RESEARCH ARTICLE
Composition is the Core Driver of the
Language-selective Network
Francis Mollica1*, Matthew Siegelman2*, Evgeniia Diachek3, Steven T. Piantadosi4,
Zachary Mineroff 5, Richard Futrell6, Hope Kean7, Peng Qian7, and Evelina Fedorenko7,8,9
1School of Psychological Sciences, University of Melbourne
2Psychology Department, 哥伦比亚大学
3Psychology Department, Vanderbilt University
4Psychology Department, UC Berkeley
5The METALS Program, 卡内基梅隆大学
6Linguistics Department, UC Irvine
7Brain & Cognitive Sciences Department, 和
8McGovern Institute for Brain Research, 和
9Psychiatry Department, Massachusetts General Hospital
*Equal contributors.
关键词: 功能磁共振成像, compositionality, 语义学, syntax, information theory, mutual information
抽象的
The frontotemporal language network responds robustly and selectively to sentences. 但是
features of linguistic input that drive this response and the computations that these language
areas support remain debated. Two key features of sentences are typically confounded in
natural linguistic input: words in sentences (A) are semantically and syntactically combinable
into phrase- and clause-level meanings, 和 (乙) occur in an order licensed by the language’s
grammar. Inspired by recent psycholinguistic work establishing that language processing is
robust to word order violations, we hypothesized that the core linguistic computation is
作品, 和, 因此, can take place even when the word order violates the grammatical
constraints of the language. This hypothesis predicts that a linguistic string should elicit a
sentence-level response in the language network provided that the words in that string can
enter into dependency relationships as in typical sentences. We tested this prediction across
two fMRI experiments (total N = 47) by introducing a varying number of local word swaps
into naturalistic sentences, leading to progressively less syntactically well-formed strings.
Critically, local dependency relationships were preserved because combinable words
remained close to each other. As predicted, word order degradation did not decrease the
magnitude of the blood oxygen level–dependent response in the language network, 除了
when combinable words were so far apart that composition among nearby words was highly
不太可能. This finding demonstrates that composition is robust to word order violations,
and that the language regions respond as strongly as they do to naturalistic linguistic input,
providing that composition can take place.
介绍
A left-lateralized network of anatomically and functionally interconnected brain regions selec-
tively supports language processing (例如, Fedorenko, Behr, & Kanwisher, 2011). The regions
of this “language network” respond to both (A) word meanings and (乙) combinatorial semantic/
syntactic processing (例如, Bautista & Wilson, 2016; Fedorenko, Hsieh, Nieto-Castañón,
开放访问
杂志
引文: Mollica, F。, Siegelman, M。,
Diachek, E., 匹安多糖, S. T。, Mineroff,
Z。, 富特雷尔, R。, … Fedorenko, 乙. (2020).
Composition is the core driver of the
language-selective network.
Neurobiology of Language, 1(1), 104–134.
https://doi.org/10.1162/nol_a_00005
DOI:
https://doi.org/10.1162/nol_a_00005
支持信息:
https://doi.org/10.1162/nol_a_00005
已收到: 09 七月 2019
公认: 19 十二月 2019
利益争夺: 作者有
声明不存在竞争利益
存在.
Corresponding Authors:
Francis Mollica
mollicaf@gmail.com
Evelina Fedorenko
evelina9@mit.edu
处理编辑器:
Kate Watkins
版权: © 2020
麻省理工学院.
在知识共享下发布
归因 4.0 国际的 (抄送 4.0)
执照.
麻省理工学院出版社
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
n
哦
/
我
/
我
A
r
t
我
C
e
–
p
d
F
/
/
/
/
1
1
1
0
4
1
8
6
7
9
0
2
n
哦
_
A
_
0
0
0
0
5
p
d
.
/
我
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
Composition drives the language-selective network
Whitfield-Gabrieli, & Kanwisher, 2010; Fedorenko, Nieto-Castañón, & Kanwisher, 2012;
凯勒, Carpenter, & Just, 2001). The magnitude of neural responses in these regions, as mea-
sured with diverse brain imaging techniques, appears to scale with how language-like the input
是, with strongest responses elicited by sentences, and progressively lower responses elicited by
短语, lists of unconnected words, pseudowords, and foreign/indecipherable speech (例如, Bedny,
Pascual-Leone, Dodell-Feder, Fedorenko, & Saxe, 2011; Fedorenko et al., 2010; Fedorenko et al.,
2016; Hultén, 锄, Uddén, Lam, & Hagoort, 2019; Pallier, Devauchelle, & 德阿内, 2011;
斯科特, Gallée, & Fedorenko, 2017; Vagharchakian, Dehaene-Lambertz, Pallier, & 德阿内,
2012). But what features of the linguistic stimulus and what associated linguistic computations
drive the language network’s response? 尤其, sentences—its preferred stimulus—both
(A) contain word pairs that are semantically and syntactically combinable into phrases and clauses,
和 (乙) have the word order constrained by the rules of the language. Here we evaluate a hypoth-
esis that the core linguistic computation has to do with combining words into phrases and clauses,
and that this computation does not depend on word order (IE。, can take place even when the word
order is not licensed by the language’s grammar). A key prediction of this hypothesis is that a lin-
guistic string should elicit a sentence-level response in the language network, providing that the
words in that string are combinable.
The motivation for this hypothesis is twofold. 第一的, all languages reflect the structure of the
世界 (例如, 米科洛夫, 吸勺, 陈, 科拉多, & 院长, 2013; Pennington, Socher, &
曼宁, 2014), including both broad generalizations (例如, that properties can apply to ob-
jects or entities, that entities can engage in actions, or that some actions can affect objects) 和
particular contingencies (例如, which specific properties apply to which objects, which specific
entities engage in which actions, 等等). This knowledge of the world along with lexical
知识 (knowledge of word meanings) determines which words in the linguistic input are
combined to form phrases and clauses during language comprehension. 例如, 这
words tasty (a property, denoted by an adjective) and apple (an object, denoted by a noun)
are combinable into a phrase, in this case one with a plausible meaning, but the words tasty
and ate (an action, denoted by a past tense verb) cannot be combined because adjectives
are not typically dependents of verbs like taste. 相比之下, although many accounts of
syntactic representation and processing have emphasized word order as a key cue to build-
ing syntactic structures (例如, 贝弗, 1970; Kimball, 1973), languages across the world vary
widely in the rigidity of their word order constraints, with many languages exhibiting highly
flexible orderings, pointing to a more limited role of word order, at least in those languages
(例如, Dryer & Haspelmath, 2013; 黑尔, 1983; 杰肯道夫 & Wittenberg, 2014). 因此,
combinability of words into phrases and clauses, but not strict word order, appears to be a
universal feature of linguistic input that our language-processing mechanisms must be able
to handle.
第二个, recent work in psycholinguistics has shown that our sentence interpretation
mechanisms are well designed for coping with errors—including morphosyntactic agreement
errors and word swaps—providing that a plausible meaning is recoverable (例如, 费雷拉,
贝利, & 费拉罗, 2002; 吉布森, 卑尔根, & 匹安多糖, 2013; 征收, 2008乙; 征收, 比克内尔,
斯莱特里, & 雷纳, 2009; Traxler, 2014). These coping mechanisms are sufficiently pervasive
to interfere with our ability to detect errors during proofreading (例如, Schotter, 比克内尔,
霍华德, 征收, & 雷纳, 2014) and to make grammaticality judgments for sentences with
easily correctable syntactic errors compared to clearly grammatical/ungrammatical sentences
(Mirault, Snell, & Grainger, 2018). 因此, if the core linguistic computation implemented
in the language-selective cortex has to do with combining words into phrases and clauses,
form-based errors may be irrelevant, providing that they do not impede this process.
Neurobiology of Language
105
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
n
哦
/
我
/
我
A
r
t
我
C
e
–
p
d
F
/
/
/
/
1
1
1
0
4
1
8
6
7
9
0
2
n
哦
_
A
_
0
0
0
0
5
p
d
/
.
我
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
Composition drives the language-selective network
数字 1. A sample item from the critical experiment; colors are used to illustrate the increasing
degradedness (IE。, the color spectrum becomes progressively more discontinuous with more
swaps). A. The schematic of the procedure used to create the scrambled-sentence conditions in
实验 1. 乙. A sample stimulus from the ScrLowPMI condition in Experiment 2.
为了检验这个假设, we used a novel manipulation to examine neural responses to sen-
tences where word order is degraded (to varying extents) but local dependency relationships
are preserved. 尤其, naturalistic sentences were gradually degraded by increasing the
number of local word swaps (数字 1), which broke syntactic dependencies and led to pro-
gressively less syntactically well-formed strings (桌子 1). Critically, local semantic and syntac-
tic relationships were preserved. The degree of local combinability can be formally estimated
using tools from information theory (Shannon & Weaver, 1963). Naturalistic linguistic input is
characterized by relatively high pointwise mutual information (采购经理人指数) among words within a
local linguistic context, and it falls off for word pairs spanning longer distances (例如, 富特雷尔,
Qian, 吉布森, Fedorenko, & 空白的, 2019; 李, 1990; 林 & Tegmark, 2017). Our local-word-
swap manipulation maintained approximately the same level of local mutual information as
that observed in typical linguistic input. As can be seen in Figure 2e, the conditions with 1-, 3-,
5-, and even 7-word swaps (Scr1, Scr3, Scr5, and Scr7) have local PMI levels that are similar to
the intact condition (see Methods for details). To evaluate the importance of locality for
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
n
哦
/
我
/
我
A
r
t
我
C
e
–
p
d
F
/
/
/
/
1
1
1
0
4
1
8
6
7
9
0
2
n
哦
_
A
_
0
0
0
0
5
p
d
/
.
我
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
桌子 1. Description of the stimuli in Experiment 1 and 2a
Mean words moved
0 (0)
Proportion crossing a
syntactic boundary
0
Length of largest
meaningful substring
12 (0)
Int
Scr1
Scr3
Scr5
Scr7
1 (0)
2.49 (0.502)
3.39 (0.611)
4.29 (0.691)
70.7
77.3
83.3
87.0
8.47 (2.04)
5.82 (1.97)
4.62 (1.64)
4.46 (1.67)
106
Data are mean (标清).
Neurobiology of Language
Composition drives the language-selective network
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
n
哦
/
我
/
我
A
r
t
我
C
e
–
p
d
F
/
/
/
/
1
1
1
0
4
1
8
6
7
9
0
2
n
哦
_
A
_
0
0
0
0
5
p
d
.
/
我
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
A. The schematic of the language localizer task used to define the language-responsive areas. 乙. The parcels used to define the
数字 2.
language-responsive areas. In each participant, the top 10% of most localizer-responsive voxels within each parcel were taken as that par-
ticipant’s region of interest. Replicating prior work (Fedorenko et al., 2010), the localizer effect—estimated using across-runs cross-validation
to ensure independence—was highly robust in both experiments ( p’s < 0.0001). c, d. Neural responses (in % BOLD signal change relative to
fixation) to the conditions of the language localizer and Experiments 1 (n = 16) and 2 (n = 32). e. The formula for computing PMI (see Methods
for details), and average positive PMI values for the materials in Experiments 1 and 2. (N.B.: Slightly different scramblings of the materials for
the Scr1, Scr3, and Scr5 conditions were used in the two experiments; hence two bars (left = Experiment 1) for each of these conditions.) BOLD
— blood oxygen-level dependent, Int — intact, Scr — scrambled, PMI — pointwise mutual information.
building dependency relationships, in one condition (in Experiment 2), we scrambled words
within each sentence in a way so as to minimize local PMI and thus break local interword
relationships. In this condition, local PMI is comparable to that of a list of unconnected words
(see ScrLowPMI and Word-list conditions in Figure 2e). Participants read these materials—
presented one word at a time—while undergoing fMRI, and blood oxygenation level–dependent
(BOLD) responses were examined in language-selective regions defined using a separate localizer
task (Fedorenko et al., 2010; Figure 2a).
Neurobiology of Language
107
Composition drives the language-selective network
If the core function of the language-processing mechanisms is to combine words into
phrases and clauses, and this process is robust to word order violations, we would expect
the neural response to remain high as long as local PMI is similar to that observed in natural-
istic linguistic input, but to drop for the condition where local PMI is low. If, on the other hand,
composition critically depends on word order, such that it is hindered or altogether blocked in
cases where the word order violates the grammatical rules of the language, or if the core lin-
guistic computation has to do with word-order–based parsing, then we would expect the neu-
ral response to decrease as the word order becomes more degraded. It is also possible, based
on this hypothesis, that there would be a nonlinearity in the response across conditions, with
an increase for conditions with a small number of word swaps, which are relatively easily
correctable with the cost carried by the language areas, and then a drop for conditions with
a larger number of swaps.
To foreshadow the key results, we found that the fMRI BOLD response in the language
areas does not decrease relative to the response to its preferred stimulus (sentences), providing
that mutual information among nearby words remains as high as in typical linguistic input,
allowing for composition. However, scrambling a sentence so as to minimize local mutual
information, and thus block composition, leads to the response dropping to the level of that
for a list of unconnected words. These results support the idea that composition is the core
computation implemented in the language network, and that this computation is robust to
word order violations.
METHODS
Participants
Forty-seven individuals (age 18–48, average age 22.8; 31 female) participated for payment
(Experiment 1: n = 16; Experiment 2: n = 32; one individual participated in both Experiment 1
and Experiment 2, and one individual participated in Experiment 2 twice, once in version a and
once in version b, as described in subsequent text, for a total of 49 scanning sessions across the 47
participants. For the participant who participated in Experiment 2 twice, the data were combined
across the two sessions. We included twice as many participants in Experiment 2 to ensure that the
critical result in Experiment 1 was not due to insufficient power.) Forty-one participants were right-
handed, as determined by the Edinburgh handedness inventory (Oldfield, 1971), or by self-report;
the remaining six left-handed/ambidextrous individuals showed typical left-lateralized language
activation in the language localizer task (see Willems, van der Haegen, Fisher, & Francks,
2014, for arguments to include left-handers in cognitive neuroscience research). All participants
were native speakers of English from the Boston community. Four additional participants were
scanned (for Experiment 2) but were excluded from the analyses due to excessive head motion
or sleepiness, and/or failure to perform the behavioral task. All participants gave written
informed consent in accordance with the requirements of MIT’s Committee on the Use of
Humans as Experimental Subjects.
Experimental Design and Materials
In both experiments, each participant completed (a) a version of the language localizer task
(see Figure 2a; Fedorenko et al., 2010), which was used to identify language-responsive areas
at the individual-subject level, and (b) the critical sentence comprehension task (in 30 of 49
scanning sessions, participants completed the localizer task in the same session as the critical
task, and for the remaining 19 sessions, the localizer came from an earlier session; see
Mahowald & Fedorenko, 2016, for evidence of the stability of the localizer activation maps
Neurobiology of Language
108
l
D
o
w
n
o
a
d
e
d
f
r
o
m
h
t
t
p
:
/
/
d
i
r
e
c
t
.
m
i
t
.
e
d
u
n
o
/
l
/
l
a
r
t
i
c
e
-
p
d
f
/
/
/
/
1
1
1
0
4
1
8
6
7
9
0
2
n
o
_
a
_
0
0
0
0
5
p
d
/
.
l
f
b
y
g
u
e
s
t
t
o
n
0
7
S
e
p
e
m
b
e
r
2
0
2
3
Composition drives the language-selective network
across sessions). In addition, each participant completed a spatial working memory task
(Fedorenko et al., 2011), used in some control analyses to characterize brain regions that
are sensitive to sentence scrambling, as described in subsequent text. Some participants further
completed one or two additional tasks for unrelated studies. The language localizer task was
always completed first; the order of all other tasks varied across participants. The entire scan-
ning session lasted ∼2 hr.
