A L D O N A D O A N D J E N N I F E R C U L B E R T S O N
The experiment had two training phases followed by a testing phase, the structure of which
was exactly as described for Experiment 1B. The only difference was in the person categories
instantiated by the highlighting (见图 8). Briefly, the two training phases were composed
of exposure and What if . . . 试验; the testing phase involved trials in which a referent set was
highlighted and participants had to choose the corresponding form. Participants were given feed-
back after exposure and What if . . . 试验, but not after testing trials. The order of presentation
of meanings was fully randomized within phases for each participant.
In the first training phase (16 试验), participants were trained on two singular (filler) 形式.
After this first training, participants were asked to type in a meaning for the two words they had
学到了, and they were given feedback on their answers. Unlike in Experiment 2, 参与者
were not excluded on the basis of their performance with these singular items in this phase, 自从
they were told what they meant and they were not used for extrapolation. In the second training
phase (28 试验), participants were trained on both filler and critical meanings. 最后, 在里面
testing phase the critical held-out meanings were added. 有 24 trials in this phase, 8 的
which were repetitions of the held-out meanings. As in previous experiments, participants com-
pleted a debrief questionnaire at the end of the experiment. A summary of the procedure is given
in the online appendix (table A.3).
4.2.4 Participants A group of 259 English-speaking adults (Condition 1: 74, Condition 2: 86,
Condition 3: 99) who had not participated in one of our previous experiments were recruited via
Amazon Mechanical Turk. Five additional participants were excluded for not being self-reported
native speakers of English, 和 30 for failing to pass the two attention checks included in the
实验: one at the beginning of the experiment (before starting the training), 和第二个
before starting with the testing. Per our preregistration, participants who did not pass both attention
checks did not contribute to our sample. Per preestablished exclusion criteria (as in Experiment
1), participants who failed to perform accurately in at least two-thirds of each training category
(4/6) during the testing phase were excluded from the analyses. The data from 152 参与者
were kept for the analyses (Condition 1: 48, Condition 2: 54, Condition 3: 50). All participants
were paid 2.5 USD for their participation, which lasted approximately 15 minutes.
4.3 结果
Recall that participants were taught two pronominal forms (coded as forms A and B), which they
had to use to describe both a critical set of two trained categories and a third, held-out meaning.
数字 9 shows the proportion of trials on which participants chose the pronominal form coded
as A during the testing phase, for each category and condition. For Condition 1, figure 9 节目
a mixed pattern of responses for the held-out meaning (first person plural): some participants
used the trained second person form (coded as A), some used the third person form (coded as
乙), and some behaved randomly (IE。, no consistent pattern of response). 相比之下, 参与者
in Conditions 2 和 3 appear to have inferred a consistent paradigm: 在这两种情况下, 参与者
largely used the same form (coded as A) for second and third person meanings (不管
what the trained meaning was).
我
D
哦
w
n
哦
A
d
e
d
F
r
哦
米
H
t
t
p
:
/
/
d
我
r
e
C
t
.
米
我
t
.
e
d
你
/
我
我
/
n
G
A
r
t
我
C
e
–
p
d
我
F
/
/
/
/
5
3
2
2
9
5
2
0
1
4
2
4
4
/
我
我
n
G
_
A
_
0
0
4
0
6
p
d
.
F
乙
y
G
你
e
s
t
t
哦
n
0
7
S
e
p
e
米
乙
e
r
2
0
2
3
P E R S O N O F I N T E R E S T
325
(A) Exposure 1PL category
我
D
哦
w
n
哦
A
d
e
d
F
r
哦
米
H
t
t
p
:
/
/
d
我
r
e
C
t
.
米
我
t
.
e
d
你
/
我
我
/
n
G
A
r
t
我
C
e
–
p
d
我
F
/
/
/
/
5
3
2
2
9
5
2
0
1
4
2
4
4
(乙) Testing 2PL category
数字 8
Illustration of exposure and testing trials in Experiment 3
To compare performance across conditions, we ran a logistic mixed-effects model predicting
form A choices in held-out trials by condition (3 级别). Recall that the meaning of the pronominal
form coded as A differed depending on condition (see Mapping in table 9). The model used
treatment coding, with Condition 2 as baseline, and included random by-participant intercepts.
The model revealed that participants in Condition 2 (基线) were significantly above
chance in selecting form A for the held-out category ((西德:4)(西德:4) 4.15, p (西德:5) .001***). In this condition,
the trained meaning of form A was third person, and the held-out category was second person;
所以, this result confirms that participants inferred a 2(西德:5)3 (rather than a 1(西德:5)2) homophony
pattern. In Condition 3, the trained meaning of form A was second person, and the held-out
category was third person. 所以, if participants in this condition were not significantly differ-
/
我
我
n
G
_
A
_
0
0
4
0
6
p
d
.
F
乙
y
G
你
e
s
t
t
哦
n
0
7
S
e
p
e
米
乙
e
r
2
0
2
3
326
M O R A M A L D O N A D O A N D J E N N I F E R C U L B E R T S O N
数字 9
Proportion of form A choices by condition during the testing phase. Choice of the same form across categories
indicates homophony. Error bars represent standard error on by-participant means; dots represent individual
participants’ means.
ent from the baseline, they presumably also inferred 2(西德:5)3 homophony, and to the same degree as
participants in Condition 2. This is confirmed by the model ((西德:4)(西德:4) 0.614, p (西德:4) .047). 相比之下,
the model revealed a significantly lower proportion of form A responses for held-out items in
Condition 1 compared with the baseline ((西德:4)(西德:4) (西德:6)4.8, p (西德:5) .001***). In this condition, the trained
meaning of form A was second person, and the held-out category was first person. 所以,
we can conclude that participants did not infer 1(西德:5)2 homophony to the same degree that participants
in Conditions 2 和 3 inferred 2(西德:5)3 homophony.
