Partial Truths: Adults Choose to Mention Agents - 麻省理工学院人工智能研究专业

Partial Truths: Adults Choose to Mention Agents
and Patients in Proportion to Informativity, Even If
It Doesn’t Fully Disambiguate the Message

Melissa Kline1, Laura Schulz1, and Edward Gibson1

1Brain and Cognitive Sciences, 麻省理工学院

关键词: pragmatics, argument structure, common ground, 沟通

开放访问

杂志

抽象的

How do we decide what to say to ensure our meanings will be understood? The Rational
Speech Act model (RSA; Frank & 古德曼, 2012) asserts that speakers plan what to say by
comparing the informativity of words in a particular context. We present the ﬁrst example of
an RSA model of sentence-level (who-did-what-to-whom) 含义. In these contexts, 这
set of possible messages must be abstracted from entities in common ground (people and
物体) to possible events (Jane eats the apple, Marco peels the banana), with each word
contributing unique semantic content. How do speakers accomplish the transformation from
context to compositional, informative messages? In a communication game, 参与者
described transitive events (例如, Jane pets the dog), with only two words, in contexts where
two words either were or were not enough to uniquely identify an event. Adults chose
utterances matching the predictions of the RSA even when there was no possible fully
“successful” utterance. Thus we show that adults’ communicative behavior can be described
by a model that accommodates informativity in context, beyond the set of possible entities in
common ground. This study provides the ﬁrst evidence that adults’ language production is
做作的, at the level of argument structure, by the graded informativity of possible utterances
in context, and suggests that full-blown natural speech may result from speakers who model
and adapt to the listener’s needs.

引文: Kline, M。, Schulz, L。, &
吉布森, 乙. (2017). Partial Truths:
Adults Choose to Mention Agents and
Patients in Proportion to Informativity,
Even If It Doesn’t Fully Disambiguate
the Message. 开放的心态: Discoveries
in Cognitive Science, 2(1), 1–13.
https://doi.org/10.1162/opmi_a_00013

DOI:
https://doi.org/10.1162/opmi_a_00013

介绍

补充材料:
https://doi.org/10.17605/OSF.IO/AWGH6

已收到: 28 十月 2016
公认: 28 七月 2017

利益争夺: 作者有
no competing interests to declare.

通讯作者:
Melissa Kline
mekline@mit.edu

版权: © 2017
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

Communication requires continually making decisions about what information to include and
exclude. It is not always necessary to fully describe an event: If someone asks, What are you
doing? then I’m eating might be sufﬁcient, and possibly preferable to longer alternatives like
I’m eating a sandwich or I’m eating a grilled cheese sandwich. For a speaker to successfully
communicate with a listener in this way, the two need to implicitly agree on some shared
principles of communication. Grice ( 1975) codiﬁed these conversational assumptions as a
series of “maxims,” including the maxims of Quantity (“give as much information as is needed,
but no more”) and Relevance (“say something that furthers the goal of the conversation”). 因此
a speaker can refer to a sandwich alone if the alternative is a salad, but should refer to a grilled
cheese sandwich if the alternative is peanut butter and jelly.

As listeners, adults understand language in part by using statistical information to
predict upcoming words and structures (阿尔特曼 & Kamide, 1999; 征收, 2008; MacDonald,
2013; MacDonald, Pearlmutter, & 塞登伯格, 1994; Tanenhaus, Spivey-Knowlton, 艾伯哈德,
& Sedivy, 1995; 比照. Kuperberg & Jaeger, 2016, for a recent review). How does this predicting

我

D
哦
w
n
哦
A
d
e
d

F
r
哦
米
H

t
t

:
/
/

d
我
r
e
C
t
.