Language localizer
Participants passively read sentences and lists of pronounceable nonwords in a blocked de-
sign. The Sentences > Nonwords contrast targets brain regions sensitive to high-level linguistic
加工 (Fedorenko et al., 2010). The robustness of this contrast to materials, modality of
presentation, 语言, and task has been established previously (Fedorenko et al., 2010;
Fedorenko, 2014; Scott et al., 2017). In this version of the localizer, the sentences were con-
structed to vary in content and structures used, and the nonwords were created using the
Wuggy software (Keuleers & Brysbaert, 2010), to match the phonotactic properties of the non-
words to those of the words used in the Sentence condition. Each trial started with 100 多发性硬化症
pretrial fixation, followed by a 12-word-long sentence or a list of 12 nonwords presented
on the screen, one word/nonword at a time, at the rate of 450 ms per word/nonword. 然后,
a line drawing of a hand pressing a button appeared for 400 多发性硬化症, and participants were in-
structed to press a button whenever they saw this icon, and finally a blank screen was shown
为了 100 多发性硬化症, for a total trial duration of 6 s. The simple button-press task was included to help
participants stay awake and focused. Each block consisted of three trials and lasted 18 s. 每个
run consisted of 16 experimental blocks (8 per condition), 和 5 fixation blocks (14 s each), 为了
a total duration of 358 s (5 min 58 s). Each participant performed two runs. Condition order
was counterbalanced across runs.
Spatial working memory task (used in some control analyses)
Participants had to keep track of four (easy condition) or eight (hard condition) locations in a 3 ×
4 grid (Fedorenko et al., 2011). In both conditions, participants performed a two-alternative,
forced-choice task at the end of each trial to indicate the set of locations that they had just seen.
The Hard > Easy contrast targets brain regions that are sensitive to general executive demands
(例如, Duncan, 2010; Duncan & 欧文, 2000). Fedorenko, Duncan, and Kanwisher (2013) (看
also Hugdahl, Raichle, 米特拉, & Specht, 2015) have shown that the regions activated by this task
are also activated by a wide range of tasks that contrast a difficult versus an easier condition.
Each trial lasted 8 s (see Fedorenko et al., 2011, 欲了解详情). Each block consisted of four trials
and lasted 32 s. Each run consisted of 12 experimental blocks (6 per condition), 和 4 fixation
blocks (16 s each), for a total duration of 448 s (7 min 28 s). Forty-five participants performed two
runs; the remaining two participants performed one run. Condition order was counterbalanced
across runs when participants performed two runs.
Critical task in Experiment 1
Design and materials. Participants read sentences with correct word order (Intact [Int]) 和
sentences with progressively more scrambled word orders created by an increasing number
(之间 1 和 7) of local word swaps (Scrambled [Scr] 1, 3, 5, 和 7; 见图 1), 还有
as two control conditions: lists of unconnected words and lists of nonwords. At the end of each
审判, participants were presented with a word (in the sentence and word-list conditions) 或一个
nonword (in the nonword-list condition) and asked to decide whether this word/nonword
appeared in the preceding trial.
Neurobiology of Language
109
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
n
哦
/
我
/
我
A
r
t
我
C
e
–
p
d
F
/
/
/
/
1
1
1
0
4
1
8
6
7
9
0
2
n
哦
_
A
_
0
0
0
0
5
p
d
.
/
我
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
Composition drives the language-selective network
To create the sentence materials, we extracted 150, 12-word-long sentences from the
British National Corpus (BNC; Burnard, 2000). We then permuted the word order in each sen-
tence via local swaps, to create the scrambled conditions. 尤其, a word was chosen at
random and switched with one of its immediate neighbors. This process was repeated a spec-
ified number of times. Because one random swap can directly undo a previous swap, 我们
ensured that the manipulation was successful by calculating the edit distance. (The code used
to create the scrambled conditions is available at Open Science Framework [OSF; Mollica
等人。, 2019]: https://osf.io/y28fz/.) We chose versions with 1, 3, 5, 和 7 swaps in order to
limit the number of sentence conditions to five, 尽管, 同时, covering a range of
degradedness levels. The materials thus consisted of 150 sentences with five versions each (Int,
Scr1, Scr3, Scr5, and Scr7), for a total of 750 strings. These were distributed across five exper-
imental lists following a Latin square design, so that each list contained only one version of a
sentence and 30 trials of each of the five conditions. Any given participant saw the materials
from just one experimental list, and each list was seen by two to four participants.
To characterize the sentence materials in greater detail, as critical for interpretation (看
Discussion section), we performed three analyses on the materials used in Experiments 1
and 2a (见表 1). 第一的, we manually annotated the number of words that were moved
in each scrambled condition (where a move is defined as a rightward or leftward movement
of a word across one or more words). 例如, if the dog chased the cat was scrambled by
three swaps to dog the cat chased the, two words (the and cat) have moved; and if it was
scrambled by three swaps to the chased the cat dog, only one word (狗) has moved. As ex-
pected, this value increased gradually from the least to the most scrambled condition (IE。,
从 1 in the Scr1 condition to 4.29 in the Scr7 condition), suggesting that there were more
opportunities to break syntactic dependencies as the number of swaps increased. 第二, 我们
manually annotated the stimuli for the proportion of swaps that crossed a constituent boundary
in the original sentence. This number increased gradually from 70.7% in the Scr1 condition to
87% in the Scr7 condition. This analysis ensures that the scrambling procedure broke syntactic
dependencies, even in the condition with a single swap, and did not simply swap words
within constituents. And finally, we computed the length of the largest contiguous grammatical
and meaningful substring (whether or not that substring was present in the original sentence).
This value decreased gradually from 8.47 words in the Scr1 condition to 4.46 words in the
Scr7 condition.
The word-list condition consisted of sequences of 12 real words (173 unique words: 55.5%
nouns, 15.6% 动词, 22.5% 形容词, 和 6.4% adverbs; average word length: 7.19 河粉-
nemes, 标准差 [标清] 1.43 [Weide, 1998]; average log frequency: 1.73, 标清 0.80
[Brysbaert, 新的, & Keuleers, 2012]), and the nonword-list condition consisted of sequences
的 12 nonwords (there were actually four different nonword-list conditions—a manipulation
not of interest to the current study; we averaged the responses across the four nonword-list
conditions in the analyses). The nonwords used in this experiment were generated differently
from the nonwords used in the language localizer task. 尤其, they were created from
real words by introducing some number of letter replacements keeping local phonotactics
intact. We do not make any direct comparisons between nonword conditions across
实验, so this difference is of no consequence. The word-list and nonword-list materials
were the same across participants. (All the materials are available at OSF [Mollica et al.,
2019].)
Computing mutual information values. To estimate the likelihood of dependencies among nearby
字, we used pointwise mutual information (or PMI), a metric from information theory
Neurobiology of Language
110
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
n
哦
/
我
/
我
A
r
t
我
C
e
–
p
d
F
/
/
/
/
1
1
1
0
4
1
8
6
7
9
0
2
n
哦
_
A
_
0
0
0
0
5
p
d
.
/
我
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
Composition drives the language-selective network
(教会 & Hanks, 1990; Fano, 1961), which measures the mutual dependence between
变量 (在这种情况下, 字). Positive PMI values suggest a dependence between words
based on their overlap in contexts of use. Negative and near-zero PMI values suggest the
absence of a dependence. Following word2vec (Mikolov et al., 2013), we used a sliding
four-word window to extract local word pairs from each 12-word string. This is equivalent to
collecting the bigrams, 1-skip-grams, and 2-skip-grams from each string.
For each word pair, we calculated PMI as follows:
(西德:1)
PMI wi; wj
(西德:3)
¼ log
(西德:3)
(西德:1)
P wj; wi
(西德:1)
ÞP wj
ð
P wi
(西德:3)
Probabilities were estimated using the Google N-gram corpus (Michel et al., 2010) and ZS
Python library (史密斯, 2014) with Laplace smoothing (α = 0.1). For each 12-word string, 我们
averaged across the positive PMI values for all word pairs occurring within a four-word sliding
window. (The code for computing PMI is available at OSF [Mollica et al., 2019].) Although PMI
encompasses both semantic and syntactic dependence, it down-weighs the contribution of high
频率, closed-class words, like determiners, pronouns, and prepositions, given that it reflects
interword association beyond the simple frequency of co-occurrence. As can be seen in
Figure 2e, local PMI across the four scrambled conditions (Scr1, Scr3, Scr5, and Scr7) is as high
as that in the intact (Int) 状况. (Given that the sentences in the intact condition were drawn
from a corpus, their local PMI values likely reflect average local PMI in typical linguistic input.)
This operationalization is a coarse measure that collapses over finer-grained distinctions that may
affect the formation of semantic and syntactic dependencies (例如, Bemis & Pylkkänen, 2011;
Pylkkänen, Bemis, & Elorrieta, 2014; also see Pylkkänen, 2016, 2019, for reviews). 然而,
to the extent that this operationalization can account for patterns of neural (在这种情况下,
大胆的) responses and thus yield insights about the workings of the language system, it holds the-
oretical and empirical value.
程序. Participants read sentences, scrambled sentences, word lists, and nonword lists in
an event-related fMRI design. Each trial lasted 8 s and consisted of the presentation of the
刺激 (a sequence of 12 words/nonwords presented one at a time in the center of the
screen with no punctuation, 为了 500 ms each, in black capital letters on a white background),
followed by a blank screen for 300 多发性硬化症, followed by a memory probe presented in blue font for
1,200 多发性硬化症, followed again by a blank screen for 500 多发性硬化症. The memory probe came from the
preceding stimulus on half of the trials. For the sentences, the probes were uniformly
distributed across the beginning (first four words), 中间 (middle four words), or end (最后的
four words) of the sentence; for the word and nonword lists, the probes were uniformly
distributed across the 12 positions. Incorrect probes were the shuffled correct probes from
other sequences in the same condition.
The trials in each experimental list (300 全部的; 30 trials per condition, where the conditions
included the intact sentence condition, four scrambled sentence conditions, the word-list con-
迪迪, and four nonword-list conditions) were divided into six subsets corresponding to six
runs. Each run lasted 480 s (8 min) and consisted of 8 s * 50 试验 (5 per condition) 和 80 s
of fixation. The optseq2 algorithm (戴尔, 1999) was used to create condition orderings and to
distribute fixation among the trials so as to optimize our ability to de-convolve responses to the
different conditions. Condition order varied across runs and participants. Most participants (n =
13) performed five runs; the remaining three participants performed four or three runs due to
time constraints.
Neurobiology of Language
111
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
n
哦
/
我
/
我
A
r
t
我
C
e
–
p
d
F
/
/
/
/
1
1
1
0
4
1
8
6
7
9
0
2
n
哦
_
A
_
0
0
0
0
5
p
d
.
/
我
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
Composition drives the language-selective network
Critical task in Experiment 2
Design and materials. 实验 2 was designed both (A) to assess the robustness of the results
in Experiment 1, in line with increasing emphasis on replicability in cognitive neuroscience
(例如, Poldrack et al., 2017; Siegelman, 空白的, Mineroff, & Fedorenko, 2019; Uddén et al., 2019),
和 (乙) to directly evaluate the locality constraint on semantic composition. 尤其, 作为
discussed in the preceding text, in typical linguistic input, semantic and syntactic dependencies
tend to be local (例如, 富特雷尔, Mahowald, & 吉布森, 2015). 因此, our linguistic processing
mechanisms are plausibly optimized for building complex meanings within local linguistic
上下文. 例如, if words tasty and apple occur within the same sentence, but are
separated by eight other words, we may be less likely to combine them then we would in
cases where tasty and apple occur in proximity to one another. To evaluate the importance of
locality for the engagement of the composition mechanisms, we included a manipulation
where words were scrambled within a sentence in a way that minimizes local PMI. If locality
很重要, this condition should elicit a lower neural response compared to the conditions
with high local PMI because participants would not be engaging in composition.
As in Experiment 1, participants read sentences with correct word order (Int) and sentences
with progressively more scrambled word orders (Scr 1, 3, 和 5). The materials for these
scrambled conditions were identical to those in Experiment 1 for half of the participants,
and different permutations of the same intact stimuli for the other half. 因为, as expected,
the results were almost identical across these two versions of the materials, we report the re-
sults for all participants together.
The condition with seven word swaps (Scr7) was replaced by a condition in which each
pair of nearby content words was separated as much as possible within the 12-word string, 所以
as to minimize local mutual information (见图 1). We focused on separating nearby con-
tent words because those carry the most information in the signal (Shannon & Weaver, 1963)
and contribute to positive PMI values, as noted earlier. 拿, 例如, one of our intact
句子: Larger firms and international companies tended to offer the biggest pay rises. 第一的,
the content words were given a fixed order that maximized the sum of the distances between
adjacent content words (two content words are considered adjacent in the original string if
they have no content words between them): 例如, larger international tended biggest
rises firms companies offer pay. This process was repeated for the function words (例如, 和
到). 然后, the ordered function words were embedded in the center of the ordered content
字 (IE。, larger international tended biggest rises and the to firms companies offer pay),
which maximizes the distances between adjacent content words in the original sentence.
(The code is available at OSF [Mollica et al., 2019].) The manipulation was effective, leading
to a significant drop in local mutual information (see Figure 2e). If locality is important for building
interword relationships, then minimizing the likelihood of dependency formation within local
contexts should lead to a drop in the neural response, similar to what is observed during the
processing of unconnected word-lists (例如, Fedorenko et al., 2010; Pallier et al., 2011).
In addition to the five sentence conditions, we included five word-list conditions that were
matched in terms of their lexical properties word-for-word to the sentence conditions. In par-
针状的, 每一个 876 unique words in the sentence conditions was replaced by a different word
of the same syntactic category (using the following set: nouns, 动词, 形容词, adverbs, 和
closed-class words), similar in length (± 0.03 phonemes, 一般 [Weide, 1998]) and fre-
quency (± 0.23 log frequency (lf ), 一般 [Brysbaert et al., 2012]). (Due to a script error,
11 words were replaced by the same word as the original word, 和 6 words were replaced by
a word of a different part of speech.) We included the same number of word-list conditions as
sentence conditions to match the distribution of sentence and word-/nonword-list conditions
Neurobiology of Language
112
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
n
哦
/
我
/
我
A
r
t
我
C
e
–
p
d
F
/
/
/
/
1
1
1
0
4
1
8
6
7
9
0
2
n
哦
_
A
_
0
0
0
0
5
p
d
/
.