These results were further confirmed by two separate, intercept-only models for Conditions
1 和 3. For Condition 3, the proportion of form A responses was significantly above chance ((西德:4)
(西德:4) 3.78, p (西德:5) .001***), again indicating that as in Condition 2, Condition 3 participants inferred
A 2(西德:5)3 rather than a 1(西德:5)2 系统. 相比之下, the proportion of form A responses was not above
chance for Condition 1 ((西德:4)(西德:4) (西德:6)0.72; p (西德:4) .29), indicating that Condition 1 participants did not
show a clear preference between 1(西德:5)2 和 1(西德:5)3 homophony.24
4.4 讨论
Results from Experiment 3 show that participants consistently inferred paradigms that feature
2(西德:5)3 homophony whenever their training was compatible with this (Conditions 2 和 3). 那是,
24 This is further confirmed by a debrief questionnaire where participants were asked to provide information about
the meanings of each of the pronouns. Most participants in Conditions 2 和 3 reported that the second and third person
meanings were mapped into the same form.
我
D
哦
w
n
哦
A
d
e
d
F
r
哦
米
H
t
t
p
:
/
/
d
我
r
e
C
t
.
米
我
t
.
e
d
你
/
我
我
/
n
G
A
r
t
我
C
e
–
p
d
我
F
/
/
/
/
5
3
2
2
9
5
2
0
1
4
2
4
4
/
我
我
n
G
_
A
_
0
0
4
0
6
p
d
.
F
乙
y
G
你
e
s
t
t
哦
n
0
7
S
e
p
e
米
乙
e
r
2
0
2
3
P E R S O N O F I N T E R E S T
327
learners preferred systems that collapse the second and third person categories to alternative
homophony patterns. 什么时候 2(西德:5)3 homophony was not available to participants (Condition 1), 那里
was no stable pattern of responses: some participants used the same form for first and second
人, some used the same form for first and third, and some alternated randomly between the
二.
This result is in direct contrast to the prediction we derived from Harley and Ritter’s (2002)
理论. This theory is designed to derive 1(西德:5)2 homophony, 并不是 2(西德:5)3 homophony. 更远, 它
does not distinguish 2(西德:5)3 从 1(西德:5)3 homophony. Ackema and Neeleman’s (2018) theory fares
slightly better, as it posits that second and third person form a natural class (DIST) that excludes
first person. This predicts that 2(西德:5)3 homophony patterns should be preferred over 1(西德:5)3 图案, 在
accordance with our results (Forchheimer 1953, Bobaljik 2008; see also binary-features accounts).
然而, by also positing a feature shared between first and second person (PROX), this theory
predicts an asymmetry between 1(西德:5)2 和 1(西德:5)3 homophony, not attested in our findings (相似的
predictions are made by Harbour (2016)). 因此, while the pattern of results in our experiment is
quite clear, it does not straightforwardly match the predictions of either account.
Our results, 然而, do mirror what one might deem the most obvious typological asymme-
try in figure 7: paradigms featuring 2(西德:5)3 homophony are the most common (见表 9). Interest-
英利, an analogous pattern is found in the distribution of person systems across sign languages,
where second and third person are consistently homophonous both in pointing signs and in other
grammatical constructions (Meier 1990, Neidle et al. 2000), as well as in the spatial deixis domain
in languages across both modalities: most locative systems (例如, here/there in English) rely on
a distinction between spaces in the vicinity of the speaker (1) and spaces not in the vicinity of
the speaker (2(西德:5)3) (Harbour 2016).25 This match between our results and typology suggests the
possibility that some additional force is at play, which distinguishes first and second person, 甚至
if they do form a natural class.26
总结一下, recall that at issue here was the special status of the [(西德:3)参与者] distinction
in driving homophony. Theories based on the traditional set of binary features, like Harley and
Ritter’s (2002), are designed to capture this particular natural class, thus predicting systematic
25 原则, the English-speaking participants in our experiments might be extrapolating their experience with
the locatives here and there, which make a 2(西德:5)3 partition of the space, to the person space (Daniel Harbour, pers. comm.).
然而, in our view this would be rather surprising.
26 It is worth briefly discussing two alternative interpretations of our experimental results. 第一的, it could be that the
little sister was treated as a speech act participant rather than as a third person referent—for example, because she is
spatially close to the speaker and hearer in the illustrations (Peter Ackema, pers. comm.). To address this possibility, 我们
reran one condition of this experiment with a different set of images where the little sister was seated in the background
together with the parents. The results replicate the findings reported here, suggesting this is not an issue (see the online
appendix, figure A.3). 第二, participants might think the speaker (the girl with the hat) is in some cases directly
addressing the others (例如, parents) instead of the intended hearer (the girl asking the questions). If an addressee shift
of this sort were at play, there would be no “true” third person, explaining why participants derive what seems to be a
2(西德:5)3 pattern (as suggested by Klaus Abels, pers. comm.). 然而, in the postexperiment debrief, where participants had
to provide meanings for each of the words they had learned, most participants (across all three conditions) described the
pronominal form they used to refer to nonparticipants in the conversation (例如, the parents) as meaning them in English.
This shows that they were representing one of the pronominal forms as a true third person, and did not interpret the
speaker as directly addressing the others.
我
D
哦
w
n
哦
A
d
e
d
F
r
哦
米
H
t
t
p
:
/
/
d
我
r
e
C
t
.
米
我
t
.
e
d
你
/
我
我
/
n
G
A
r
t
我
C
e
–
p
d
我
F
/
/
/
/
5
3
2
2
9
5
2
0
1
4
2
4
4
/
我
我
n
G
_
A
_
0
0
4
0
6
p
d
.