米

我
t
.

e
d
你
哦
p
米

我
/

我

A
r
t
我
C
e
–
p
d

F
/

2
1
1
1
8
6
8
3
0
7
哦
p
米
_
A
_
0
0
0
1
3
p
d

我

乙
y
G
你
e
s
t

哦
n
0
7
S
e
p
e
米
乙
e
r
2
0
2
3

Partial Truths Kline, Schulz, 吉布森

listener operate? Listeners could simply expect similar language to occur in similar contexts,
without regard to the speaker’s motives. But more speciﬁcally, they could expect speakers to
behave predictably because they expect them to behave helpfully. A recent formalization of
this latter hypothesis is the Rational Speech Act model (RSA), which is based around a coop-
erative speaker-listener pair (Frank & 古德曼, 2012). Speakers attempt to maximize the
information transferred to the listener, and listeners succeed by assuming that the speaker is
doing this. Rational Speech Act models successfully predict a variety of phenomena in prag-
matics including the interpretation of scalar implicatures, hyperbole, and metaphor (古德曼
& Frank, 2016; 古德曼 & Stuhlmüller, 2013; Kao, 征收, & 古德曼, 2013; Kao, 吴,
卑尔根, & 古德曼, 2014). Are listeners warranted in making these generous assump-
tions about speakers? Many features of language production seem to be shaped to improve
the chances of successful communication. Formal approaches based on information theory
(Shannon, 1949) have been used to successfully explain reduction and omission phenom-
ena in natural language production including phonological reduction, lexical choice (例如,
math/mathematics), and inclusion of optional arguments (Aylett & Turk, 2004; Jaeger, 2010;
Mahowald, Fedorenko, 匹安多糖, & 吉布森, 2013; Resnik, 1996; van Son & van Santen,
2005; though see Keysar, Barr, & Horton, 1998).

If production is driven by the value to the listener rather than the costs to the speaker,
then the speaker should ﬂexibly adapt when the (linguistic or nonlinguistic) context changes.
For the speciﬁc case of referring expressions (那, that big sandwich), there is a large body of
work showing that speakers’ choices are related to available nonlinguistic information (例如,
Brennan & 克拉克, 1996; Brown-Schmidt & Tanenhaus, 2008; Nadig & Sedivy, 2002; Pogue,
Kurumada & Tanenhaus, 2016; Sedivy, 2003). This is taken as evidence of an awareness of
listeners’ needs because the language production cost of that big sandwich is presumably the
same across contexts, while the beneﬁt to the listener is considerable when there are many
sandwiches, but null if the listener can already pick out the lone sandwich. Speakers do this
even when a listener would need to make inferences about a speaker’s intention to succeed:
In a context with a blue circle as a target with a blue square and a green square as distractors,
adults limited to a single word produce CIRCLE to identify the target object, not BLUE: 虽然
BLUE is a good description of the target in isolation, it could also refer to the blue square
(Frank & 古德曼, 2012).

But human language goes beyond referring expressions for objects: Sentences express
entire propositions about the world (Ben is eating my grilled cheese sandwich). Deriving the set
of possible propositions (not just possible object referents) would seem to require an extensive
understanding of both world knowledge and the ways that conversations tend to unfold (比照.
Ginzburg, 1996). Even once a particular proposition has been chosen, we have many choices
about how to encode it in a sentence. We make choices about argument structure and verb
identify (he ate it/he put it in his mouth), and language provides many ways to omit or limit how
much we say in conveying a proposition, including pronouns (Ben/he ate the sandwich), ellip-
姐姐 (Ben ate the sandwich, and then a cookie), passive constructions (The sandwich was eaten),
and optional arguments (Ben ate the sandwich [with a fork and knife]). These options are used
务实地: Speakers tend to (A) omit or reduce information the listener can retrieve from
linguistic context, (乙) converge with dialog partners on syntactic alternations, 和 (C) 包括
optional material when listeners might otherwise go astray (Brennan & Hanna, 2009; 加拉塔
& Brennan, 2010; Horton, 2005; Kurumada & Jaeger, 2015; 皮克林 & Garrod, 2004).

Relatively little attention has been paid to how speakers use nonlinguistic information to
produce informative sentences. Do we attempt to communicate sentence-level meaning using

开放的心态: 认知科学的发现

我

D
哦
w
n
哦
A
d
e
d

F
r
哦
米
H

t
t

:
/
/

d
我
r
e
C
t
.