我
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
Composition drives the language-selective network
in Experiment 1. 然而, in the analyses, we averaged the responses across the five word-list
状况, given that there is no reason to expect differences among them.
The materials were distributed across five experimental lists; any given participant saw the ma-
terials from just one list, and each list was seen by five to seven participants. As in Experiment 1, 在
the end of each trial, participants were presented with a word and asked to decide whether this
word appeared in the preceding trial (see Results for behavioral performance).
程序. The procedure was identical to that in Experiment 1 except that the memory probe
was uniformly distributed across the 12 positions in every condition. Most participants (n = 30)
performed five or six runs; the remaining two participants performed four or three runs due to
time constraints.
fMRI data acquisition
Structural and functional data were collected on the whole-body 3 Tesla Siemens Trio scanner
with a 32-channel head coil at the Athinoula A. Martinos Imaging Center at the McGovern
Institute for Brain Research at MIT. T1-weighted structural images were collected in 179 sag-
ittal slices with 1 mm 各向同性体素 (repetition time [TR] = 2,530 多发性硬化症, echo time [TE] = 3.48 多发性硬化症).
Functional, BOLD data were acquired using an echo planar imaging sequence (with a 90° flip
angle and using Generalized Autocalibrating Partially Parallel Acquisitions [GRAPPA] with an ac-
celeration factor of 2), with the following acquisition parameters: 31 4-mm-thick near-axial slices,
acquired in an interleaved order with a 10% distance factor; 2.1 mm × 2.1 mm in-plane resolu-
的; field of view of 200 mm in the phase encoding anterior to posterior (A > P) 方向; 矩阵
大小 96 mm × 96 毫米; TR of 2,000 多发性硬化症; and TE of 30 多发性硬化症. Prospective acquisition correction
(Thesen, Heid, Mueller, & Schad, 2000) was used to adjust the positions of the gradients based on
the participant’s motion one TR back. 首先 10 s of each run were excluded to allow for steady-
state magnetization.
fMRI data preprocessing and first-level analysis
First-level analyses were conducted in Statistical Parametric Mapping 5 (SPM5) (we used an
older version of the software here due to the use of these data in other projects spanning many
years and hundreds of subjects); critical second-level analyses were performed using custom
MATLAB and R scripts. Data from each participant were motion corrected (realignment to the
mean image using second-degree b-spline interpolation) and normalized into a common brain
空间, the Montreal Neurological Institute (MNI) template (normalization was estimated for
the mean image using trilinear interpolation) and resampled into 2 mm 各向同性体素. 这
data were then smoothed with a 4 mm Gaussian filter and high-pass filtered (在 200 s). The task
effects in both the language localizer task and the critical experiment were estimated using a
General Linear Model (GLM) in which each experimental condition was modeled with a box-
car function (corresponding to a block or event) convolved with the canonical hemodynamic
response function (HRF). The model also included first-order temporal derivatives of these ef-
fects, as well as nuisance regressors representing entire experimental runs and offline-estimated
motion parameters.
Language functional region of interest definition and response estimation
For each participant, functional regions of interest (fROIs) were defined using the Group-
constrained Subject-Specific (GSS) 方法 (Fedorenko et al., 2010; Julian, Fedorenko,
韦伯斯特, & Kanwisher, 2012), whereby a set of parcels or “search spaces” (IE。, brain areas within
Neurobiology of Language
113
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
n
哦
/
我
/
我
A
r
t
我
C
e
–
p
d
F
/
/
/
/
1
1
1
0
4
1
8
6
7
9
0
2
n
哦
_
A
_
0
0
0
0
5
p
d
.
/
我
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
Composition drives the language-selective network
which most individuals in prior studies showed activity for the localizer contrast) is combined
with each individual participant’s activation map for the same contrast. To define the language
fROIs, we used six parcels (Figure 2b) derived from a group-level representation of data for the
Sentences > Nonwords contrast in 220 参与者 (a set of participants scanned in our lab).
These parcels included three regions in the left frontal cortex: two located in the inferior frontal
gyrus (LIFG and LIFGorb), and one located in the middle frontal gyrus (LMFG); and three regions
in the left temporal and parietal cortices spanning the entire extent of the lateral temporal lobe
and extending into the angular gyrus (LAntTemp, LPostTemp, and LAngG). (These parcels were
similar to the parcels reported originally in Fedorenko et al. [2010], except that the two anterior
temporal regions were collapsed together, and the two posterior temporal regions were collapsed
together.) Following much prior work in our group, individual fROIs were defined by selecting—
within each parcel—the top 10% of most localizer-responsive voxels based on the t-values for the
Sentences > Nonwords contrast. Responses (in percent BOLD signal change units) to the relevant
critical experiment’s conditions, relative to the fixation baseline, were then estimated in these
fROIs. So the input to the critical statistical analyses consisted of—for each participant—a value
(percent BOLD signal change) for each of 10 conditions in each of the six language fROIs.
此外, for Experiment 1, responses were averaged across the four nonword-list conditions,
leaving a total of seven conditions; and for Experiment 2, responses were averaged across the five
word-list conditions, leaving a total of six conditions. In the critical analyses (Figure 2c, d), 我们
consider the language network as a whole (treating regions as random effects; see fMRI data in
实验 1 和 2 in Results) given the abundant evidence that the regions of this network form
anatomically (例如, Axer, Klingner, & Prescher, 2013; Saur et al., 2008) and functionally integrated
系统, as evidenced by strong interregional correlations during rest and language comprehen-
锡安 (例如, 空白的, Kanwisher, & Fedorenko, 2014; Paunov, 空白的, & Fedorenko, 2019) 并由
correlations in effect sizes across the regions (Mineroff, 空白的, Mahowald, & Fedorenko,
2018), but we also report the individual profiles of the six language fROIs and associated statistics
(数字 3 and fMRI data in Experiments 1 和 2 in Results). 此外, to facilitate comparisons
with other data sets, we include for all individual participants, the whole-brain contrast maps for
all the individual conditions relative to the fixation baseline on OSF (Mollica et al., 2019).
Statistical tests
To compare the average change in BOLD response across conditions, we conducted a mixed-
effect linear regression model with maximal random effect structure (Barr, 征收, Scheepers, &
Tily, 2013), predicting the level of response with a fixed effect and random slopes for
Condition, and random effects for ROI and Participant. To further compare the average change
in BOLD response across conditions in each ROI separately, we conducted a mixed-effect
linear regression model with maximal random-effect structure, predicting the level of response
with a fixed effect and random slopes for Condition, and random effects for Participant.
Condition was dummy-coded with Intact sentences as the reference level. Models were fit
separately for Experiment 1 and Experiment 2 using the brms package (布尔克纳, 2017) 在R中
(R Team, 2017) to interface with Stan (Stan Development Team, 2018).
Behavioral naturalness rating study
To ensure that our scrambling manipulation was successful (in that human comprehenders
would show sensitivity to it in some behavioral measure), 76 participants recruited through
Amazon.com’s Mechanical Turk rated the naturalness of the sentence stimuli used in
实验 1 on a 7-point scale (从 1 = unnatural to 7 = natural). On each trial, 参与者
were presented with a single stimulus on the screen along with the scale. The end points of the
Neurobiology of Language
114
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
n
哦
/
我
/
我
A
r
t
我
C
e
–
p
d
F
/
/
/
/
1
1
1
0
4
1
8
6
7
9
0
2
n
哦
_
A
_
0
0
0
0
5
p
d
/
.
我
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
Composition drives the language-selective network
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
n
哦
/
我
/
我
A
r
t
我
C
e
–
p
d
F
/
数字 3. Neural responses (在 % BOLD signal change relative to fixation) to the conditions of the language localizer and Experiments 1 (顶部
panel) 和 2 (bottom panel) in each of the six language functional regions of interest (fROIs). LAntTemp — left anterior temporal lobe, LIFG — left
inferior frontal gyrus, LIFGorb — left orbital inferior frontal gyrus, LPostTemp — left posterior temporal lobe, LMFG — left middle frontal gyrus,
LAngG — left angular gyrus, BOLD — blood oxygen-level dependent, Int — intact, Scr — scrambled.
/
/
/
1
1
1
0
4
1
8
6
7
9
0
2
n
哦
_
A
_
0
0
0
0
5
p
d
/
.
我
scale were labeled. Participants responded by selecting a discrete point on the scale and then
pressing the “Enter” key on their keyboard to move to the next trial. As in the fMRI study, 这
materials were distributed across five experimental lists (150 trials each) following a Latin
square design. Each list contained only one version of a sentence and 30 trials of each of
the five conditions (Int, Scr1, Scr3, Scr5, and Scr7). Any given participant saw the materials
from just one experimental list. Due to a computer error, one list was administered to 16 par-
ticipants; other lists were seen by 15 participants each.
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
The ratings were analyzed using a mixed-effect linear regression model with a fixed effect
and random slopes for Condition, and random effects for Participant and Item. To demonstrate
the effectiveness of the manipulation at every level, Condition was backwards difference
coded. As can be seen in Figure 4a and Table 2, every increase in degradation was associated
with a significant decrease in perceived naturalness, although with diminishing returns. 因此
participants were robustly sensitive to the scrambling manipulation. (The presentation code
and data are available at OSF [Mollica et al., 2019].)
Behavioral sentence reconstruction study
To assess the extent to which participants might be able to reconstruct the original sentence
from its scrambled version, 180 additional participants recruited through Mechanical Turk
were presented with the scrambled stimuli and asked to try to create a well-formed and
Neurobiology of Language
115
Composition drives the language-selective network
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
n
哦
/
我
/
我
A
r
t
我
C
e
–
p
d
F
/
/
/
/
1
1
1
0
4
1
8
6
7
9
0
2
n
哦
_
A
_
0
0
0
0
5
p
d
.
/
我
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
数字 4. Behavioral data from norming and experiments. A. The average naturalness rating (higher =
more natural). 乙. The average reconstruction accuracy. C, d. The average memory probe accuracy from
实验 1 和 2. All error bars reflect 95% bootstrapped confidence intervals. Int — intact, Scr —
scrambled, PMI — pointwise mutual information.
meaningful sentence out of the words. As part of the instructions, several simple examples
were provided. Participants were instructed that the actual stimuli would be more difficult
and that they should try their best before moving on. For control purposes, we included
one of the word-list conditions from Experiment 2, but we do not analyze those data here.
Similar to the rating study, the materials were distributed across experimental lists (six lists
in this study, 150 trials each) following a Latin square design. Each list contained only one
version of a sentence and 25 trials of each of the six conditions (Scr1, Scr3, Scr5, Scr7, 和
ScrLowPMI, word-list). Any given participant saw the materials from just one experimental list.
Each list was seen by 30 参与者. On each trial, participants were presented with a single
stimulus on the screen along with a text box. Participants’ responses were automatically con-
strained to include only words in the stimulus; 然而, due to a script error, participants were
allowed to use some words from the stimulus multiple times or to omit words. In the analyses,
we excluded all trials in which a response was not the same length as the stimuli, 导致
17% overall data loss (Scr1: 6%, Scr3: 11%, Scr5: 14%, Scr7: 19%; ScrLowPMI: 34%). 这
Neurobiology of Language
116
Composition drives the language-selective network
桌子 2. The results of a mixed-effect linear regression for the acceptability rating data
Grand mean
Scr1 – Int
Scr3 – Scr1
Scr5 – Scr3
Scr7 – Scr5
Estimate
3.54*
−2.04*
−1.42*
−0.56*
−0.23*
Est. Error
0.08
0.11
0.08
0.05
0.04
95% CI
3.38
−2.25
−1.57
−0.67
−0.30
3.69
−1.83
−1.26
−0.46
−0.15
*Denotes significant difference. Int — intact, Scr — scrambled.
distribution of data loss over conditions is itself a reflection of increasing reconstruction diffi-
culty as the number of swaps increases.
Reconstruction accuracy was analyzed using a logistic mixed-effect linear regression model
with a fixed effect and random slopes for Condition, and random effects for Participant and
Item. As in the naturalness rating study, Condition was backwards difference coded. As can be
seen in Figure 4b and Table 3, every increase in degradation was associated with a significant
decrease in the ability to reconstruct the sentence. This result suggests that it is unlikely that
participants were able to reconstruct a full-fledged sentence-level meaning, especially given
the word-by-word presentation and time demands of our task in the scanner compared to the
unlimited time participants were given in the web-based reconstruction task. We return to this
point in the Discussion section.
Discovering and characterizing brain regions sensitive to the sentence-scrambling manipulation. 给定
that in the behavioral naturalness rating study we found robust sensitivity to the scrambling
manipulation, we asked whether any parts of the brain work harder when we process
scrambled sentences. To search for brain regions that are sensitive to scrambling, we performed
a GSS whole-brain analysis (Fedorenko et al., 2010; Julian et al., 2012). This analysis searches
for spatially consistent (across individuals) patterns of activation while taking into account
interindividual variability in the precise loci of activations, which increases sensitivity relative to
traditional random-effects analyses that assume voxel-wise correspondence across people
(Nieto-Castañón & Fedorenko, 2012). We chose a contrast between the most scrambled
condition that was shared between the two experiments (IE。, Scr5) and the Intact condition.
Pooling data across experiments (n = 47; for the participant who took part in both Experiments 1
和 2, we used the data from Experiment 1; for the participant who took part in Experiment 2
twice, we used the data from the first session), we took individual whole-brain activation maps
桌子 3. The results of a mixed-effect logistic regression for the reconstruction accuracy data
Grand mean
Scr3 – Scr1
Scr5 – Scr3
Scr7 – Scr5
ScrLowPMI – Scr7
Estimate
0.41*
−0.71*
−0.59*
−0.32*
−1.46*
Est. Error
0.13
0.07
0.06
0.06
0.07
95% CI
0.15
−0.85
−0.70
−0.44
−1.61
*Denotes significant difference. Int — intact, Scr — scrambled, PMI — pointwise mutual information.
0.69
−0.57
−0.48
−0.21
−1.33
117
Neurobiology of Language
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
n
哦
/
我
/
我
A
r
t
我
C
e
–
p
d
F
/
/
/
/
1
1
1
0
4
1
8
6
7
9
0
2
n
哦
_
A
_
0
0
0
0
5
p
d
.
/
我
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
Composition drives the language-selective network
for the Scr5 > Int contrast and binarized them so that voxels that show a reliable effect
(significant at p < 0.05, uncorrected at the whole-brain level) were turned into 1’s and all
other voxels were turned into 0’s. ( We chose a liberal threshold for the individual activation
maps to maximize our chances of detecting regions of interest; as explained below, however,
the resulting regions were subsequently evaluated using statistically conservative criteria.) We
overlaid these maps to create a probabilistic activation overlap map, thresholded this map to
include only voxels where at least 4 of the 47 participants showed activation, and divided it into
“parcels” using a watershed image parcellation algorithm (see Fedorenko et al., 2010, for
details). Finally, we identified parcels that—when intersected with the individual activation
maps—contained suprathreshold (i.e., significant for our contrast of interest at p < 0.05,
uncorrected) voxels in at least half of the individual participants.