F
乙
y
G
你
e
s
t
t
哦
n
0
7
S
e
p
e
米
乙
e
r
2
0
2
3
328
M O R A M A L D O N A D O A N D J E N N I F E R C U L B E R T S O N
homophony between speech act participants, 那是, first and second person. Our results do not
bear this out. 相当, behavior appears to be consistent with a preference to distinguish the speaker
来自别人的: person systems that distinguish between participants and nonparticipants—making
use of the PROX or [(西德:3)参与者] feature—seem to be dispreferred over systems that make use
of the DIST or [(西德:3)speaker] feature. The preference for 2(西德:5)3 systems in our experiment might then
lie in participants’ unwillingness to lose a distinction they have learned between forms that include
and forms that exclude the speaker.27 This bias for speaker distinctiveness was partly discussed
in the context of Zwicky’s generalization and Experiment 2 多于. We therefore return to this
issue in the next section.
5 General Discussion
The main aim of the experiments reported here was to bring behavioral evidence from learning
to bear on how person systems are represented and whether they are subject to universal constraints
on possible partitions. We see our results as making three main contributions: (A) we provide
confirmatory evidence that the person space is represented in terms of primitive features, (乙) 我们
point to the need for restrictions on patterns of inclusive homophony based on weak biases rather
than hard-and-fast constraints, (C) we provide new evidence for an asymmetry between 2(西德:5)3 和
other two-way homophony patterns involving 1 (EXCL(西德:5)INCL), 2, 和 3 person categories. We first
summarize each of these in turn and then discuss a number of broader issues raised by our results
as a whole.
In Experiment 1, we sought to provide evidence that the person space is represented as an
interacting set of universal features. While this idea forms the basis of most theories of person
in the theoretical linguistics literature, to date the link with learnability has been left implicit.
然而, these theories make testable predictions. If learners represent the person space as the
interaction of a (primitive) set of person features, then categories that form a natural class should
be readily mapped to the same phonological form (as formalized in Pertsova 2011, among many
其他的). The learner must simply determine which feature(s) are underspecified. A set of cells
that cannot be characterized as constituting a natural class arguably requires learners to first learn
the different categories (IE。, as feature bundles) and then independently learn to map the relevant
set to a phonological form. 因此, a feature-based theory of the person space predicts that paradigms
with this kind of random homophony should be less readily learned than homophony among
meanings with a shared feature. 相比之下, if learners perceive the person space as a simple
conjunction of four person categories (第一的, 包括的, 第二, 第三个), then there is no basis
for predicting that some homophony patterns should be more learnable than others.
The predictions of feature-based approaches to person were borne out in Experiment 1:
homophony patterns among forms sharing a feature were indeed easier to learn and were more
27 The extrapolation paradigm we use is designed to target precisely this kind of effect. An ease-of-learning design,
相比之下, may be more likely to tap purely into learners’ perceptions of how natural it is for two categories (例如, 第一的
and second persons) to share a form. For some evidence suggesting this, see https://osf.io/9nc36/?view_only(西德:4)09b711597
8e14ae0898d4b8fb47fb4b6.
我
D
哦
w
n
哦
A
d
e
d
F
r
哦
米
H
t
t
p
:
/
/
d
我
r
e
C
t
.
米
我
t
.
e
d
你
/
我
我
/
n
G
A
r
t
我
C
e
–
p
d
我
F
/
/
/
/
5
3
2
2
9
5
2
0
1
4
2
4
4
/
我
我
n
G
_
A
_
0
0
4
0
6
p
d
.
F
乙
y
G
你
e
s
t
t
哦
n
0
7
S
e
p
e
米
乙
e
r
2
0
2
3
P E R S O N O F I N T E R E S T
329
likely to be inferred when extrapolating a known form to a new meaning. This was the case even
though the specific features involved are not active in the participants’ native language, suggesting
that this is how these systems are represented. These findings draw an obvious parallel with work
in phonology showing that rules that can be characterized on the basis of features (or natural
类) are easier to learn than rules that group together a set of featurally distinct segments (例如,
Saffran and Thiessen 2003, Cristia` and Seidl 2008, Moreton and Pater 2012). This suggests that
feature-based representations are relevant for learning across domains, although the set of relevant
features is of course domain-specific.
In Experiment 2, we turned to Zwicky’s (1977) well-known observation that person partitions
without an inclusive distinction assimilate inclusive with first person (rather than second or third).
Different approaches to person treat this apparently strong asymmetry differently. Harbour’s
(2016) and Ackema and Neeleman’s (2018) theories are purpose-built to derive first-inclusive
tripartitions only (though Ackema and Neeleman’s may also be able to generate second-inclusive
homophony).28 然而, approaches like Harley and Ritter’s (2002) can derive both first-inclu-
sive and second-inclusive (though the latter is more marked).
再次, translating the predictions of these theories in terms of learnability, we tested whether
homophony of inclusive with first, 第二, or third person was treated distinctly by learners. 我们
found that third-inclusive homophony was particularly problematic, while first-inclusive had only
a weak advantage over second-inclusive. In accordance with Harley and Ritter’s (2002) 理论,
this suggests that there is no hard constraint against systems that systematically collapse second
and inclusive persons; 相当, inclusive can form a natural class with both first and second person,
虽然不是 (readily) with third.
最后, in Experiment 3 we provided new evidence for a learning-based asymmetry between
partitions with homophony between second and third person (2(西德:5)3) 一方面, and both
first-second (1(西德:5)2) and first-third (1(西德:5)3) homophony patterns on the other. While the former were
readily inferred by our participants, the latter were not. This result roughly matches an asymmetry
found in the typological distribution of pronominal systems. 然而, it is problematic under
the assumption that all homophony patterns targeting meanings that share a feature should be
equally possible. 的确, theories like Harley and Ritter’s (2002), which posit that first and second
person form a salient natural class—of speech act participants—would predict that 1(西德:5)2 homoph-
ony should if anything have a learnability advantage. Alternative approaches such as Harbour’s
(2016) and Ackema and Neeleman’s (2018) posit common features between both first and second
类别, and between second and third; 所以, their most straightforward prediction is a
specific learnability disadvantage for 1(西德:5)3 homophony. Participants in Experiment 3 instead consis-
tently inferred 2(西德:5)3 homophony whenever this was consistent with the input they were trained
在, 尽管 1(西德:5)2 和 1(西德:5)3 were equally dispreferred.