米

我
t
.

e
d
你
哦
p
米

我
/

我

A
r
t
我
C
e
–
p
d

F
/

2
1
1
1
8
6
8
3
0
7
哦
p
米
_
A
_
0
0
0
1
3
p
d

我

乙
y
G
你
e
s
t

哦
n
0
7
S
e
p
e
米
乙
e
r
2
0
2
3

Partial Truths Kline, Schulz, 吉布森

something like the rational speaker model, tailoring what we say to the surrounding context? 在
least one study suggests this may be the case: Lockridge and Brennan (2002) had participants
describe scenes with either typical or atypical instruments (He stabbed him with a knife/an
icepick) to a naïve listener. In an unconstrained storytelling task, speakers were more likely
to mention atypical than typical instruments, especially when the listener could not see the
事件. 然而, understanding event descriptions is challenging exactly because events are
transitory—they don’t “stick around” in the context like objects do, and references to events
often occur when the event itself is in the past or future (Gleitman, 1990). 因此, while this
study suggests speakers are sensitive to how world knowledge impacts linguistic informativity,
it does not address the ﬁt between production and particular nonlinguistic contexts: 骗局-
trast in that study is between not seeing the event (the usual scenario) and seeing the event
as it is being described (which listeners usually can’t).

To begin understanding how speakers use nonlinguistic context to decide what to say
about an event, we focus on a single class of basic propositions: transitive sentences like John
feeds the dog. While this construction is used for many classes of verbs, we consider prototyp-
ical cases in which the basic meaning involves an agent performing some action on a patient.
Even assuming the speaker is referring to an event that might occur (or has recently occurred)
in the immediate context, the set of possible messages is potentially inﬁnite (Quine, 1960).
这里, we leave aside the question of word-to-concept mapping and focus on the question of
possible events given a set of possible participants. A speaker trying to design an informative
event utterance must consider not only the possible verbs, but also what referent could cor-
respond to each argument position. We can represent the number of possible events as the
product of the possible verbs, 代理人, and patients:

{约翰, Sue, 乔治, Maria, Jenny} × {feeds, chases, pets}×
{the dog, the cat} = 30 事件

(1)

We use this logic to create “toy” worlds in which there are always exactly seven entities (人们
和物体), and the messages to be communicated are interactions between these entities
(例如, John feeds the dog).

In natural speech, both agents and patients can sometimes be omitted from transitive
descriptions. Many transitive verbs can be used intransitively, 例如, We’ll eat in the
kitchen,1 and many languages also allow noun phrases in subject position to be omitted rela-
tively freely (例如, in Spanish Comió bocadillos, [他] ate sandwiches). 用英语, these kinds
of subject omissions require speciﬁc discourse context (例如, a command, Don’t eat in the
kitchen). We therefore use a production task that restricts the producer to exactly two words,
forcing participants to make the choice to omit at least one element (agent, 病人, 或动词). 在
most object reference studies (比照. Brennan & 克拉克, 1996; Brown-Schmidt & Tanenhaus, 2008;
Nadig & Sedivy, 2002; Pogue et al., 2016; Sedivy, 2003), a noun phrase like my sandwich
or my grilled-cheese sandwich is assumed to be informative when it uniquely identiﬁes one
out of several referents in the context, underinformative if it could apply to more than one ob-
ject (two such sandwiches), and overinformative if it includes additional modiﬁers ( my grilled
cheese sandwich when there is only one sandwich). Rational Speech Act models assume a
richer sense of “informativity” in which words are informative to the extent that they reduce

1 These unergative alternations can be contrasted with unaccusative intransitive alternations like John broke
the lamp/The lamp broke; here we focus solely on the inclusion or omission of agents and patients rather than
on the argument structure or syntactic behavior of particular verbs.

开放的心态: 认知科学的发现

我

D
哦
w
n
哦
A
d
e
d

F
r
哦
米
H

t
t

:
/
/

d
我
r
e
C
t
.