To characterize the functional profiles of scrambling-responsive regions in greater detail, in
each of the regions, we estimated the BOLD response magnitude to the conditions of the two
experiments. To estimate the responses to the Scr5 and Int conditions, which were used in the
localizer contrast, we used an across-runs cross-validation procedure (e.g., Nieto-Castañón &
Fedorenko, 2012), to ensure independence between the data used to define the fROIs and to
estimate the responses (Kriegeskorte, Simmons, Bellgowan, & Baker, 2009). In particular, each
parcel was intersected with each participant’s activation map for the Scr5 > Int contrast for all
but the first run of the data. The voxels within the parcel were sorted—for each participant—
based on their t-values, and the top 10% of voxels were selected as that participant’s fROI. 这
responses were then estimated using the left-out run’s data. The procedure was repeated iteratively
leaving out each of the runs. 最后, the responses were averaged across the left-out runs to derive
a single-response magnitude per subject per region per condition. To estimate the responses to the
other critical conditions, we used data from the Scr5 and Int conditions across all runs. Statistical
tests were performed on these extracted percent BOLD signal change values.
此外, we estimated the BOLD responses of scrambling-responsive regions to two other ex-
实验: (A) the language localizer and (乙) the spatial working memory experiment. Responses to
the language localizer conditions can tell us whether the scrambling-responsive regions show a sig-
nature of the language network: 那是, stronger responses to sentences than nonword sequences. 我们
have constrained our definition of the language-responsive regions in the critical analyses by a set of
parcels derived according to activations for the language localizer contrast in a large number of in-
个人 (as described earlier). Thus regions outside of this network of language-responsive regions
should not show language-responsive properties. So this analysis provides a reality check of sorts.
Responses to the conditions of the spatial working memory task tell us whether the scrambling-
responsive regions may belong to the domain-general multiple demand (医学博士) 网络, 哪个重新-
sponds robustly to this task (例如, Fedorenko et al., 2013) and which has been generally implicated
in executive functions like working memory and cognitive control (Duncan, 2010, 2013).
结果
Behavioral (Memory Probe Task) Data in Experiments 1 和 2
Response accuracy for each experiment was analyzed with a logistic mixed-effect linear re-
gression model with a fixed effect and random slopes for Condition, and random intercepts for
Participant and Item. Condition was dummy-coded with Intact Sentences as the reference
等级. For both experiments, accuracy was above chance for all conditions. In Experiment 1,
accuracies in the scrambled sentence conditions did not differ significantly from accuracy in
the intact sentence condition; 然而, accuracy was significantly lower in the word-list and
nonword-list conditions compared to the intact sentence condition (Figure 4c and Table 4), 在
Neurobiology of Language
118
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
n
哦
/
我
/
我
A
r
t
我
C
e
–
p
d
F
/
/
/
/
1
1
1
0
4
1
8
6
7
9
0
2
n
哦
_
A
_
0
0
0
0
5
p
d
.
/
我
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
Composition drives the language-selective network
桌子 4. The results of logistic mixed-effect models for Experiments 1 和 2 for the memory probe data
实验 1
实验 2
Int
Scr1 versus Int
Scr3 versus Int
Scr5 versus Int
Scr7 versus Int
ScrLowPMI versus Int
Words versus Int
Nonwords versus Int
Estimate
3.63*
Est. 错误
0.33
0.37
−0.02
−0.39
−0.45
–
−1.74*
−2.35*
0.86
0.74
0.59
0.61
–
0.38
0.33
95% CI
3.03
−0.94
−1.19
−1.39
−1.50
–
−2.51
−3.04
4.32
2.44
1.76
0.94
0.87
–
−1.01
−1.75
Estimate
1.98*
Est. 错误
0.24
−0.19
0.11
−0.27
–
−0.64*
−0.30*
–
0.22
0.27
0.24
–
0.23
0.17
–
95% CI
1.52
−0.62
−0.39
−0.71
–
−1.07
−0.64
–
2.46
0.27
0.70
0.22
–
−0.16
0.04
–
Stimulus type was dummy-coded with Intact sentences as the reference level. Int — intact, Scr — scrambled, PMI — pointwise mutual information.
*Denotes significant difference.
line with prior work (例如, Fedorenko et al., 2010). 相似地, in Experiment 2, accuracies in the
scrambled sentence conditions did not differ significantly from accuracy in the intact sentence
状况; 然而, accuracy was lower in the ScrLowPMI and the word-list conditions com-
pared to the intact sentence condition (Figure 4d and Table 4). (Data and analysis code are avail-
able at OSF [Mollica et al., 2019].)
fMRI Data in Experiments 1 和 2
In Experiment 1, replicating much prior work (Fedorenko et al., 2010; Pallier et al., 2011), 出色地-
formed sentences elicited significantly stronger BOLD responses than the word-list and nonword-
list conditions (Figure 2c, 桌子 5). 然而, degrading the sentences by introducing local word
swaps did not decrease the magnitude of the language network’s response: Even stimuli with seven
word swaps (例如, their last on they overwhelmed were day farewell by messages and gifts; 数字 1)
elicited as strong a response as fully grammatical sentences (例如, on their last day they were over-
whelmed by farewell messages and gifts; Figure 2c, 桌子 5). The results also held—both qualita-
tively and statistically—for each language ROI separately (数字 3 和表 6). This pattern of
similarly strong responses for the well-formed and degraded sentences suggests that interword de-
pendencies are being formed even when the word order violates the rules of the language, 和
supports the idea that composition is the core computation implemented in the language network.
In Experiment 2, we replicated the pattern observed in Experiment 1 for the intact sentences
and sentences with 1, 3, 或者 5 local word swaps, all of which elicited similarly strong BOLD
responses, all reliably higher than the control, word-list, 状况 (see Figure 2d, 桌子 5).
然而, the ScrLowPMI condition elicited a response that was as low as that elicited by lists
of unconnected words (see Figure 2d, 桌子 5), demonstrating that combinable words have to
occur in proximity to one another for the composition mechanisms to get triggered. 再次, 这
results held for each language ROI separately (见图 3 和表 6).
Brain Regions Sensitive to the Sentence-Scrambling Manipulation
Despite eliciting as strong a BOLD response as well-formed and meaningful sentences, 这
scrambled sentences were rated as less acceptable behaviorally (Figure 4a and Table 2),
Neurobiology of Language
119
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
n
哦
/
我
/
我
A
r
t
我
C
e
–
p
d
F
/
/
/
/
1
1
1
0
4
1
8
6
7
9
0
2
n
哦
_
A
_
0
0
0
0
5
p
d
.
/
我
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
Composition drives the language-selective network
桌子 5. The results of mixed effect linear regressions for Experiments 1 和 2
实验 1
实验 2
Estimate
1.06*
Est. 错误
0.20
0.07
0.11
0.03
0.06
–
−0.73*
−1.00*
0.09
0.10
0.11
0.11
–
0.12
0.13
95% CI
0.64
−0.11
−0.08
−0.19
−0.15
–
−0.97
−1.26
1.46
0.24
0.30
0.26
0.27
–
−0.50
−0.73
Estimate
0.90*
Est. 错误
0.18
0.03
−0.10
−0.02
–
−0.35*
−0.46*
–
0.07
0.08
0.07
–
0.08
0.08
–
95% CI
0.55
−0.11
−0.25
−0.16
–
−0.52
−0.62
–
1.27
0.17
0.04
0.12
–
−0.19
−0.30
–
Int
Scr1 versus Int
Scr3 versus Int
Scr5 versus Int
Scr7 versus Int
ScrLowPMI versus Int
Words versus Int
Nonwords versus Int
Condition was dummy-coded with Intact sentences as the reference level. Int — intact, Scr — scrambled, PMI — pointwise mutual information.
*Denotes significant difference.
suggesting that there has to be a cost to the processing of this kind of degraded linguistic input.
The whole-brain search for scrambling-sensitive areas discovered four regions, located in the
middle frontal gyrus bilaterally and in the Supplementary Motor Area (SMA) (Figure 5a).
The patterns of responses observed—averaging across the fROIs—are shown in Figure 5b.
Qualitatively, with respect to the conditions of the critical experiments, we found that the re-
sponse increased parametrically from the Int to the Scr5 condition in both experiments.
此外, in Experiment 1, the response remained high for the Scr7 condition, but in
实验 2, the response fell off for the ScrLowPMI condition. To quantify this non-monotonic
pattern, we collapsed across experiments and conducted a mixed-effect linear regression with
第一的- and second-order terms for Edit Distance (IE。, the number of swaps required to reconstruct
the original intact sentence) as a fixed effect and random slopes, and random effects for
Participant and ROI. We found a small but significant increase in the BOLD response as stimuli
become more scrambled, with a decrease in the ScrLowPMI condition (桌子 7).
With respect to the conditions of the language localizer and the spatial working memory
实验, none of the four fROIs showed a stronger response to sentences than nonword
序列 (实际上, three of the four regions showed a reliably stronger response to nonword
sequences than sentences, in line with Fedorenko et al., 2013); and all four fROIs showed a
stronger response to the Hard than Easy condition in the spatial working memory experiment.
These results suggest that the scrambling-responsive fROIs fall within the domain-general MD
cortex (Duncan, 2010, 2013). The parametric increase as a function of the degree of scram-
bling in the critical experiments is in line with the robust sensitivity of the MD cortex to effort
across domains (例如, Duncan & 欧文, 2000; Hugdahl et al., 2015). 尤其, 参与者
have to exert greater cognitive effort to extract meaning from the more scrambled sentences
(perhaps due to greater uncertainty about how words go together, as suggested by the results of
the behavioral sentence reconstruction experiment; 图4b). The fall-off in these fROIs for
the ScrLowPMI condition—which elicited a low response in the language regions as shown in
our critical analysis—is consistent with the idea that participants “give up” their attempts to
derive a meaningful representation in this condition (例如, Callicott et al., 1999; Linden et al.,
Neurobiology of Language
120
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
n
哦
/
我
/
我
A
r
t
我
C
e
–
p
d
F
/
/
/
/
1
1
1
0
4
1
8
6
7
9
0
2
n
哦
_
A
_
0
0
0
0
5
p
d
.
/
我
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
Composition drives the language-selective network
桌子 6. The results of mixed effect linear regressions for Experiments 1 和 2 for the six language functional regions of interest
实验 1
实验 2
Estimate
Est. 错误
95% CI
Estimate
Est. 错误
95% CI
Int
LIFGorb
LIFG
LMFG
LAntTemp
LPostTemp
LAngG
Scr1 versus Int
LIFGorb
LIFG
LMFG
LAntTemp
LPostTemp
LAngG
Scr3 versus Int
LIFGorb
LIFG
LMFG
LAntTemp
LPostTemp
LAngG
Scr5 versus Int
LIFGorb
LIFG
LMFG
LAntTemp
LPostTemp
LAngG
1.21*
1.43*
1.04*
0.89*
1.24*
0.61*
0.09
0.16
0.09
−0.01
0.05
0.05
0.16
0.26*
0.18*
−0.03
0.07
0.02
0.08
0.22
0.17*
−0.13*
−0.02
−0.14*
Neurobiology of Language
0.12
0.18
0.15
0.09
0.08
0.13
0.10
0.10
0.07
0.05
0.07
0.05
0.12
0.11
0.08
0.07
0.07
0.05
0.12
0.12
0.08
0.05
0.08
0.07
0.97
1.07
0.75
0.72
1.07
0.37
−0.10
−0.03
−0.05
−0.11
−0.08
−0.04
−0.08
0.05
0.02
−0.17
−0.07
−0.09
−0.16
−0.01
0.02
−0.23
−0.17
−0.28
1.44
1.80
1.32
1.07
1.41
0.87
0.27
0.35
0.23
0.10
0.18
0.16
0.39
0.48
0.34
0.11
0.20
0.12
0.31
0.45
0.32
−0.3
0.14
−0.01
0.75*
0.98*
1.11*
0.89*
1.04*
0.60*
0.01
0.04
0.11*
0.01
0.02
−0.05
−0.17
−0.12
−0.03
−0.09
−0.09
−0.13
−0.07
0.00
0.08
−0.07
−0.02
−0.08
0.18
0.17
0.16
0.10
0.10
0.14
0.09
0.07
0.06
0.05
0.06
0.06
0.09
0.08
0.07
0.06
0.06
0.08
0.10
0.07
0.07
0.06
0.06
0.06
0.40
0.64
0.81
0.70
0.85
0.33
−0.17
−0.09
0.00
−0.10
−0.09
−0.17
−0.36
−0.28
−0.17
−0.20
−0.21
−0.28
−0.26
−0.14
−0.06
−0.18
−0.14
−0.21
1.10
1.31
1.42
1.09
1.24
0.86
0.19
0.17
0.23
0.12
0.13
0.07
0.01
0.03
0.12
0.02
0.03
0.03
0.12
0.14
0.21
0.04
0.10
0.04
121
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
n
哦
/
我
/
我
A
r
t
我
C
e
–
p
d
F
/
/
/
/
1
1
1
0
4
1
8
6
7
9
0
2
n
哦
_
A
_
0
0
0
0
5
p
d
/
.
我
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
Composition drives the language-selective network
桌子 6. (continued )
实验 1
实验 2
Estimate
Est. 错误
95% CI
Estimate
Est. 错误
95% CI
Scr7 versus Int
LIFGorb
LIFG
LMFG
LAntTemp
LPostTemp
LAngG
0.13
0.20
0.16
−0.08
0.03
−0.09
ScrLowPMI versus Int
LIFGorb
LIFG
LMFG
LAntTemp
LPostTemp
LAngG
Words versus Int
LIFGorb
LIFG
LMFG
LAntTemp
LPostTemp
LAngG
–
–
–
–
–
–
−0.75*
−0.89*
−0.75*
−0.62*
−0.83*
−0.55*
Nonwords versus Int
LIFGorb
LIFG
LMFG
LAntTemp
LPostTemp
LAngG
−1.06*
−1.26*
−0.87*
−0.86*
−1.11*
−0.85*
0.13
0.15
0.11
0.07
0.08
0.06
–
–
–
–
–
–
0.10
0.14
0.14
0.08
0.07
0.09
0.11
0.16
0.11
0.07
0.08
0.13
−0.14
−0.08
−0.05
−0.23
−0.13
−0.22
–
–
–
–
–
–
−0.95
−1.17
−1.04
−0.77
−0.98
−0.73
−1.29
−1.57
−1.08
−1.00
−1.26
−1.10
0.39
0.49
0.38
0.05
0.20
0.04
–
–
–
–
–
–
−0.54
−0.62
−0.48
−0.46
−0.68
−0.38
−0.84
−0.95
−0.65
−0.72
−0.94
−0.58
–
–
–
–
–
–
−0.44*
−0.39*
−0.29*
−0.34*
−0.37*
−0.28*
−0.41*
−0.43*
−0.55*
−0.44*
−0.45*
−0.49*
–
–
–
–
–
–
–
–
–
–
–
–
0.11
0.09
0.08
0.06
0.06
0.08
0.13
0.10
0.07
0.07
0.07
0.07
–
–
–
–
–
–
–
–
–
–
–
–
−0.66
−0.57
−0.46
−0.46
−0.50
−0.43
−0.67
−0.62
−0.68
−0.58
−0.58
−0.63
–
–
–
–
–
–
–
–
–
–
–
–
−0.22
−0.21
−0.13
−0.22
−0.25
−0.13
−0.16
−0.23
−0.40
−0.31
−0.31
−0.36
–
–
–
–
–
–
Condition was dummy-coded with Intact sentences as the reference level. LAntTemp—left anterior temporal lobe, LIFG—left inferior frontal gyrus, LIFGorb—
left orbital inferior frontal gyrus, LPostTemp—left posterior temporal lobe, LMFG—left middle frontal gyrus, LAngG—left angular gyrus, Int — intact,
Scr — scrambled.