28 Ackema and Neeleman’s (2018) system does derive second-inclusive systems (violating Zwicky’s generalization)
每当 (A) PROX and PROX-PROX have different phonological realizations, 和 (乙) there is an impoverishment of DIST
in the plural when it is a dependent of PROX.
我
D
哦
w
n
哦
A
d
e
d
F
r
哦
米
H
t
t
p
:
/
/
d
我
r
e
C
t
.
米
我
t
.
e
d
你
/
我
我
/
n
G
A
r
t
我
C
e
–
p
d
我
F
/
/
/
/
5
3
2
2
9
5
2
0
1
4
2
4
4
/
我
我
n
G
_
A
_
0
0
4
0
6
p
d
.
F
乙
y
G
你
e
s
t
t
哦
n
0
7
S
e
p
e
米
乙
e
r
2
0
2
3
330
M O R A M A L D O N A D O A N D J E N N I F E R C U L B E R T S O N
总结一下, our results suggest that learners represent the person space in units that are
smaller than person categories (特征) and that instantiate natural classes. Natural-class-based
similarity therefore clearly plays an important role in determining how humans partition this
person space. 的确, a bias toward patterns based on natural class similarity pushes learners to
preferentially collapse categories that share features (when they are required to do so). Our results
support the claim that inclusive and second person should be included in the set of natural classes
in this domain. 然而, there is also reason to believe that a bias for keeping the speaker distinct
may also be at play. 的确, our findings are compatible with the idea that natural-class-based
相似, and speaker- (or ego-)based distinctiveness may be two independent forces influencing
the learnability of person systems. In Experiment 2, the partition characterized by first-inclusive
homophony was learned most readily and is consistent with both of these pressures: 它涉及
homophony among meanings that share a feature, [(西德:5)speaker] and keeps person categories impli-
cating the speaker distinct from other categories. The second-inclusive homophony pattern is the
next best, involving homophony among shared meanings but not maintaining speaker categories
as distinct from others. In Experiment 3, these two pressures led learners to infer 2(西德:5)3 每当
they could—again among the options presented to learners in this case, 2(西德:5)3 homophony is the
only one to involve both categories that are highly similar and forms that keep the speaker category
distinct. The fit is not perfect: 原则, as noted above, we might have expected the first-
inclusive advantage to be stronger, 和 1(西德:5)2 homophony to be preferred over 1(西德:5)3. One possibility
is that our failure to find these differences might be the result of particular features of our experi-
mental design. 例如, a difference in learnability between 1(西德:5)2 和 1(西德:5)3 might be revealed
in an ease-of-learning experiment (see Experiment 2), where resorting to the preferred 2(西德:5)3 pattern
is not possible.
In our discussion of Experiment 2, we suggested that a speaker distinctiveness bias may be
the result of the cognitive importance of the ego (see Dixon 1994). 的确, research on early
pronoun acquisition has argued for an egocentric bias, whereby children perceive the world as a
function of their presence in it and adopt categorization systems that carry this distinction (Charney
1980, Loveland 1984, Moyer et al. 2015). This is also supported by the fact that perspective-
taking appears to be a capacity that takes time to develop; infants and young children do not
always succeed on so-called theory-of-mind tasks that require them to recognize that their internal
knowledge states are not the same as other people’s (看, 例如, Ruffman 2014). If the pressure
for keeping the speaker distinct in pronoun systems comes from a general egocentric bias, 然后
we might expect (A) the bias to be stronger in children, 和 (乙) the bias to be stronger when
learners are active speakers of the language. The experiments reported here of course involved
adults, and our participants were never themselves the speaker. 然而, further research could
test both predictions.
An alternative is to build a speaker-based asymmetry between natural classes directly into
a theory of person—at the same level of representation as the domain-specific primitive features.
In a nutshell, the idea would be that the natural class “speaker” is somehow represented as special
within the person space. The pressure for speaker distinctiveness would then be a special type
of natural-class-based similarity that comes not directly from the set of primitive features but
我
D
哦
w
n
哦
A
d
e
d
F
r
哦
米
H
t
t
p
:
/
/
d
我
r
e
C
t
.
米
我
t
.
e
d
你
/
我
我
/
n
G
A
r
t
我
C
e
–
p
d
我
F
/
/
/
/
5
3
2
2
9
5
2
0
1
4
2
4
4
/
我
我
n
G
_
A
_
0
0
4
0
6
p
d
.
F
乙
y
G
你
e
s
t
t
哦
n
0
7
S
e
p
e
米
乙
e
r
2
0
2
3
P E R S O N O F I N T E R E S T
331
from the particular status of the speaker feature. This sketch is very speculative, and a worked-
out implementation is beyond the scope of this article. 然而, this idea is in the spirit of
Harbour (2016) and Ackema and Neeleman (2018), whose theories encode an inherent asymmetry
between speaker and addressee, built into the ontology (see footnote 15). 至关重要的是, we would
argue that this asymmetry should be treated as a bias, which shapes but does not strictly delimit
the space of possible partitions.
Before moving on, we should mention an anonymous reviewer’s concerns about whether
our results could be accommodated by a non-feature-based approach to person. This is a sensible
担心, as it questions whether our approach truly provides evidence that the person space is
represented in terms of a set of universal primitives. To address this concern, let us say a few words
about whether the results reported here are compatible with putative nonfeatural representations of
the person space.