米

我
t
.

e
d
你
哦
p
米

我
/

我

A
r
t
我
C
e
–
p
d

F
/

2
1
1
1
8
6
8
3
0
7
哦
p
米
_
A
_
0
0
0
1
3
p
d

我

乙
y
G
你
e
s
t

哦
n
0
7
S
e
p
e
米
乙
e
r
2
0
2
3

Partial Truths Kline, Schulz, 吉布森

the number of possible interpretations by any amount (Frank & 古德曼, 2012). 因此, 我们
can vary the informativity of these utterances by varying the possible events that might have
occurred in the local context, speciﬁcally by manipulating the set of possible agents and pa-
tients. We can then ask whether speakers choose informative utterances, even in cases where
a listener would be unable to identify the entire event meaning.

数字 1 shows a possible event (JOHN FEEDS THE DOG) and six sets of entities that
could participate in the event to be named. Each context set is made up of people (canonical
代理人) and either animals or inanimate objects (both of which are more likely than humans
to appear as patients). Critically, we manipulate the communicative context (and therefore the
informativity of potential utterances) by altering the set of seven entities that appear in the con-
text picture. If the context is Figure 1a, the utterance FEED DOG fails to resolve the ambiguity
(anyone could have done it); 另一方面, the utterance JOHN FEED speciﬁes the agent
and relies on an intelligent listener to identify the unique patient in context. For Figure 1f,
the reverse is true: FEED DOG resolves the ambiguity. In the intermediate cases (Figures 1b–
1e), there is no two-word utterance that can fully disambiguate the intended meaning: 那里
are multiple options for both agent and patient, and the verb cannot be uniquely inferred the
context images.

In these conditions, different words reduce ambiguity to different degrees:

Our critical hypothesis has to do with how people will behave in the four interme-
diate arrays.
在
图1e, mentioning John (and the verb) narrows down the possible events to just two alter-
natives (he feeds the dog or duck) rather than ﬁve (somebody feeds the dog). If the RSA model
extends to descriptions of argument structure relations, adults should still be able to select in-
formative utterances: When there are more agents than patients, participants should be more
likely to mention subjects, even if ambiguity between multiple messages remains. 然而,
if participants use a simpler strategy of determining just whether or not a given utterance suc-
cessfully conveys the intended event, then they should still choose informative arguments in
the deterministic cases, but perform at chance (or otherwise not differentiate the intermediate
状况) when both arguments remain ambiguous.

METHOD

参加者

Ninety-one English-speaking adults participated on Amazon’s Mechanical Turk (AMT). Partic-
ipants were screened to be located in the United States and self-reporting English as their ﬁrst
语言 (额外的 21 participants were excluded who did not meet these criteria). 不
other demographic information was collected. The task took approximately 13 minutes to
complete and participants were paid $1.00. This pay rate was based on an anticipated study
length of 10 minutes, following the 10¢/minute rule of thumb used for AMT studies in the lab
at the time these data were collected. All participants gave informed consent in accordance
with the requirements of the Massachusetts Institute of Technology’s institutional review board.

Stimuli

We created cartoon stimulus sets for each of 12 动词 (吃, 喂养, hold, drink, kick, drop,
wash, pour, 扔, 触碰, read, and roll). Each set consisted of an action picture and six
“context” pictures showing possible agents and patients who might participate in the event.
The people were generated using a character-creation website (Brooks et al., 2007) with dis-
tinct features and names on their shirts. The objects were chosen from a category (例如, 各种各样的

开放的心态: 认知科学的发现

我

D
哦
w
n
哦
A
d
e
d

F
r
哦
米
H

t
t

:
/
/

d
我
r
e
C
t
.

米

我
t
.

e
d
你
哦
p
米

我
/

我

A
r
t
我
C
e
–
p
d

F
/

2
1
1
1
8
6
8
3
0
7
哦
p
米
_
A
_
0
0
0
1
3
p
d

我

乙
y
G
你
e
s
t

哦
n
0
7
S
e
p
e
米
乙
e
r
2
0
2
3

Partial Truths Kline, Schulz, 吉布森

我

D
哦
w
n
哦
A
d
e
d

F
r
哦
米
H

t
t

:
/
/

d
我
r
e
C
t
.