*Denotes significant difference.
Neurobiology of Language
122
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
n
哦
/
我
/
我
A
r
t
我
C
e
–
p
d
F
/
/
/
/
1
1
1
0
4
1
8
6
7
9
0
2
n
哦
_
A
_
0
0
0
0
5
p
d
.
/
我
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
Composition drives the language-selective network
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
n
哦
/
我
/
我
A
r
t
我
C
e
–
p
d
F
/
/
/
/
1
1
1
0
4
1
8
6
7
9
0
2
n
哦
_
A
_
0
0
0
0
5
p
d
/
.
我
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
数字 5.
Scrambling sensitive brain regions and their BOLD response profile. A. The parcels used to define the scrambling-responsive areas. 在
each participant, the top 10% of most localizer-responsive voxels within each parcel were taken as that participant’s region of interest. 乙. Neural
responses (在 % BOLD signal change relative to fixation) to the conditions of Experiments 1 (顶部) 和 2 (底部), as well as the language localizer
and spatial working memory task. BOLD — blood oxygen-level dependent, MD — multiple demand, Int — intact, Scr — scrambled, PMI —
pointwise mutual information.
2003; 比照. Wen, 米切尔, & Duncan, 2018). 尤其, because participants no longer have
the evidence in the input that nearby words are combinable, they stop engaging their compo-
sition mechanisms.
讨论
In this study, we evaluated a hypothesis that that the core linguistic computation implemented
in the language-selective cortex has to do with combining words into phrases and clauses, 和
that this computation can take place even when the word order is not licensed by the lan-
guage’s grammar. Across two fMRI experiments, we examined the processing of stimuli where
桌子 7. The results of mixed effect linear regression models for the scrambling-responsive regions
Intercept
Edit distance
Edit distance2
Estimate
0.844
0.046*
−0.002*
Est. 错误
0.491
0.024
0.001
95% CI
−0.125
0.012
−0.003
*Denotes significant difference.
1.835
0.086
−0.0005
123
Neurobiology of Language
Composition drives the language-selective network
the word order was degraded, via a novel parametric manipulation (varying numbers of local
word swaps), making word-order–based parsing difficult or impossible, but semantic and syn-
tactic dependencies could still be formed among nearby words. Using behavioral measures in
independent groups of participants, we established robust sensitivity to the scrambling
manipulation: Sentences with more word swaps, and correspondingly more syntactic depen-
dencies disrupted (见表 1), were rated as less natural (见表 2, 图4a), and it was
more difficult to reconstruct the original sentence from the scrambled versions (见表 3,
图4b). 然而, scrambled sentences, even the conditions with a large number (5 和 7)
word swaps, elicited BOLD responses in the language areas that were as strong as the response
elicited by naturalistic sentences. Only when interword dependencies could not be formed
among nearby words did the BOLD response in the language areas drop to the level of that
elicited by lists of unconnected words. These results suggest that the ability to form local
dependencies is necessary and sufficient for eliciting the maximal BOLD response in the
language-selective brain network, where maximal is defined as the BOLD response to the pre-
ferred stimulus—well-formed and meaningful sentences. We interpret these findings as suggest-
ing that composition is the core linguistic computation driving the neural responses in the
language-selective cortex, and that this computation does not depend on word order (看
Bornkessel-Schlesewsky, Schlesewsky, 小的, & Rauschecker, 2015, for a related proposal).
Our analyses of the experimental materials and the behavioral sentence reconstruction
study help rule out two alternative explanations of these findings. One possibility is that
conditions with 1-, 3-, 5-, and even 7-word swaps (Scr1, Scr3, Scr5, and Scr7) but not the
ScrLowPMI condition contained a sufficiently long well-formed and meaningful substring,
and that such substrings are sufficient to elicit a BOLD response similar to that elicited by a
fully well-formed sentence. To rule out this possibility, we turn to an earlier fMRI study by
Pallier et al. (2011). They examined responses to 12-word–long sequences that varied in their
composition between a sentence, two 6-word–long substrings, three 4-word–long substrings,
four 3-word–long substrings, six 2-word–long substrings, and a list of 12 unconnected words.
The BOLD response was shown to fall off as a function of the length of the substrings: A 12-
word–long sentence elicited a stronger response than a sequence composed of two 6-word–
long substrings, 哪个, 反过来, elicited a stronger response than a sequence composed of three
4-word–long substrings, 等等. We replicate this finding in our work (Mollica et al., 和-
published data). The analysis of our experimental materials (见表 1) revealed that the
length of the longest well-formed and meaningful substring decreases with each scrambling
level and drops to 4.46 words on average for the condition with 7-word swaps. 因此, 这
alternative hypothesis considered here predicts a gradual fall-off in the BOLD response from
the Int condition to the Scr7 condition, which is not the pattern we observe.
Another possibility is that participants were able to reconstruct the original sentence in all
the scrambled conditions except for the ScrLowPMI condition. This possibility is unlikely
given that in the behavioral sentence reconstruction study the ability to reconstruct the orig-
inal sentence dropped off with each additional scrambling level (see Figure 4b and Table 3).
And this pattern was observed despite access by participants to the entire stimulus string and
were not limited time-wise (比照. the word-by-word relatively fast presentation in the scanner).
此外, we model the BOLD response during the entire trial, 和, by design, 参与者
do not have access to all the words until after the last word has been presented. 因此, 这
similarly strong BOLD response across the Int through Scr7 conditions is unlikely due to par-
ticipants successfully “unscrambling” the stimuli and processing them as such. (当然,
some local unscrambling could still take place. 然而, it is important to note that this un-
scrambling was apparently not carried out by the language areas, given that there was no
Neurobiology of Language
124
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
n
哦
/
我
/
我
A
r
t
我
C
e
–
p
d
F
/
/
/
/
1
1
1
0
4
1
8
6
7
9
0
2
n
哦
_
A
_
0
0
0
0
5
p
d
.
/
我
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
Composition drives the language-selective network
increase in neural response to the scrambled compared to intact stimuli in these areas. 因此,
whatever computation is performed by the language areas proceeds in the same way in the
intact and the scrambled conditions.)
Having ruled out these two alternatives, we argue that during the incremental processing of
linguistic strings, participants form dependency relationships among words within a moving
local context of a few words. This process results in the construction of phrase- and clause-
level meanings. Composition is driven by the lexicosemantic and syntactic (part of speech and
morphological endings) properties of the input words combined with a plausibly Bayesian
inductive inference process (例如, Steyvers, Griffiths, & 丹尼斯, 2006). 尤其, when lin-
guistic data underconstrain interpretation, participants likely make their best guesses about the
intended meaning by combining the information in the input with their prior semantic and
linguistic knowledge (看, 例如, Chater & 曼宁, 2006; 吉布森等人。, 2013, for applications
of the general Bayesian framework to linguistic interpretation).
In addition to the consistently high BOLD response across the scrambled conditions, 这
behavioral data from the memory probe task performed in the scanner (Figure 4c, d and
桌子 4) provide indirect evidence that complex meanings were formed during the processing
of all the sentence conditions except for the ScrLowPMI condition. 尤其, a classic find-
ing in the memory literature is that people’s memory for phrases and sentences is superior to
their memory for lists of unconnected words (例如, Baddeley, Hitch, & 艾伦, 2009; Brener,
1940), which has been attributed either to the fact that people represent sentences in terms
of their meaning/gist extracted during comprehension, and that gist can later be used to regen-
erate the specific word-forms (例如, Potter & Lombardi, 1990), or to the automatic engagement
of long-term memory mechanisms during sentence-level comprehension, which leads to
more-effective binding of information within the episodic buffer (Baddeley, 2000; Baddeley,
艾伦, & Hitch, 2011). We found that participants’ performance on the memory probe task did
not decline as a function of the scrambling manipulation. It only dropped in the ScrLowPMI
状况. This consistently high memory probe performance can be used to indirectly infer
that participants successfully formed complex meaning representations in the scrambled con-
版本, as they did when processing well-formed and meaningful sentences.
In the remainder of the article, we discuss three issues that our results speak to.
The Relationship Between Lexicosemantic and Syntactic Processes
In this study, we showed that word order—one component of syntax—does not appear
to affect the basic composition process carried out by the core frontotemporal language-
selective network: Provided that dependencies can be formed between nearby words in
linguistic strings, the composition mechanisms get engaged as they do when we process
naturalistic linguistic input. Throughout the article, we have described the composition
process as encompassing both semantic composition and syntactic structure building. 这
relationship between the two has been treated differently across proposals in the theoretical
linguistic literature. In mainstream generative grammar and formal semantics (例如, Chomsky,
1965, 1981; Montague, 1974; Partee, 1975, 1995; Partee, ter Meulen, & Wall, 1990),
semantic composition is considered to be a special case of syntactic composition, and syntax
determines the meaning of a phrase or a clause. 然而, according to an alternative
看法, semantic composition can proceed (partially or fully) independently from syn-
tactic structure building (例如, Culicover & 杰肯道夫, 2006, Culicover & 杰肯道夫, 2005;
杰肯道夫, 2007, 2010; 杰肯道夫 & 杰肯道夫, 2002; 杰肯道夫 & Wittenberg, 2017;
Kuperberg, 2007). Baggio (2018) refers to this idea as “autonomous semantics.”
Neurobiology of Language
125
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
n
哦
/
我
/
我
A
r
t
我
C
e
–
p
d
F
/
/
/
/
1
1
1
0
4
1
8
6
7
9
0
2
n
哦
_
A
_
0
0
0
0
5
p
d
.
/
我
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
Composition drives the language-selective network
According to his proposal, words are bound into “relational structures” based on associa-
主动的, categorical, and logical relationships (see Michalon & Baggio, 2019, for evidence of
computational feasibility). We are sympathetic to the latter view. We think of semantics as
an independent computational system that obeys its own rules for how words are bound
together during language comprehension. 当然, many of these rules have correlates
in syntax, but nevertheless we conceive of semantic composition as a process that can
take place independently from syntactic structure building.
然而, at the implementation level, it does not appear to be the case that semantic com-
position and syntactic structure building are spatially separable in the brain, at least at the
resolution accessible to current imaging techniques. Many have searched for and claimed
to have observed a dissociation between brain regions that support (lexico-)semantic process-
ing and those that support syntactic processing (例如, Cooke et al., 2006; Dapretto &
Bookheimer, 1999; Embick, Marantz, Miyashita, O’Neil, & Sakai, 2000; Friederici, Opitz, &
von Cramon, 2000; Noppeney & Price, 2004, inter alia). 然而, some of these classic find-
ings do not appear robust to replication (Siegelman et al., 2019). And in general, taking the
available evidence from cognitive neuroscience en masse, the picture that has emerged does
not support a double dissociation between lexicosemantic and syntactic processes.
第一的, the specific regions that have been argued to support (lexico-)semantic versus syntac-
tic processing, and the precise construal of these regions’ contributions, differ widely across
studies and proposals (例如, Baggio & Hagoort, 2011; Bemis & Pylkkänen, 2011; Duffau,
Moritz-Gasser, & Mandonnet, 2014; Friederici, 2011, 2012; Matchin & Hickok, 2019; Tyler
等人。, 2011; Ullman, 2004, 2016). 第二, although diverse paradigms have been used across
studies to probe semantic versus syntactic processing, any given study (比照. Fedorenko, 空白的,
Siegelman, & Mineroff, 2020) has typically used a single paradigm, raising the possibility that
the results reflect paradigm-specific differences between conditions rather than a general
difference between semantic and syntactic computations. 此外, given the tight link
between meaning and structure, results from some syntactic manipulations may, 实际上, 是
due to parallel semantic composition processes. 最后, a number of neuroimaging studies
have failed to observe a double dissociation between semantic and syntactic processing,
reporting instead overlapping areas of activation (例如, Bautista & Wilson, 2016; Fedorenko
等人。, 2010; Keller et al., 2001; Röder, Stock, Neville, Bien, & Rösler, 2002). 尤其,
any brain region that shows sensitivity to syntactic processing appears to be at least as sensitive
to individual word meanings and semantic composition. It is notable that there do exist brain
areas—in the left anterior temporal lobe/temporal pole—that respond to word meanings, or abstract
conceptual representations, according to some accounts, but not syntactic/combinatorial process-
英 (例如, Lambon Ralph, Jefferies, 帕特森, & 罗杰斯, 2017; Mesulam et al., 2013; 帕特森,
Nestor, & 罗杰斯, 2007; Schwartz et al., 2009; 施瓦茨, Marin, & 藏红花, 1979; Visser,
Jefferies, & Lambon Ralph, 2010; 比照. Westerlund & Pylkkänen, 2017). 总之, 看起来
syntactic processing (A) is not focally carried out in a particular brain region within the language
network contra some proposals (例如, Berwick, Friederici, Chomsky, & Bolhuis, 2013; Brennan
等人。, 2012; Friederici, Bahlmann, Heim, Schubotz, & 安旺德, 2006; Matchin & Hickok,
2019; Tyler et al., 2011), but is distributed across the left lateral frontal and temporal areas (例如,
空白的, Balewski, Mahowald, & Fedorenko, 2016); 和 (乙) is supported by the very same brain
regions that support the processing of word meanings and semantic composition.
We would further argue that semantic composition, not syntactic structure building—to the
extent that the two are separable—is primary in language comprehension and is the core op-
eration driving the language-selective areas (see also Fedorenko et al., 2016; Pylkkänen &
Brennan, in press). On the theoretical side, this argument is motivated by a key function of
Neurobiology of Language
126
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
n
哦
/
我
/
我
A
r
t
我
C
e
–
p
d
F
/
/
/
/
1
1
1
0
4
1
8
6
7
9
0
2
n
哦
_
A
_
0
0
0
0
5
p
d
/
.
我
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
Composition drives the language-selective network
language—to communicate meanings (例如, 比照. Chomsky, 2002; Goldberg, 2006; 杰肯道夫,
2011). Abundant evidence now suggests that many properties of human languages—from the
sound systems, to lexicons, to grammars—have been shaped by communicative pressures, 到
optimize information transfer (see Gibson et al., 2019, for review). 因此, it seems likely
that our language-processing mechanisms would be optimized for extracting meaning from
the signal. On the empirical side, we know that meaningful sentences elicit stronger responses
in the language areas than structured but meaningless stimuli, like Jabberwocky sentences or
nonsensical sentences (例如, Fedorenko et al., 2010; 汉弗莱斯, Binder, Medler, & Liebenthal,
2007; 比照. Pallier et al., 2011; Scott et al., 2017) although the lack of a difference in the mean
response to real versus Jabberwocky sentences in some language areas does not appear to be
replicable, and is likely driven by a between-subject comparison in the original study
(Dehaene and Pallier, personal communication), suggesting that syntactic structure building
alone cannot explain the response properties of the language areas. 然而, future studies
should aim to further evaluate the relative importance of semantic versus syntactic composi-
tion in language comprehension.