Our experiments show that learners are sensitive to whether certain semantic values (例如,
speaker, addressee) are included in the pronominal reference, regardless of the specific person
类别. This suggests that person categories form natural classes along these semantic dimen-
西翁. 这里, we have treated these natural classes as defined by primitive features. 可以说, 一个
选择, non-feature-based approach would be one that does not decompose person categories
into smaller units, but treats them as primitives themselves; 例如, 第一的, 包括的, 第二,
and third would be four different primitives, as well as all other plural combinations (see Cysouw
2007 供讨论). We see a problem with this type of alternative view: in order to account
for our main findings, this approach would need to explain why it is that learners are more likely
to collapse some person categories than others, even when these categories are nonnative.
As a response to this issue, the same reviewer conjectures that some additional pragmatic
mechanism might result in more accurate learning predictions. For the time being, 然而, 我们
remain agnostic about how this proposal could be further developed into a full theory of person
that accounts for our results (and for the typology). 反而, we maintain that, to the extent that
learners are shown to be sensitive to the semantic overlap between categories, a feature-based
account of our results is more parsimonious (as well as in line with already established theories
of person).
6 结论
Person systems have been extensively explored from a theoretical standpoint: a number of ap-
proaches have been proposed, each of which constrains the set of possible person partitions that
humans can represent, with the aim of explaining the prevalence of certain partitions of the person
space crosslinguistically. The experiments reported here inform these theoretical approaches by
generating direct behavioral evidence for the impact of hypothesized representations and con-
straints on the learnability of different person partitions. 的确, our results constitute the first
experimental evidence for learnability differences in this domain.
具体来说, we have provided evidence that there is a universal basis for a set of primitives
organizing the person space that learners are sensitive to regardless of their native language.
Looking more closely into the nature of these primitive features, we have shown that a theory
我
D
哦
w
n
哦
A
d
e
d
F
r
哦
米
H
t
t
p
:
/
/
d
我
r
e
C
t
.
米
我
t
.
e
d
你
/
我
我
/
n
G
A
r
t
我
C
e
–
p
d
我
F
/
/
/
/
5
3
2
2
9
5
2
0
1
4
2
4
4
/
我
我
n
G
_
A
_
0
0
4
0
6
p
d
.
F
乙
y
G
你
e
s
t
t
哦
n
0
7
S
e
p
e
米
乙
e
r
2
0
2
3
332
M O R A M A L D O N A D O A N D J E N N I F E R C U L B E R T S O N
of the person space needs to account for the semantic similarity between inclusive and second
人, 一方面, and between second and third persons, 在另一. Each of these pairs
of categories was treated as a natural class by learners, suggesting that they have features in
常见的 (例如, an addressee feature possibly shared between second and inclusive persons). 是-
sides a preference for feature-based patterns, there was evidence across our experiments that
participants have an additional bias toward partitions of the person space where the speaker is
distinct from other categories. 因此, even though both speakers and addressees are participants
in the conversation, there is an inherent asymmetry in how learners treat them. We sketched two
possible accounts of this, the first in terms of a general egocentric bias, and the second in terms
of a special type of natural-class-based similarity.
更普遍, our results suggest that the experimental methods developed here provide
a novel tool for testing theoretically motivated questions about how languages carve up the person
空间. While these kinds of methods have gained traction in investigating learning biases in a
number of linguistic domains, here we have highlighted the sparsity of typological data as under-
scoring the need for new sources of evidence in building theories of person.
参考
Ackema, 彼得, and Ad Neeleman. 2013. Person features and syncretism. Natural Language and Linguistic
理论 31:901–950.
Ackema, 彼得, and Ad Neeleman. 2018. Features of person: From the inventory of persons to their morpho-
logical realization. 剑桥, 嘛: 与新闻界.
Baerman, 马修, and Dunstan Brown. 2013. Syncretism in verbal person/number marking. In The world
atlas of language structures online, 编辑. by Matthew S. Dryer and Martin Haspelmath. 莱比锡: Max
Planck Institute for Evolutionary Anthropology.
Baerman, 马修, Dunstan Brown, and Greville G. Corbett. 2005. The syntax–morphology interface: A
study of syncretism. 剑桥: 剑桥大学出版社.
Barwise, Jon, and Robin Cooper. 1981. Generalized quantifiers and natural language. Linguistics and Philoso-
物理层 4:159–219.
Bates, Douglas, Martin Maechler, Ben Bolker, and Steven Walker. 2014. Lme4: Linear mixed-effects models
using Eigen and S4. R包版本 1:1–23.
Be´jar, Susana. 2003. Phi-syntax: A theory of agreement. Doctoral dissertation, 多伦多大学.
Benveniste, Emile. 1971. Problems in general linguistics. Miami, FL: University of Miami Press.
Bickel, Balthasar. 2008. A refined sampling procedure for genealogical control. Language Typology and
Universals 61:221–233.
Boas, Franz. 1911. Introduction to the handbook of North American Indians. Smithsonian Institution Bulletin
40:1–83.
Bobaljik, Jonathan David. 2008. Missing persons: A case study in morphological universals. The Linguistic
审查 25:203–230.
棕色的, Cynthia A. 1997. Acquisition of segmental structure: Consequences for speech perception and second
language acquisition. Doctoral dissertation, 麦吉尔大学.
Carstensen, Alexandra, Jing Xu, Cameron Smith, and Terry Regier. 2015. Language evolution in the lab
tends toward informative communication. In Proceedings of the 37th Annual Meeting of the Cognitive
科学社, 编辑. by David C. 诺艾尔, Rick Dale, Anne S. Warlaumont, Jeff Yoshimi, Teenie
Matlock, Carolyn D. Jennings, and Paul P. Maglio, 303–308. Austin, TX: 认知科学学会.
我
D
哦
w
n
哦
A
d
e
d
F
r
哦
米
H
t
t
p
:
/
/
d
我
r
e
C
t
.