米

我
t
.

e
d
你
哦
p
米

我
/

我

A
r
t
我
C
e
–
p
d

F
/

2
1
1
1
8
6
8
3
0
7
哦
p
米
_
A
_
0
0
0
1
3
p
d

我

乙
y
G
你
e
s
t

哦
n
0
7
S
e
p
e
米
乙
e
r
2
0
2
3

数字 1. Context and event images for JOHN FEEDS THE DOG. We refer to each context con-
dition by the number of people (代理人) 和物体 (患者) 展示, 例如, Figure 1a is
notated as [1:6], Figure 1b as [5:2], 等等.

foods) appropriate for each verb. The total number of agents and patients in each context
sums to 7, yielding six variations (IE。, [6:1] 到 [1:6]) for each of the 12 stimulus sets. 全部
刺激, 代码, and analyses are available in the Supplemental Materials for this article (Kline,
Schulz, & 吉布森, 2017).

开放的心态: 认知科学的发现

Partial Truths Kline, Schulz, 吉布森

程序

Stimuli were presented using Python and the EconWillow package (Weel, 2008), 访问过
through AMT. Participants were told that they were providing descriptions for another (sham)
参与者. Participants saw the trials in a random order, with two items presented at each
context type. On each trial, they saw the context picture for ten seconds, read a sentence
describing the action they would see (例如, “John feeds the dog”), and then saw the action
picture for ten seconds. 最后, the context picture reappeared and participants were given two
separate text boxes to enter their description; if they entered more than two words (screened
by checking for spaces, 例如, “baby rolls” in one box), they were told to try again. To encourage
participants to answer quickly, their response speed in seconds was shown after every trial.

Data Coding

A total of 1,092 responses were collected from the 91 参与者, 182 responses in each
状况. Responses were ﬁrst checked for minor variations such as capitalization and verb
形式 (例如, “Eaten” was coded as “eat”). The majority of these responses (84%) consisted of
two of the possible three content words in the sentence (例如, JOHN FEED, FEED DOG, 或者
JOHN DOG). In the remaining responses, participants deviated from these exact lexical items;
in these cases we checked if the word used could refer to a unique entity (例如, she in an array
with a female agent among only male distractors). A full record of this coding is available in the
补充材料 (Kline, Schulz, & 吉布森, 2017); 只是 20 responses (1.8%) consisted
of two unclassiﬁed words and thus were excluded from analyses. Because not every response
contained two codable words, we present analyses below for the presence of agents, 患者,
and verbs in each response.