Our results also speak to a differential role of language statistics in syntax versus semantics.
一方面, language statistics are relevant because humans plausibly store and contin-
ually update an implicit predictive model of linguistic forms that they use to anticipate upcom-
ing linguistic elements during comprehension (Christiansen & Chater, 2016; 黑尔, 2001; 征收,
2008A). 的确, a wealth of evidence demonstrates that expectations over linguistic forms af-
fect language processing (例如, 戴尔 & 张, 2014; Federmeier, 2007; Kuperberg & Jaeger,
2016; 皮克林 & Garrod, 2013). 另一方面, as discussed in the Introduction section,
language statistics are relevant because they reflect the distributional properties of objects and
events in the world, albeit with a bias toward objects and events that are worth encoding and
communicating through language (Andrews, Vigliocco, & Vinson, 2009; Griffiths, Steyvers, &
Tenenbaum, 2007). Our work, along with a recent computational model of the N400
(Rabovsky, 汉森, & 麦克莱兰, 2018), demonstrates that the brain is sensitive to language
statistics as a proxy for both world states and, perhaps more clearly, the implicit semantic de-
pendencies in world states (例如, which properties are likely to apply to which objects, 哪个
entities are likely engage in which actions, 等等). Keeping track of these kinds of depen-
dencies may subsume at least some of the syntactic information. 例如, Rabovsky et al.
(2018) show that a model trained on semantic dependencies alone captures word-order effects
observed in the N400 component.
The Temporal Receptive Window of the Language Areas
An important notion has been gaining ground in the recent literature: the idea of a temporal
receptive window (TRW) of a brain unit (cell, voxel, brain area) (例如, Hasson, 哪个, Vallines,
Heeger, & 鲁宾, 2008; Lerner, 蜂蜜, Silbert, & Hasson, 2011; Overath, 麦克德莫特, Zarate,
& Poeppel, 2015). A TRW is defined by Hasson and colleagues as “the length of time before a
response during which sensory information may affect that response,” although the amount of
information rather than time may be more relevant, especially for higher-level areas (例如,
Vagharchakian et al., 2012). What is the size of the TRW of the core language areas?
We have known for some time that discourse-level processing—connecting sentences into
coherent texts—is carried out by regions outside of the frontotemporal language network (例如,
Ferstl, 诺伊曼, Bogler, & von Cramon, 2008; Ferstl & von Cramon, 2001; Kuperberg et al., 2006;
Lerner et al., 2011; see Jacoby & Fedorenko, 2018 for evidence of insensitivity to discourse-
level processing in the functionally defined language areas of the core frontotemporal
Neurobiology of Language
127
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
n
哦
/
我
/
我
A
r
t
我
C
e
–
p
d
F
/
/
/
/
1
1
1
0
4
1
8
6
7
9
0
2
n
哦
_
A
_
0
0
0
0
5
p
d
.
/
我
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
Composition drives the language-selective network
网络). 例如, Lerner et al. (2011) presented participants with an auditory story as well
as the same story scrambled at different grains of information (at the paragraph level, 在
sentence level, and at the word level). In a whole-brain voxel-wise analysis of intersubject cor-
关系 (Hasson et al., 2008), which can be used to draw inferences about the size of the TRW
of a voxel, 他们发现 (A) brain areas sensitive to paragraph-level structure and above
resemble the default mode network (例如, 巴克纳, Andrews-Hanna, & Schacter, 2008) 或者
the network that supports social cognition (例如, Saxe & Kanwisher, 2003), 和 (乙) brain areas
sensitive to word- and sentence-level processing (but not to structure above the sentence level)
resemble the core language network. The intersubject correlations were higher for the
sentence-scrambled condition than the word-scrambled condition (see also Blank & Fedorenko,
2019), but where exactly between a single word and a sentence does the TRW of the language
areas fall?
The study of Pallier et al. (2011) discussed earlier showed that the response in the language
network appears to increase gradually from same-length sequences composed of single words
to 2-word phrases, to 3-word phrases, to 4-word phrases, to 6-word phrases, with an addi-
的, albeit smaller, increase for full sentences. Our results suggest that when combinable words
are separated by ∼8 words (as previously adjacent content words are in the ScrLowPMI condition;
average separation is 8.33 字), resulting in low average local PMI, composition does not take
地方, as evidenced by a low response in the language areas. The TRW of the language areas
therefore appears to be in the 5- to 7-word range. As alluded to in the Introduction section, 这
relatively local linguistic processing is likely driven by the statistical properties of natural lan-
规格, where most semantic/syntactic dependencies are local (例如, 富特雷尔等人。, 2015), 和
PMI falls off quite sharply as a function of interword distance (例如, 林 & Tegmark, 2017). 我们
can further speculate that linguistic chunks of this size are sufficient to express clause-level
含义, where clauses describe events—salient and meaningful semantic units in our expe-
rience with the world (例如, Zacks & Tversky, 2001). 当然, we can detect and process
syntactic and anaphoric dependencies that span much longer windows than ∼6 words, 和
these types of nonlocal dependencies have been investigated extensively in the psycholin-
guistic literature (例如, 吉布森, 1998, 2000; Lewis & 琼斯, 1996; 磨坊主 & Chomsky, 1963;
Yngve, 1960, inter alia). How exactly the processing of such dependencies is carried out in
the brain remains debated, in part because the most commonly used method in cognitive
神经科学 (功能磁共振成像) lacks the temporal resolution needed to track the dynamics of dependency
形成. We do not take our results as inconsistent with the human ability to process nonlocal
dependencies; 反而, we take them to suggest that our language-processing mechanisms
may be optimized for dealing with particular-size packages of linguistic information.
Sensitivity of Domain-General Executive Mechanisms to the Scrambling Manipulation
Although the BOLD responses of the language-selective regions were robust to the scrambling
manipulation, in a behavioral rating study, more scrambled sentences elicited lower natural-
ness ratings (see Figure 4a and Table 2), suggesting that such sentences should incur a greater
processing cost. What cognitive and neural mechanisms handle this extra cost? We found a
number of brain regions that appear to fall within the domain-general MD network (Duncan,
2010, 2013), which has been implicated broadly in goal-directed behavior and linked to ex-
ecutive functions, like working memory and cognitive control. These regions expended more
energy when participants processed sentences with scrambled word orders compared to intact
句子. The level of BOLD response increased as the degree of scrambling increased, 直到
participants were no longer able to form local semantic dependencies (as evidenced by a drop in
the BOLD response in the language network), which occurred in the ScrLowPMI condition. 这些
Neurobiology of Language
128
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
n
哦
/
我
/
我
A
r
t
我
C
e
–
p
d
F
/
/
/
/
1
1
1
0
4
1
8
6
7
9
0
2
n
哦
_
A
_
0
0
0
0
5
p
d
.
/
我
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
Composition drives the language-selective network
results suggest that the cost associated with the processing of scrambled sentences is carried by
domain-general executive regions that support diverse demanding tasks across domains (例如,
Duncan & 欧文, 2000; Hugdahl et al., 2015).
The importance and the precise role of the MD network in language comprehension remains
debated (例如, Blank et al., 2014; 坎贝尔 & Tyler, 2018; Diachek, 空白的, Siegelman, &
Fedorenko, 2019; 赖特, Randall, Marslen-Wilson, & Tyler, 2011). A number of prior studies
have reported activation in the MD areas during the processing of acoustically degraded speech
(例如, Peelle, 2018) or sentences with syntactic errors (例如, Kuperberg et al., 2003), suggesting
that the MD network may be important for coping with signal corruption, perhaps performing
specific operations aimed at “repairing” the input. 然而, other studies have reported MD
activity during conditions that do not involve corrupted input, both in the domain of language
(例如, Hoffman, Loginova, & 拉塞尔, 2018; Whitney, 柯克, O’Sullivan, Lambon Ralph, & Jefferies,
2012), and for many nonlinguistic tasks (例如, Crittenden & Duncan, 2012; Duncan & 欧文,
2000; Fedorenko et al., 2013; Hugdahl et al., 2015), suggesting perhaps that the contribution is
more general in nature (例如, providing more attentional or working memory resources). At this
时间, it is difficult to put forward mechanistic-level accounts of the MD networks’ contribution to
processing noisy linguistic input.
To conclude, we have provided evidence that constructing complex meanings appears to
be the core linguistic computation implemented in the language-selective frontotemporal net-
工作: providing that computation is engaged (as determined by the combination of input prop-
erties and a plausibly Bayesian inference process) and language brain areas are as active as
when they process their preferred stimulus—well-formed meaningful sentences. 而且,
combinable words have to occur in proximity to one another for the composition mechanisms
to get triggered. Many important questions about linguistic composition remain. 例如,
how strongly is composition driven by our prior experience with particular words versus the
underlying concepts? Is the span over which high mutual information is detected and affects
composition determined by language statistics or by our general memory limitations? And is it
similar between the visual and auditory modalities? How exactly do bottom-up lexicosemantic
and syntactic cues trade off with top-down inferential processes that take into account our
knowledge of language and the world? And how are we able to quickly re-map our world-
knowledge priors when we process fictional or otherwise implausible scenarios (例如,
Nieuwland & van Berkum, 2006)? Despite all these open questions, current work brings us
one step closer to a mechanistic-level account of the computations that the language network
plausibly supports.
致谢
We thank Zuzanna Balewski for help with creating the experimental script for Experiment 1;
EvLab members for help with scanning and helpful discussions; Nancy Kanwisher, Adele
Goldberg, Ray Jackendoff, Leon Bergen, Josh Tenenbaum, Roger Levy, Giousue Baggio, 这
audience at the CUNY2017 Sentence Processing conference, two anonymous reviewers, 和
especially Ted Gibson for comments on this line of work and earlier drafts of the manuscript;
and Martin Schneider for help with collecting the behavioral data for the sentence reconstruc-
tion study. The authors would also like to acknowledge the Athinoula A. Martinos Imaging
Center at the McGovern Institute for Brain Research at MIT, and the support team (Steven
Shannon and Atsushi Takahashi). This work was additionally supported by the Department
of Brain and Cognitive Sciences and the McGovern Institute for Brain Research at MIT.
Neurobiology of Language
129
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
n
哦
/
我
/
我
A
r
t
我
C
e
–
p
d
F
/
/
/
/
1
1
1
0
4
1
8
6
7
9
0
2
n
哦
_
A
_
0
0
0
0
5
p
d
/
.
我
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
Composition drives the language-selective network
资金信息
乙. Fedorenko, 美国国立卫生研究院 (http://dx.doi.org/10.13039/100000002), 奖
IDs: R00-HD-057522, R01-DC-016607 and R01-DC016950.
作者贡献
F. Mollica: 概念化; 形式分析; 可视化; Writing—original draft; Writing—
编辑. 中号. Siegelman: 概念化; 形式分析; 数据采集; Project administra-
的; Writing—editing. 乙. Diachek: 数据采集; 项目管理; Writing—editing.
S. 匹安多糖: 概念化; Writing—editing. Z. Mineroff: 概念化; 数据
收藏; 项目管理. 右. 富特雷尔: 概念化; Writing—editing. H. Kean:
数据采集; 项目管理; Writing—editing. 磷. Qian: 形式分析; Writing—
编辑. 乙. Fedorenko: 概念化; Formal acquisition; 资源; 监督; Writing—
original draft; Writing—editing.
DATA AVAILABILITY
数据, 刺激, and analysis scripts are hosted on the Open Science Framework https://osf.io/y28fz/.
参考
Andrews, M。, Vigliocco, G。, & Vinson, D. (2009). Integrating expe-
riential and distributional data to learn semantic representations.
心理评论, 116(3), 463.
Axer, H。, Klingner, C. M。, & Prescher, A. (2013). Fiber anatomy of
dorsal and ventral language streams. Brain and Language, 127(2),
192–204.
Baddeley, A. (2000). The episodic buffer: A new component of
工作记忆? 认知科学的趋势, 4(11), 417–423.
Baddeley, A. D ., 艾伦, 右. J。, & Hitch, G. J. (2011). Binding in visual work-
ing memory: The role of the episodic buffer. Neuropsychologia,
49(6), 1393–1400.
Baddeley, A. D ., Hitch, G. J。, & 艾伦, 右. J. (2009). Working
memory and binding in sentence recall. Journal of Memory and
语言, 61(3), 438–456.
Baggio, G. (2018). Meaning in the brain. 与新闻界.
Baggio, G。, & Hagoort, 磷. (2011). The balance between memory
and unification in semantics: A dynamic account of the n400.
Language and Cognitive Processes, 26(9), 1338–1367.
Barr, D. J。, 征收, R。, Scheepers, C。, & Tily, H. J. (2013). Random
effects structure for confirmatory hypothesis testing: Keep it max-
imal. 记忆与语言杂志, 68(3), 255–278.
Bautista, A。, & Wilson, S. 中号. (2016). Neural responses to grammat-
ically and lexically degraded speech. 语言, Cognition and
神经科学, 31(4), 567–574.
Bedny, M。, Pascual-Leone, A。, Dodell-Feder, D ., Fedorenko, E., &
Saxe, 右. (2011). Language processing in the occipital cortex of
congenitally blind adults. 国家科学院院刊
科学系, 108(11), 4429–4434.
Bemis, D. K., & Pylkkänen, L. (2011). Simple composition: A
magnetoencephalography investigation into the comprehension
of minimal linguistic phrases. 神经科学杂志, 31(8),
2801–2814.
Berwick, 右. C。, Friederici, A. D ., Chomsky, N。, & Bolhuis, J. J.
(2013). 进化, 脑, and the nature of language. 趋势
认知科学, 17(2), 89–98.
贝弗, 时间. G. (1970). The cognitive basis for linguistic structures.
Cognition and the Development of Language, 279(362), 1–61.
空白的, 我。, Balewski, Z。, Mahowald, K., & Fedorenko, 乙. (2016).
Syntactic processing is distributed across the language system.
神经影像, 127, 307–323.
空白的, 我。, Kanwisher, N。, & Fedorenko, 乙. (2014). A functional dis-
sociation between language and multiple-demand systems
revealed in patterns of bold signal fluctuations. Journal of Neuro-
physiology, 112(5), 1105–1118.
空白的, 我. A。, & Fedorenko, 乙. (2019). No evidence for functional
distinctions across frontotemporal language regions in their
temporal receptive windows. bioRxiv, 712372.
Bornkessel-Schlesewsky, 我。, Schlesewsky, M。, 小的, S. L。, &
Rauschecker, J. 磷. (2015). Neurobiological roots of language in
primate audition: Common computational properties. 趋势
认知科学, 19(3), 142–150.
Brener, 右. (1940). An experimental investigation of memory span.
实验心理学杂志, 26(5), 467.