米
我
t
.
e
d
你
/
我
我
/
n
G
A
r
t
我
C
e
–
p
d
我
F
/
/
/
/
5
3
2
2
9
5
2
0
1
4
2
4
4
/
我
我
n
G
_
A
_
0
0
4
0
6
p
d
.
F
乙
y
G
你
e
s
t
t
哦
n
0
7
S
e
p
e
米
乙
e
r
2
0
2
3
P E R S O N O F I N T E R E S T
333
Charney, Rosalind. 1980. Speech roles and the development of personal pronouns. Journal of Child Language
7:509–528.
Chemla, Emmanuel, Brian Buccola, and Isabelle Dautriche. 2019. Connecting content and logical words.
Journal of Semantics 36:531–547.
Chomsky, Noam. 1965. Aspects of the theory of syntax. 剑桥, 嘛: 与新闻界.
Corbett, Greville G., Dunstan Brown, and Matthew Baerman. 2002. Surrey syncretism database. https://
www.smg.surrey.ac.uk/syncretism/.
Cowper, Elizabeth, and Daniel Currie Hall. 2009. Argumenthood, pronouns, and nominal feature geometry.
Determiners: Universals and Variation 147:97–120.
Cristia`, Alejandrina, and Amanda Seidl. 2008. Is infants’ learning of sound patterns constrained by phonologi-
cal features? Language Learning and Development 4:203–227.
Culbertson, Jennifer. 2012. Typological universals as reflections of biased learning: Evidence from artificial
language learning. Language and Linguistics Compass 6:310–329.
Culbertson, Jennifer. To appear. Artificial language learning. In Oxford handbook of experimental syntax,
编辑. by Jon Sprouse. 纽约: 牛津大学出版社.
Culbertson, Jennifer, and David Adger. 2014. Language learners privilege structured meaning over surface
频率. 美国国家科学院院刊 111:5842–5847.
Culbertson, Jennifer, Annie Gagliardi, and Kenny Smith. 2017. Competition between phonological and
semantic cues in noun class learning. 记忆与语言杂志 92:343–358.
Culbertson, Jennifer, Paul Smolensky, and Ge´raldine Legendre. 2012. Learning biases predict a word order
universal. 认识 122:306–329.
Cysouw, 迈克尔. 2003. The paradigmatic structure of person marking. 牛津: 牛津大学出版社.
Cysouw, 迈克尔. 2005. Quantitative methods in typology. In Quantitative Linguistik: Ein internationales
Handbuch, 编辑. by Reinhard Ko¨hler, Gabriel Altmann, and Rajmund G. Piotrowski, 554–578. 柏林:
De Gruyter Mouton.
Cysouw, 迈克尔. 2007. Building semantic maps: The case of person marking. In New challenges in typology,
编辑. by Matti Miestamo and Bernhard Wa¨lchli, 225–248. 柏林: De Gruyter Mouton.
Cysouw, 迈克尔. 2010. On the probability distribution of typological frequencies. In The mathematics of
语言, 编辑. by David Hutchison, Takeo Kanade, Josef Kittler, Jon M. Kleinberg, Friedemann
Mattern, John C. 米切尔, Moni Naor, 等人。, 29–35. 柏林: 施普林格.
Cysouw, 迈克尔. 2013. Inclusive/Exclusive distinction in independent pronouns. In The world atlas of
language structures online, 编辑. by Matthew S. Dryer and Martin Haspelmath. 莱比锡: Max Planck
Institute for Evolutionary Anthropology.
Daniel, 迈克尔. 2005. Understanding inclusives. In Clusivity, 编辑. by Elena Filimonova, 3–48. 阿姆斯特丹:
John Benjamins.
De Leeuw, Joshua R. 2015. jsPsych: A JavaScript library for creating behavioral experiments in a Web
browser. Behavior Research Methods 47:1–12.
狄克逊, 右. 中号. 瓦. 1994. Ergativity. 剑桥: 剑桥大学出版社.
Dunn, 迈克尔, Simon J. Greenhill, Stephen C. Levinson, and Russell D. Gray. 2011. Evolved structure of
language shows lineage-specific trends in word-order universals. 自然 473:79–82.
Fedzechkina, Maryia, Florian T. Jaeger, and Elissa L. Newport. 2012. Language learners restructure their
input to facilitate efficient communication. 美国国家科学院院刊 109:
17897–17902.
Forchheimer, 保罗. 1953. The category of person in language. 柏林: de Gruyter.
吉布森, 爱德华, Richard Futrell, Julian Jara-Ettinger, Kyle Mahowald, Leon Bergen, Sivalogeswaran Rat-
nasingam, Mitchell Gibson, Steven T. 匹安多糖, and Bevil R. 康威. 2017. Color naming across
languages reflects color use. 美国国家科学院院刊 114:10785–10790.
Greenberg, Joseph H. 1988. The first person inclusive dual as an ambiguous category. Studies in Language
12:1–18.
我
D
哦
w
n
哦
A
d
e
d
F
r
哦
米
H
t
t
p
:
/
/
d
我
r
e
C
t
.
米
我
t
.
e
d
你
/
我
我
/
n
G
A
r
t
我
C
e
–
p
d
我
F
/
/
/
/
5
3
2
2
9
5
2
0
1
4
2
4
4
/
我
我
n
G
_
A
_
0
0
4
0
6
p
d
.
F
乙
y
G
你
e
s
t
t
哦
n
0
7
S
e
p
e
米
乙
e
r
2
0
2
3
334
M O R A M A L D O N A D O A N D J E N N I F E R C U L B E R T S O N
黑尔, Kenneth. 1973. Person marking in Walbiri. In A festschrift for Morris Halle, 编辑. by Stephen R.
Anderson and Paul Kiparsky, 308–344. 纽约: 霍尔特, 莱因哈特和温斯顿.