结果

We code the main effect of interest numerically, representing the key condition of context type
in the model as the number of potential agents in the context image (recall that the number
of agents and patients in these context images are inversely related, always summing to seven
全部的). The effect of the number of agents vs. patients on whether participants mentioned the
agent in their response was highly signiﬁcant by a mixed-effect logistic regression2 with
random slopes and intercepts for both item and participant (β = 0.55, SE = 0.15, Z = 3.79,
p < .001; LRT: χ2 = 10.7, d f = 8, p < .005). The same was true for patients (β = −0.55, SE = 0.13, Z = −4.38, p < .001; LRT: χ2 = 17.2, d f = 8, p < .001). These patterns are as predicted—as more agent distractors (and thus fewer patient distractors) were present, partic- ipants were more likely to mention the agent and less likely to mention the patient (Figure 2). We also found that participants overall were somewhat more likely to mention patients than agents: On the subset of trials (74%) where participants mentioned only one of the two, there were signiﬁcantly more patients than agents (p < .001, binomial test). To test whether participants gave graded responses to the intermediate arrays (e.g., [2:5]), we also examined the effects of array type after removing trials for which a “deterministic” answer could be given ([6:1], [1:6]). The effects of array type on both agent and patient mention were both signiﬁcant when evaluating only these intermediate cases (agent mention: 2 In addition to reporting beta statistics, we evaluate these models with likelihood ratio tests (LRT) by com- parison with a model with the same random effects and only the effect of interest omitted from the ﬁxed effects structure. Exact model speciﬁcations can be found in the analysis ﬁle named MD_turk.R in the Supplemental Materials for this article (Kline, Schulz, & Gibson, 2017). OPEN MIND: Discoveries in Cognitive Science 6 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 o p m i / l a r t i c e - p d f / / / / / 2 1 1 1 8 6 8 3 0 7 o p m _ a _ 0 0 0 1 3 p d . i 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 Partial Truths Kline, Schulz, Gibson Figure 2. Trials on which participants included subjects, objects, and verbs. Error bars represent 95% bootstrapped conﬁdence intervals. β = 0.30, SE = 0.10, Z = 2.97, p < .005, LRT: χ2 = 5.89, d f = 8, p < .05; patient mention: β = −0.33, SE = 0.16, Z = −2.01, p < .05, LRT: χ2 = 3.88, d f = 8, p < .05). MODEL COMPARISONS To evaluate how human performance might reﬂect pragmatic choices, we compared three computational models (with two additional variations shown in the Supplemental Materials [Kline, Schulz, & Gibson, 2017]). Each of these models generates (unordered) two-word ut- terances (“AV”—agent and verb, “VP”—verb and patient, or “AP”—agent and patient ) at each of the conditions in the experiment; we compare model predictions to participants’ responses of these types (omitting the ∼15% of responses that included some other word). Below, we describe the common assumptions the models share, deﬁne the particulars of each model, and then compare them to human performance. In all models, the shared context is the set of possible events that might occur given the set of agents, patients, and plausible verbs (we assume the prior probability of picking any particular event e in E is uniform). We assume that each object/person in the scene is classiﬁed unambiguously as an agent or patient (wrong in general, but true in our experimental context). For the verbs, we assume participants are considering some set of possible interactions between the agents and patients (e.g., petting, feeding). In principle, the notion of “possible verb set” could be estimated empirically by asking naïve participants to list possible actions between the agent–patient stimulus sets directly. Here, we simply assume the set is relatively small and does not vary with the number of agents and patients.3 Thus, the shared context of possible events E is E = {(a 1, v 1, p 1 ) , (a 1, v 1, p 2) , (a 1, v 2, p 1 ) . . . }, (2) 3 We tested just two values for the number of verbs (k): 5 and 50; we use k = 5 in all models. The effect of increasing k is to increase the relative likelihood of including the verb in an utterance. OPEN MIND: Discoveries in Cognitive Science 7 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 o p m i / l a r t i c e - p d f / / / / / 2 1 1 1 8 6 8 3 0 7 o p m _ a _ 0 0 0 1 3 p d . i 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 Partial Truths Kline, Schulz, Gibson with the number of possible events e in E calculated as follows: A = {a 2, . . . , ai} 1, a 2, . . . , pj} P = {p 1, p V = {v 2, . . . , vk} 1, v |E| = |A| ∗ |P| ∗ |V| . (3) Next, we consider a set of possible descriptions (D) that could be used for some target event e. Each of these descriptions might also apply to other events in the possible set; the number of events some description d can refer to is notated as |d|. We assume that there is a single, unambiguous label available for each agent, patient, and verb. This means that for a single-word description, |d| can be deﬁned easily: For instance, the word referring to the agent can refer to any of the events in E that include that agent plus some patient and verb: (4) For these models, we consider the set of utterances D = [“AV,” “VP,” “AP”] (i.e., two words, produced in any order). A two-word description like “AV” can refer to any event in E that contains that agent, that verb, and then some patient: d = “A” → |d| = j ∗ k. d = “AV” → |d| = k. (5) All of the following models use these same assumptions about the relationship between a particular context (containing possible agents, patients, and verbs) and the number of events a particular utterance could refer to; they differ in how utterances are produced given this information. The outputs of the models are shown in Figure 3. Nonpragmatic “Cost only” model ( pNP) We begin with a baseline model that does not take any aspect of the context in which an utterance was produced into account. In our dataset, we found an overall difference in the frequency with which certain utterances are produced (in particular, “AV” sequences are less frequent than “VP” sequences), averaging across contexts. We can consider this as reﬂecting a differential cost to the speaker of producing each of these utterances. The corresponding model just produces utterances at this base rate ( phuman), with no effects of the context in which the utterance is produced. Because the human participants sometimes produced a word that did not refer to the agent, verb, or patient, we estimate these probabilities for the model by renormalizing over only the “standard” responses (~84% of all responses). The probability of producing each two-word description d for this model is simply this global likelihood: ⎧ ⎪⎨ ⎫ ⎪⎬ phuman (AV) if d = “AV” phuman (VP) if d = “VP” phuman (AP) if d = “AP” . ⎪⎭ (6) pNP (d | e, E) = pNP(d) = ⎪⎩ To avoid overﬁtting, we evaluate this model by randomly splitting the human data in half to calculate these parameters from the data, and evaluating each set of predictions against the other half of the human dataset. Pragmatic “succeed/fail” heuristic (pSF ) Many common-ground-type experiments (e.g., Sedivy, 2003) tacitly assume that an utterance is pragmatically helpful when it uniquely identiﬁes the target referent. For this model, we assume the verb is always mentioned because (unlike the possible agents and patients) there OPEN MIND: Discoveries in Cognitive Science 8 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 o p m i / l a r t i c e - p d f / / / / / 2 1 1 1 8 6 8 3 0 7 o p m _ a _ 0 0 0 1 3 p d . i 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 Partial Truths Kline, Schulz, Gibson 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 o p m i / l a r t i c e - p d f / / / / / 2 1 1 1 8 6 8 3 0 7 o p m _ a _ 0 0 0 1 3 p d . i 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 Figure 3. Predictions of the generative models for including Agent, Patient, and Verb in event descriptions. Human performance (reproduced from Figure 2) is shown at the bottom right for comparison. is no direct information about the verb present in the context image, and therefore this word will always be highly informative. We thus assume that the possible utterances are simply (unordered) “AV” or “VP,” with the probabilities of producing the two utterances summing to one. The probability of each utterance is given by: (cid:9) 1 i f ε i f ∑d(cid:5) ∈ D (cid:10) |d| = 1 |d| > 1
p(d(西德:5) |e, 乙) .