Brennan, J。, 尼尔, Y。, Hasson, U。, 马拉奇, R。, Heeger, D. J。, &
Pylkkänen, L. (2012). Syntactic structure building in the anterior
temporal lobe during natural story listening. Brain and Language,
120(2), 163–173.
Brysbaert, M。, 新的, B., & Keuleers, 乙. (2012). Adding part-of-speech
information to the subtlex-us word frequencies. Behavior Research
方法, 44(4), 991–997.
巴克纳, 右. L。, Andrews-Hanna, J. R。, & Schacter, D. L. (2008). 这
brain’s default network: Anatomy, function, and relevance to dis-
舒适. Annals of the New York Academy of Sciences, 1124, 1.
布尔克纳, 个人电脑. (2017). 布尔姆斯: 用于贝叶斯多级的 R 包
使用 Stan 的模型. 统计软件杂志, 80(1), 1–28.
Burnard, L. (2000). Reference guide for the British National Corpus
( World Edition). Oxford University Computing Services.
Callicott, J. H。, Mattay, V. S。, Bertolino, A。, 芬恩, K., Coppola, R。,
Frank, J. A。, … Weinberger, D. 右. (1999). Physiological charac-
teristics of capacity constraints in working memory as revealed
by functional MRI. 大脑皮层, 9(1), 20–26.
坎贝尔, K. L。, & Tyler, L. K. (2018). Language-related domain-
specific and domain-general systems in the human brain. 当前的
Opinion in Behavioral Sciences, 21, 132–137.
Neurobiology of Language
130
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
n
哦
/
我
/
我
A
r
t
我
C
e
–
p
d
F
/
/
/
/
1
1
1
0
4
1
8
6
7
9
0
2
n
哦
_
A
_
0
0
0
0
5
p
d
.
/
我
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
Composition drives the language-selective network
Chater, N。, & 曼宁, C. D. (2006). Probabilistic models of lan-
guage processing and acquisition. 认知科学的趋势,
10(7), 335–344.
Chomsky, 氮. (1965). Aspects of the theory of syntax (卷. 11). 和
按.
Chomsky, 氮. (1981). Lectures on government and binding. 多德雷赫特:
Foris.
Chomsky, 氮. (2002). On nature and language. 剑桥大学
按.
Christiansen, 中号. H。, & Chater, 氮. (2016). The now-or-never bot-
tleneck: A fundamental constraint on language. Behavioral and
Brain Sciences, 39.
教会, K. W., & Hanks, 磷. (1990). 词语关联规范, 在-
tual information, 和词典编纂. 计算语言学,
16(1), 22–29.
库克, A。, Grossman, M。, DeVita, C。, Gonzalez-Atavales, J。,
摩尔, P。, 陈, W., … Detre, J. (2006). Large-scale neural
network for sentence processing. Brain and Language, 96(1),
14–36.
Crittenden, 乙. M。, & Duncan, J. (2012). Task difficulty manipulation
reveals multiple demand activity but no frontal lobe hierarchy.
大脑皮层, 24(2), 532–540.
Culicover, 磷. W., & 杰肯道夫, 右. (2005). Simpler syntax. 牛津
University Press on Demand.
Culicover, 磷. W., & 杰肯道夫, 右. (2006). The simpler syntax
假设. 认知科学的趋势, 10(9), 413–418.
戴尔, A. 中号. (1999). Optimal experimental design for event related
功能磁共振成像. 人脑图谱, 8(23), 109–114.
Dapretto, M。, & Bookheimer, S. 是. (1999). Form and content:
Dissociating syntax and semantics in sentence comprehension.
神经元, 24(2), 427–432.
戴尔, G. S。, & 张, F. (2014). The p-chain: Relating sentence pro-
duction and its disorders to comprehension and acquisition.
英国皇家学会哲学汇刊 B: Biological
科学, 369(1634).
Diachek, E., 空白的, 我。, Siegelman, M。, & Fedorenko, 乙. (2019). 这
domain-general multiple demand (医学博士) network does not sup-
port core aspects of language comprehension: 一个大规模的
fMRI investigation. bioRxiv, 744094.
Dryer, 中号. S。, & Haspelmath, 中号. (2013). The world atlas of language
结构. Max Planck Digital Library.
Duffau, H。, Moritz-Gasser, S。, & Mandonnet, 乙. (2014). A re-examination
of neural basis of language processing: Proposal of a dynamic ho-
dotopical model from data provided by brain stimulation map-
ping during picture naming. Brain and Language, 131, 1–10.
Duncan, J. (2010). The multiple-demand (医学博士) system of the pri-
mate brain: Mental programs for intelligent behaviour. Trends
in Cognitive Sciences, 14(4), 172–179.
Duncan, J. (2013). The structure of cognition: Attentional episodes
in mind and brain. 神经元, 80(1), 35–50.
Duncan, J。, & 欧文, A. 中号. (2000). Common regions of the human
frontal lobe recruited by diverse cognitive demands. 趋势
神经科学, 23(10), 475–483.
Embick, D ., Marantz, A。, Miyashita, Y。, O’Neil, W., & Sakai, K. L.
(2000). A syntactic specialization for Broca’s area. 会议记录
of the National Academy of Sciences, 97(11), 6150–6154.
Fano, 右. 中号. (1961). Transmission of information: A statistical theory
of communications. 与新闻界.
Federmeier, K. D. (2007). Thinking ahead: The role and roots of
prediction in language comprehension. Psychophysiology, 44(4),
491–505.
Fedorenko, 乙. (2014). The role of domain-general cognitive control
in language comprehension. 心理学前沿, 5, 335.
Fedorenko, E., Behr, 中号. K., & Kanwisher, 氮. (2011). Functional
specificity for high-level linguistic processing in the human brain.
美国国家科学院院刊.
Fedorenko, E., 空白的, 我。, Siegelman, M。, & Mineroff, Z. (2020). Lack
of selectivity for syntax relative to word meanings throughout the
language network. bioRxiv. https://doi.org/10.1101/477851
Fedorenko, E., Duncan, J。, & Kanwisher, 氮. (2013). Broad domain
generality in focal regions of frontal and parietal cortex. Pro-
ceedings of the National Academy of Sciences, 201315235.
Fedorenko, E., Hsieh, P.-J., Nieto-Castañón, A。, Whitfield-Gabrieli,
S。, & Kanwisher, 氮. (2010). New method for fMRI investigations
语言的: Defining ROIs functionally in individual subjects.
神经生理学杂志, 104(2), 1177–1194.
Fedorenko, E., Nieto-Castañón, A。, & Kanwisher, 氮. (2012). 词汇
and syntactic representations in the brain: An fMRI investigation with
multi-voxel pattern analyses. Neuropsychologia, 50(4), 499–513.
Fedorenko, E., 斯科特, 时间. L。, Brunner, P。, Coon, 瓦. G。, 普里切特, B.,
Schalk, G。, & Kanwisher, 氮. (2016). Neural correlate of the con-
struction of sentence meaning. 国家会议录
Academy of Sciences, 113(41), E6256–E6262.
费雷拉, F。, 贝利, K. G。, & 费拉罗, V. (2002). 足够好的代表-
语言理解中的反感. 当前方向
心理科学, 11(1), 11–15.
Ferstl, 乙. C。, 诺伊曼, J。, Bogler, C。, & von Cramon, D. 是. (2008).
The extended language network: A meta-analysis of neuro-
imaging studies on text comprehension. 人脑图谱,
29(5), 581–593.
Ferstl, 乙. C。, & von Cramon, D. 是. (2001). The role of coherence
and cohesion in text comprehension: An event-related fMRI
学习. Cognitive Brain Research, 11(3), 325–340.
Friederici, A. D. (2011). The brain basis of language processing: 从
structure to function. Physiological Reviews, 91(4), 1357–1392.
Friederici, A. D. (2012). The cortical language circuit: From audi-
tory perception to sentence comprehension. Trends in Cognitive
科学, 16(5), 262–268.
Friederici, A. D ., Bahlmann, J。, Heim, S。, Schubotz, 右. 我。, & 安旺德,
A. (2006). The brain differentiates human and non-human grammars:
Functional localization and structural connectivity. 会议记录
of the National Academy of Sciences, 103(7), 2458–2463.
Friederici, A. D ., Opitz, B., & von Cramon, D. 是. (2000).
Segregating semantic and syntactic aspects of processing in the
人脑: An fMRI investigation of different word types.
大脑皮层, 10(7), 698–705.
富特雷尔, R。, Mahowald, K., & 吉布森, 乙. (2015). Large-scale
evidence of dependency length minimization in 37 语言.
美国国家科学院院刊, 112(33),
10336–10341.
富特雷尔, R。, Qian, P。, 吉布森, E., Fedorenko, E., & 空白的, 我. (2019).
Syntactic dependencies correspond to word pairs with high mutual
信息. In Proceedings of the Fifth International Conference on
Dependency Linguistics (PP. 3–13).
吉布森, 乙. (2000). The dependency locality theory: A distance-
based theory of linguistic complexity. 图像, 语言, Brain, 2000,
95–126.
吉布森, 乙. (1998). Linguistic complexity: Locality of syntactic de-
pendencies. 认识, 68(1), 1–76.
吉布森, E., 卑尔根, L。, & 匹安多糖, S. 时间. (2013). Rational integra-
tion of noisy evidence and prior semantic expectations in sen-
tence interpretation. 美国国家科学院院刊
科学, 201216438.
吉布森, E., 富特雷尔, R。, Piandadosi, S. T。, Dautriche, 我。, Mahowald,
K., 卑尔根, L。, & 征收, 右. (2019). How efficiency shapes human
语言. 认知科学的趋势, 23(5), 389–407.
Neurobiology of Language
131
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
n
哦
/
我
/
我
A
r
t
我
C
e
–
p
d
F
/
/
/
/
1
1
1
0
4
1
8
6
7
9
0
2
n
哦
_
A
_
0
0
0
0
5
p
d
/
.
我
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
Composition drives the language-selective network
Goldberg, A. 乙. (2006). Constructions at work: The nature of gen-
eralization in language. Oxford University Press on Demand.
Griffiths, 时间. L。, Steyvers, M。, & Tenenbaum, J. 乙. (2007). Topics in
语义表示. Psychological Reviews, 114(2), 211.
黑尔, J. (2001). A probabilistic early parser as a psycholinguistic
模型. In Proceedings of the Second Meeting of the North American
Chapter of the Association for Computational Linguistics on Lan-
guage Technologies (PP. 1–8).
黑尔, K. (1983). Warlpiri and the grammar of non-configurational
语言. 自然语言 & Linguistic Theory, 1(1), 5–47.
Hasson, U。, 哪个, E., Vallines, 我。, Heeger, D. J。, & 鲁宾, 氮. (2008).
A hierarchy of temporal receptive windows in human cortex.
神经科学杂志, 28(10), 2539–2550.
Hoffman, P。, Loginova, E., & 拉塞尔, A. (2018). Poor coherence in
older people’s speech is explained by impaired semantic and
executive processes. 电子生活, 7, e38907.
Hugdahl, K., Raichle, 中号. E., 米特拉, A。, & Specht, K. (2015). 上
existence of a generalized non-specific task-dependent network.
Frontiers in Human Neuroscience, 9, 430.
Hultén, A。, 锄, J.-M。, Uddén, J。, Lam, 氮. H。, & Hagoort, 磷.
(2019). How the brain makes sense beyond the processing of
single words–An MEG study. 神经影像, 186, 586–594.
汉弗莱斯, C。, Binder, J. R。, Medler, D. A。, & Liebenthal, 乙. (2007).
Time course of semantic processes during sentence comprehen-
锡安: An fMRI study. 神经影像, 36(3), 924–932.
杰肯道夫, 右. (2010). Meaning and the lexicon: The parallel archi-
tecture 1975–2010. 牛津: 牛津大学出版社.
杰肯道夫, 右. (2007). A parallel architecture perspective on lan-
guage processing. 脑研究, 1146, 2–22.
杰肯道夫, 右. (2011). What is the human language faculty? 二
意见. 语言, 586–624.
杰肯道夫, R。, & 杰肯道夫, 右. S. (2002). Foundations of language:
Brain, 意义, grammar, evolution. 牛津: 牛津大学出版社.
杰肯道夫, R。, & Wittenberg, 乙. (2014). What you can say without
syntax: A hierarchy of grammatical complexity. 在F中. Newmeyer
& L. 普雷斯顿 (编辑。), Measuring linguistic complexity (PP. 65–82).
牛津: 牛津大学出版社.
杰肯道夫, R。, & Wittenberg, 乙. (2017). Linear grammar as a pos-
sible stepping-stone in the evolution of language. Psychonomic
Bulletin & 审查, 24(1), 219–224.
Jacoby, N。, & Fedorenko, 乙. (2018). Discourse-level comprehension
engages medial frontal theory of mind brain regions even for expos-
itory texts. 语言, Cognition and Neuroscience, 1–17.
Julian, J. B., Fedorenko, E., 韦伯斯特, J。, & Kanwisher, 氮. (2012). 一个
algorithmic method for functionally defining regions of interest in
the ventral visual pathway. 神经影像, 60(4), 2357–2364.
凯勒, 时间. A。, Carpenter, 磷. A。, & Just, 中号. A. (2001). The neural bases
of sentence comprehension: A fMRI examination of syntactic and
lexical processing. 大脑皮层, 11(3), 223–237.
Keuleers, E., & Brysbaert, 中号. (2010). Wuggy: A multilingual pseu-
doword generator. Behavior Research Methods, 42(3), 627–633.
Kimball, J. (1973). Seven principles of surface structure parsing in
自然语言. 认识, 2(1), 15–47.
Kriegeskorte, N。, Simmons, 瓦. K., Bellgowan, 磷. S。, & 贝克, C. 我.
(2009). Circular analysis in systems neuroscience: the dangers of
double dipping. 自然神经科学, 12(5), 535.
Kuperberg, G. 右. (2007). Neural mechanisms of language compre-
hension: Challenges to syntax. 脑研究, 1146, 23–49.
Kuperberg, G. R。, Caplan, D ., Sitnikova, T。, Eddy, M。, & Holcomb,
磷. J. (2006). Neural correlates of processing syntactic, semantic,
and thematic relationships in sentences. Language and Cognitive
Processes, 21(5), 489–530.
Kuperberg, G. R。, Holcomb, 磷. J。, Sitnikova, T。, Greve, D ., 戴尔,
A. M。, & Caplan, D. (2003). Distinct patterns of neural modula-
tion during the processing of conceptual and syntactic anoma-
谎言. 认知神经科学杂志, 15(2), 272–293.
Kuperberg, G. R。, & Jaeger, 时间. F. (2016). What do we mean by pre-
diction in language comprehension? 语言, Cognition and
神经科学, 31(1), 32–59.
Lambon Ralph, 中号. A。, Jefferies, E., 帕特森, K., & 罗杰斯, 时间. 时间.
(2017). The neural and computational bases of semantic cogni-
的. 自然评论神经科学, 18(1), 42.
Lerner, Y。, 蜂蜜, C. J。, Silbert, L. J。, & Hasson, U. (2011). Topographic
mapping of a hierarchy of temporal receptive windows using a
narrated story. 神经科学杂志, 31(8), 2906–2915.