Halle, 莫里斯, and Alec Marantz. 1994. Some key features of Distributed Morphology. In Papers on phonol-
ogy and morphology, 编辑. by Andrew Carnie and Heidi Harley, with Tony Bures, 275–288. 和
Working Papers in Linguistics 21. 剑桥, 嘛: 和, MIT Working Papers in Linguistics.
Hanson, Rebecca. 2000. Pronoun acquisition and the morphological feature geometry. Calgary Working
Papers in Linguistics 22. https://prism.ucalgary.ca/handle/1880/51442.
Hanson, Rebecca, Heidi Harley, and Elizabeth Ritter. 2000. Underspecification and universal defaults for
person and number features. In Actes du Congre`s annuel de l’ACL/CLA Annual Conference Proceed-
英格斯, 111–122. Ottawa: 渥太华大学, Cahiers Linguistiques d’Ottawa.
Harbour, Daniel. 2008. On homophony and methodology in morphology. Morphology 18:75–92.
Harbour, Daniel. 2011. Descriptive and explanatory markedness. Morphology 21:223–245.
Harbour, Daniel. 2014. Paucity, abundance, and the theory of number. 语言 90:185–229.
Harbour, Daniel. 2016. Impossible persons. 剑桥, 嘛: 与新闻界.
Harley, Heidi. 2008. When is a syncretism more than a syncretism? Impoverishment, metasyncretism, 和
underspecification. In Phi theory: Phi-features across modules and interfaces, 编辑. by Daniel Harbour,
David Adger, and Susana Be´jar, 251–294. 牛津: 牛津大学出版社.
Harley, Heidi, and Elizabeth Ritter. 2002. Person and number in pronouns: A feature-geometric analysis.
语言 78:482–526.
Ingram, 大卫. 1978. Typology and universals of personal pronouns. In Universals of human language.
卷. 3, Word structure, 编辑. by Joseph H. Greenberg, 查尔斯·A. Ferguson, and Edith A. Moravcsik,
213–248. 斯坦福大学, CA: 斯坦福大学出版社.
Kay, 保罗, and Terry Regier. 2007. Color naming universals: The case of Berinmo. 认识 102:289–298.
肯普, 查尔斯, and Terry Regier. 2012. Kinship categories across languages reflect general communicative
原则. 科学 336:1049–1054.
Ladd, D. 罗伯特, Sea´n G. 罗伯茨, and Dan Dediu. 2015. Correlational studies in typological and historical
语言学. Annual Review of Linguistics 1:221–241.
Loveland, Katherine A. 1984. Learning about points of view: Spatial perspective and the acquisition of
‘I/you’. Journal of Child Language 11:535–556.
Maldonado, Mora, and Jennifer Culbertson. 2019. Something about us: Learning first person pronoun sys-
特姆斯. In Proceedings of the 41st Annual Conference of the Cognitive Science Society, 编辑. by Ashok
Goel, Colleen Seifert, and Christian Freksa, 749–755. 蒙特利尔, QB: 认知科学学会.
Martı´, Luisa. 2020. Inclusive plurals and the theory of number. Linguistic Inquiry 51:37–74.
马丁, 亚历山大, and Sharon Peperkamp. 2020. Phonetically natural rules benefit from a learning bias:
A re-examination of vowel harmony and disharmony. Phonology 37:65–90.
马丁, 亚历山大, Theeraporn Ratitamkul, Klaus Abels, David Adger, and Jennifer Culbertson. 2019. 叉-
linguistic evidence for cognitive universals in the noun phrase. Linguistics Vanguard 5(1). https://
doi.org/10.1515/lingvan-2018-0072.
麦金尼斯, Martha. 2005. On markedness asymmetries in person and number. 语言 81:699–718.
Meier, Richard P. 1990. Person deixis in American Sign Language. Theoretical Issues in Sign Language
研究 1:175–190.
Moreton, Elliott. 2008. Analytic bias and phonological typology. Phonology 25:83–127.
Moreton, Elliott, and Joe Pater. 2012. Structure and substance in artificial-phonology learning. 第一部分: Struc-
真实. Language and Linguistics Compass 6:686–701.
Moyer, 摩根, Kaitlyn Harrigan, Valentine Hacquard, and Jeffrey Lidz. 2015. 2-year-olds’ comprehension
of personal pronouns. https://www.bu.edu/bucld/files/2015/06/Moyer.pdf.
Neidle, Carol, Judy Kegl, Dawn MacLaughlin, Benjamin Bahan, and Robert G. 李. 2000. The syntax of
American Sign Language: Functional categories and hierarchical structure. 剑桥, 嘛: 和
按.
我
D
哦
w
n
哦
A
d
e
d
F
r
哦
米
H
t
t
p
:
/
/
d
我
r
e
C
t
.
米
我
t
.
e
d
你
/
我
我
/
n
G
A
r
t
我
C
e
–
p
d
我
F
/
/
/
/
5
3
2
2
9
5
2
0
1
4
2
4
4
/
我
我
n
G
_
A
_
0
0
4
0
6
p
d
.
F
乙
y
G
你
e
s
t
t
哦
n
0
7
S
e
p
e
米
乙
e
r
2
0
2
3
P E R S O N O F I N T E R E S T
335
Nevins, 安德鲁, Cilene Rodrigues, and Kevin Tang. 2015. The rise and fall of the L-shaped morphome:
Diachronic and experimental studies. Probus 27:101–155.
Noyer, Rolf. 1992. Features, positions and affixes in autonomous morphological structure. Doctoral disserta-
的, 和.
Pagel, 标记, Quentin D. Atkinson, and Andrew Meade. 2007. Frequency of word-use predicts rates of
lexical evolution throughout Indo-European history. 自然 449:717–720.
Perfors, Amy, Joshua B. Tenenbaum, and Terry Regier. 2011. The learnability of abstract syntactic principles.
认识 118:306–338.
Pertsova, Katya. 2011. Grounding systematic syncretism in learning. Linguistic Inquiry 42:225–266.
匹安多糖, Steven T., and Edward Gibson. 2014. Quantitative standards for absolute linguistic universals.
认知科学 38:736–756.
匹安多糖, Steven T., Joshua B. Tenenbaum, and Noah D. 古德曼. 2016. The logical primitives of thought:
Empirical foundations for compositional cognitive models. 心理评论 123:392–424.
R核心团队. 2018. 右: A language and environment for statistical computing. R Foundation for Statistical
计算, 维也纳, 奥地利.
Rodrigues, Aryon D. 1990. You and I (西德:4) Neither you nor I: The personal system of Tupinamba. In Amazonian
语言学: Studies in lowland South American languages, 编辑. by Doris L. 佩恩, 393–405. Austin:
University of Texas Press.
Rouse, Steven V. 2015. A reliability analysis of Mechanical Turk data. Computers in Human Behavior
43:304–307.
Ruffman, Ted. 2014. To belief or not belief: Children’s theory of mind. Developmental Review 34:265–293.
藏红花, Jenny R., and Erik D. 蒂森. 2003. Pattern induction by infant language learners. Developmental
心理学 39:484–494.
Saldana, Carmen, Yohei Oseki, and Jennifer Culbertson. 2019. Do cross-linguistic patterns of morpheme
order reflect a cognitive bias? In Proceedings of the 41st Annual Conference of the Cognitive Science
社会, 编辑. by Ashok Goel, Colleen Seifert, and Christian Freksa, 994–1000. 蒙特利尔, QB: Cogni-
tive Science Society.
Sauerland, Uli, and Jonathan David Bobaljik. 2013. Syncretism distribution modeling: Accidental homoph-
ony as a random event. In Proceedings of GLOW in Asia IX, 编辑. by Nobu Goto, Koichi Otaki,
Atsushi Sato, and Kensuke Takita, 31–53. Tsu, 日本: Mie University.
施瓦茨, Linda J., and Timothy Dunnigan. 1986. Pronouns and pronominal categories in Southwestern
Ojibwe. In Pronominal systems, 编辑. by Ursula Wiesemann, 285–322. Tu¨bingen: Gunter Narr Verlag.
Silverstein, 迈克尔. 1976. Hierarchy of features and ergativity. In Grammatical categories in Australian
语言, 编辑. by R. 中号. 瓦. 狄克逊, 112–171. 堪培拉: Australian National University.
Sokolovskaja, Natalja K. 1980. Nekotorye semanti?eskie universalii v sisteme li?nyx mestoimenij. In Teorija
i tipologija mestoimenij, 编辑. by Igor F. Vardul, 84–103. 莫斯科: Nauka.
Sonnaert, Jolijn. 2018. The atoms of person in pronominal paradigms. 阿姆斯特丹: Netherlands Graduate
School of Linguistics (LOT).
Steinert-Threlkeld, Shane, and Jakub Szymanik. To appear. Ease of learning explains semantic universals.
认识.
Tabullo, Angel, Mariana Arismendi, Alejandro Wainselboim, Gerardo Primero, Sergio Vernis, Enrique
Segura, Silvano Zanutto, and Alberto Yorio. 2012. On the learnability of frequent and infrequent
word orders: An artificial language learning study. Quarterly Journal of Experimental Psychology
65:1848–1863.
托马斯, 大卫. 1955. Three analyses of the Ilocano pronoun system. Word 11:204–208.
Wechsler, 斯蒂芬. 2010. What ‘you’ and ‘I’ mean to each other: Person indexicals, self-ascription, 和
心理理论. 语言 86:332–365.
白色的, James. 2017. Accounting for the learnability of saltation in phonological theory: A maximum entropy
model with a P-map bias. 语言 93:1–36.
我
D
哦
w
n
哦
A
d
e
d
F
r
哦
米
H
t
t
p
:
/
/
d
我
r
e
C
t
.
米
我
t
.
e
d
你
/
我
我
/
n
G
A
r
t
我
C
e
–
p
d
我
F
/
/
/
/
5
3
2
2
9
5
2
0
1
4
2
4
4
/
我
我
n
G
_
A
_
0
0
4
0
6
p
d
.
F
乙
y
G
你
e
s
t
t
哦
n
0
7
S
e
p
e
米
乙
e
r
2
0
2
3
336
M O R A M A L D O N A D O A N D J E N N I F E R C U L B E R T S O N
Wilson, Colin. 2006. Learning phonology with substantive bias: An experimental and computational study
of velar palatalization. 认知科学 30:945–982.
Zaslavsky, Noga, Charles Kemp, Naftali Tishby, and Terry Regier. 2019. Color naming reflects both percep-
tual structure and communicative need. 认知科学主题 11:207–219.
Zwicky, Arnold M. 1977. Hierarchies of person. In Papers from the Thirteenth Regional Meeting, 芝加哥
Linguistic Society, 编辑. by Woodford A. Beach, Samuel E. 狐狸, and Shulamith Philosoph, 714–733.
芝加哥: 芝加哥大学, Chicago Linguistic Society.
Mora Maldonado
Centre for Language Evolution
爱丁堡大学
mora.maldonado@ed.ac.uk
Jennifer Culbertson
Centre for Language Evolution
爱丁堡大学
jennifer.culbertson@ed.ac.uk
我
D
哦
w
n
哦
A
d
e
d
F
r
哦
米
H
t
t
p
:
/
/
d
我
r
e
C
t
.
米
我
t
.
e
d
你
/
我
我
/
n
G
A
r
t
我
C
e
–
p
d
我
F
/
/
/
/
5
3
2
2
9
5
2
0
1
4
2
4
4
/
我
我
n
G
_
A
_
0
0
4
0
6
p
d
.
F
乙
y
G
你
e
s
t
t
哦
n
0
7
S
e
p
e
米
乙
e
r
2
0
2
3