征收, 右. (2008A). 基于期望的句法理解.
认识, 106(3), 1126–1177.
征收, 右. (2008乙). A noisy-channel model of rational human sen-
tence comprehension under uncertain input. 在诉讼程序中
the conference on empirical methods in natural language pro-
cessing (PP. 234–243).
征收, R。, 比克内尔, K., 斯莱特里, T。, & 雷纳, K. (2009). Eye movement
evidence that readers maintain and act on uncertainty about past
linguistic input. 美国国家科学院院刊,
106(50), 21086–21090.
Lewis, D. D ., & 琼斯, K. S. (1996). Natural language processing for
信息检索. ACM 通讯, 39(1), 92–101.
李, 瓦. (1990). Mutual information functions versus correlation
功能. Journal of Statistical Physics, 60(5–6), 823–837.
林, H。, & Tegmark, 中号. (2017). Critical behavior in physics and
probabilistic formal languages. Entropy, 19(7), 299.
林登, D. E., Bittner, 右. A。, Muckli, L。, 华尔兹, J. A。, Kriegeskorte, N。,
Goebel, R。, … Munk, 中号. H. (2003). Cortical capacity constraints for
visual working memory: Dissociation of fMRI load effects in a
fronto-parietal network. 神经影像, 20(3), 1518–1530.
Mahowald, K., & Fedorenko, 乙. (2016). Reliable individual-level
neural markers of high-level language processing: A necessary
precursor for relating neural variability to behavioral and genetic
variability. 神经影像, 139, 74–93.
Matchin, W., & Hickok, G. (2019). The cortical organization of
syntax. 大脑皮层, 1–18.
Mesulam, M.-M., Wieneke, C。, Hurley, R。, Rademaker, A。, 汤普森,
C. K., Weintraub, S。, & Rogalski, 乙. J. (2013). Words and objects at
the tip of the left temporal lobe in primary progressive aphasia.
Brain, 136(2), 601–618.
Michalon, 奥。, & Baggio, G. (2019). Meaning-driven syntactic
predictions in a parallel processing architecture: Theory and
algorithmic modeling of ERP effects. Neuropsychologia, 131,
171–183.
Michel, J.-B。, 沉, 是. K., Aiden, A. P。, Veres, A。, Gray, 中号. K.,
皮克特, J. P。, …… & Pinker, S. (2010). Quantitative analysis of cul-
ture using millions of digitized books. 科学, 1199644.
米科洛夫, T。, 吸勺, 我。, 陈, K., 科拉多, G. S。, & 院长, J.
(2013). Distributed representations of words and phrases and
their compositionality. In Advances in neural information pro-
cessing systems (PP. 3111–3119).
磨坊主, G. A。, & Chomsky, 氮. (1963). Finitary models of language
用户. Finitary models of language users. 在R中. D. Luce, 右. 右. 衬套,
and E. 加兰特, (编辑。), Handbook of Mathematical Psychology
(PP. 419–492). 霍博肯, 新泽西州: 威利.
Mineroff, Z。, 空白的, 我. A。, Mahowald, K., & Fedorenko, 乙. (2018). A
robust dissociation among the language, multiple demand, 和
default mode networks: Evidence from inter-region correlations
in effect size. Neuropsychologia, 119, 501–511.
Neurobiology of Language
132
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
n
哦
/
我
/
我
A
r
t
我
C
e
–
p
d
F
/
/
/
/
1
1
1
0
4
1
8
6
7
9
0
2
n
哦
_
A
_
0
0
0
0
5
p
d
/
.
我
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
Composition drives the language-selective network
Mirault, J。, Snell, J。, & Grainger, J. (2018). You that read wrong
再次! A transposed-word effect in grammaticality judgments.
心理科学. https://doi.org/10.1177/0956797618806296
Mollica, F。, Siegelman, M。, Diachek, E., 富特雷尔, R。, 匹安多糖, S。,
Mineroff, Z。, & Fedorenko, 乙. (2019). High local mutual informa-
tion drives the response in the human language network [Code
发电机]. Retrieved from osf.io/y28fz
Montague, 右. (1974). Formal philosophy. 耶鲁大学出版社.
Nieto-Castañón, A。, & Fedorenko, 乙. (2012). Subject-specific func-
tional localizers increase sensitivity and functional resolution of
multi-subject analyses. 神经影像, 63(3), 1646–1669.
Nieuwland, 中号. S。, & Van Berkum, J. J. (2006). When peanuts fall in
love: N400 evidence for the power of discourse. 杂志
Cognitive Neuroscience, 18(7), 1098–1111.
Noppeney, U。, & Price, C. J. (2004). Retrieval of abstract semantics.
神经影像, 22(1), 164–170.
Oldfield, 右. C. (1971). The assessment and analysis of handedness:
The Edinburgh inventory. Neuropsychologia, 9(1), 97–113.
Overath, T。, 麦克德莫特, J. H。, Zarate, J. M。, & Poeppel, D. (2015).
The cortical analysis of speech-specific temporal structure revealed
by responses to sound quilts. 自然神经科学, 18(6), 903.
Pallier, C。, Devauchelle, A.-D., & 德阿内, S. (2011). Cortical rep-
resentation of the constituent structure of sentences. 会议记录
of the National Academy of Sciences, 201018711.
Partee, 乙. (1995). Lexical semantics and compositionality. 一个
Invitation to Cognitive Science: 语言, 1, 311–360.
Partee, 乙. (1975). Montague grammar and transformational grammar.
Linguistic Inquiry, 203–300.
Partee, 乙. B., ter Meulen, A. G。, & Wall, 右. (1990). Mathematical
methods in linguistics (卷. 30). Springer Science & Business Media.
帕特森, K., Nestor, 磷. J。, & 罗杰斯, 时间. 时间. (2007). Where do you know
what you know? The representation of semantic knowledge in the
人脑. 自然评论神经科学, 8(12), 976.
Paunov, A. M。, 空白的, 我. A。, & Fedorenko, 乙. (2019). Functionally
distinct language and theory of mind networks are synchronized
at rest and during language comprehension. Journal of Neuro-
physiology, 121(4), 1244–1265.
Peelle, J. 乙. (2018). Listening effort: How the cognitive consequences
of acoustic challenge are reflected in brain and behavior. Ear and
Hearing, 39(2), 204.
Pennington, J。, Socher, R。, & 曼宁, C. (2014). Glove: Global vectors
for word representation. 在诉讼程序中 2014 conference on
empirical methods in natural language processing (PP. 1532–1543).
皮克林, 中号. J。, & Garrod, S. (2013). An integrated theory of lan-
guage production and comprehension. Behavioral and Brain
科学, 36(4), 329–347.
Poldrack, 右. A。, 贝克, C. 我。, Durnez, J。, Gorgolewski, K. J。, Matthews,
磷. M。, Munafò, 中号. R。, … Yarkoni, 时间. (2017). Scanning the horizon:
Towards transparent and reproducible neuroimaging research.
自然评论神经科学, 18(2), 115.
Potter, 中号. C。, & Lombardi, L. (1990). Regeneration in the short-term re-
call of sentences. 记忆与语言杂志, 29(6), 633–654.
Pylkkänen, L. (2016). Composition of complex meaning: 国际米兰-
disciplinary perspectives on the left anterior temporal lobe. 在
Neurobiology of language (PP. 621–631). 爱思唯尔.
Pylkkänen, L. (2019). The neural basis of combinatory syntax and
语义学. 科学, 366(6461), 62–66.
Pylkkänen, L。, Bemis, D. K., & Elorrieta, 乙. 乙. (2014). 建筑
phrases in language production: An MEG study of simple com-
位置. 认识, 133(2), 371–384.
Pylkkänen, L。, & Brennan, J. 右. (in press). Composition: The neuro-
biology of syntactic and semantic structure building. In The cog-
nitive neurosciences. 与新闻界.
R Team. (2017). 右: A language and environment for statistical com-
puting [Software Tool]. 维也纳, 奥地利: R Foundation for
Statistical Computing.
Rabovsky, M。, 汉森, S. S。, & 麦克莱兰, J. L. (2018). Modelling
the n400 brain potential as change in a probabilistic representa-
tion of meaning. Nature Human Behaviour, 2(9), 693.
Röder, B., Stock, 奥。, Neville, H。, Bien, S。, & Rösler, F. (2002). Brain
activation modulated by the comprehension of normal and pseudo-
word sentences of different processing demands: A functional mag-
netic resonance imaging study. 神经影像, 15(4), 1003–1014.
Saur, D ., Kreher, 乙. W., Schnell, S。, Kümmerer, D ., Kellmeyer, P。,
Vry, M.-S., … others. (2008). Ventral and dorsal pathways for
语言. 美国国家科学院院刊,
0805234105.
Saxe, R。, & Kanwisher, 氮. (2003). People thinking about thinking
人们: The role of the temporo-parietal junction in “theory of
mind.” NeuroImage, 19(4), 1835–1842.
Schotter, 乙. R。, 比克内尔, K., 霍华德, 我。, 征收, R。, & 雷纳, K.
(2014). Task effects reveal cognitive flexibility responding to
frequency and predictability: Evidence from eye movements in
reading and proofreading. 认识, 131(1), 1–27.
施瓦茨, 中号. F。, Kimberg, D. Y。, 沃克, G. M。, Faseyitan, 奥。, Brecher,
A。, 戴尔, G. S。, & Coslett, H. 乙. (2009). Anterior temporal involve-
ment in semantic word retrieval: Voxel-based lesion-symptom
mapping evidence from aphasia. Brain, 132(12), 3411–3427.
施瓦茨, 中号. F。, Marin, 氧. S。, & 藏红花, 乙. 中号. (1979). Dissociations
of language function in dementia: 案例研究. Brain and Lan-
规格, 7(3), 277–306.
斯科特, 时间. L。, Gallée, J。, & Fedorenko, 乙. (2017). A new fun and ro-
bust version of an fMRI localizer for the frontotemporal language
系统. Cognitive Neuroscience, 8(3), 167–176.
Shannon, C. E., & Weaver, 瓦. (1963). The mathematical theory of
沟通. 1949. 伊利诺伊大学出版社.
Siegelman, M。, 空白的, 我. A。, Mineroff, Z。, & Fedorenko, 乙. (2019).
An attempt to conceptually replicate the dissociation between
syntax and semantics during sentence comprehension. Neuro-
科学, 413, 219–229.
史密斯, 氮. J. (2014). ZS: A file format for efficiently distributing, 使用,
and archiving record-oriented data sets of any size. Manuscript
submitted for publication. School of Informatics, 大学
爱丁堡. Retrieved from http://vorpus.org/papers/draft/zs-paper.pdf
Stan Development Team. (2018). RStan: The R interface to Stan.
(R包版本 2.17.2.). Retrieved from https://mc-stan.org
Steyvers, M。, Griffiths, 时间. L。, & 丹尼斯, S. (2006). Probabilistic
inference in human semantic memory. Trends in Cognitive
科学, 10(7), 327–334.
Thesen, S。, Heid, 奥。, Mueller, E., & Schad, L. 右. (2000). Prospective
acquisition correction for head motion with image-based track-
ing for real-time fMRI. Magnetic Resonance in Medicine: 一个
Official Journal of the International Society for Magnetic Reso-
nance in Medicine, 44(3), 457–465.
Traxler, 中号. J. (2014). Trends in syntactic parsing: Anticipation,
Bayesian estimation, and good-enough parsing. Trends in Cog-
nitive Sciences, 18(11), 605–611.
Tyler, L. K., Marslen-Wilson, 瓦. D ., Randall, B., 赖特, P。, Devereux, 乙. J。,
Zhuang, J。, … Stamatakis, 乙. A. (2011). Left inferior frontal cortex
and syntax: Function, structure and behaviour in patients with left
hemisphere damage. Brain, 134(2), 415–431.
Uddén, J。, Hultén, A。, Bendtz, K., Mineroff, Z。, Kucera, K. S。, Vino, A。, ……
Fisher, S. 乙. (2019). Toward robust functional neuroimaging genetics
of cognition. 神经科学杂志, 39(44), 8778–8787.
Ullman, 中号. 时间. (2004). Contributions of memory circuits to language:
The declarative/procedural model. 认识, 92(1–2), 231–270.
Neurobiology of Language
133
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
n
哦
/
我
/
我
A
r
t
我
C
e
–
p
d
F
/
/
/
/
1
1
1
0
4
1
8
6
7
9
0
2
n
哦
_
A
_
0
0
0
0
5
p
d
/
.
我
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
Composition drives the language-selective network
Ullman, 中号. 时间. (2016). The declarative/procedural model: A neuro-
biological model of language learning, 知识, and use. 在
Neurobiology of language (PP. 953–968). 爱思唯尔.
Vagharchakian, L。, Dehaene-Lambertz, G。, Pallier, C。, & 德阿内,
S. (2012). A temporal bottleneck in the language comprehension
网络. 神经科学杂志, 32(26), 9089–9102.
Visser, M。, Jefferies, E., & Lambon Ralph, 中号. (2010). Semantic pro-
cessing in the anterior temporal lobes: A meta-analysis of the func-
tional neuroimaging literature. 认知神经科学杂志,
22(6), 1083–1094.
Weide, 右. (1998). The CMU pronunciation dictionary, 发布 0.6.
卡内基梅隆大学.
Wen, T。, 米切尔, D. J。, & Duncan, J. (2018). Response of the
multiple-demand network during simple stimulus discrimina-
系统蒸发散. 神经影像, 177, 79–87.
Westerlund, M。, & Pylkkänen, L. (2017). How does the left anterior
temporal lobe contribute to conceptual combination? interdisci-
基本观点. In Compositionality and concepts in linguis-
tics and psychology (PP. 269–290). 占婆, 瑞士: 施普林格.
Whitney, C。, 柯克, M。, O’Sullivan, J。, Lambon Ralph, 中号. A。, &
Jefferies, 乙. (2012). Executive semantic processing is underpinned
by a large-scale neural network: Revealing the contribution of
left prefrontal, posterior temporal, and parietal cortex to con-
trolled retrieval and selection using TMS. 认知杂志
神经科学, 24(1), 133–147.
Willems, 右. M。, van der Haegen, L。, Fisher, S. E., & Francks, C.
(2014). 另一方面: Including left-handers in cognitive
neuroscience and neurogenetics. 自然评论神经科学,
15(3), 193.
赖特, P。, Randall, B., Marslen-Wilson, 瓦. D ., & Tyler, L. K.
(2011). Dissociating linguistic and task- related activity in the left
inferior frontal gyrus. 认知神经科学杂志, 23(2),
404–413.
Yngve, V. H. (1960). A model and an hypothesis for language struc-
真实. Proceedings of the American Philosophical Society, 104(5),
444–466.
Zacks, J. M。, & Tversky, 乙. (2001). Event structure in perception and
conception. Psychological Bulletin, 127(1), 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
n
哦
/
我
/
我
A
r
t
我
C
e
–
p
d
F
/
/
/
/
1
1
1
0
4
1
8
6
7
9
0
2
n
哦
_
A
_
0
0
0
0
5
p
d
.
/
我
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
Neurobiology of Language
134