pSF (d | e, 乙) =

(7)

那是, out of the available descriptions, the model will consider only the ones that
succeed in uniquely identifying the target event. The ε symbol represents a small number
arbitrarily close to zero, and indicates that this model is extremely unlikely to choose the
uninformative utterance if any informative ones are available. As shown in Equations 8–9, 它是
exact value does not impact the quantitative predictions the model makes. If there is a single
informative choice (|d|= 1), the model will select it approximately deterministically:

开放的心态: 认知科学的发现

(西德:11)

pSF

din f ormative| e, 乙

(西德:12)

= 1

1 + ε

∼= 1.

(8)

Partial Truths Kline, Schulz, 吉布森

But if neither utterance is fully informative (那是, 如果 |d| > 1), the two utterances are

produced at chance:4

(西德:11)

(西德:12)

pSF

dunin f ormative|e, 乙

ε
ε + ε

∼= 1
2

(9)

Rational speaker ( pRS)
We implement Frank and Goodman’s RSA model, which states that a description d will be
chosen in inverse proportion to how many events that description can apply to.5 Thus the
probability for each of the three possible two-word descriptions is:

pRS (d | e, 乙) =

|d|−1
(西德:13)
(西德:13)d(西德:5)

∑d (西德:5) ∈D

(西德:13)
(西德:13)−1 .

(10)

Model comparison

To facilitate comparison with the human results (数字 3) we plot the probabilities that a word
for a particular element (A, V, 磷) is included in the utterances generated by each model. 这
“nonpragmatic” model that considers only base rate performs relatively poorly at matching
human performance, r (36) = .63 (this and all model comparison p values are < .0001; we randomly divided the human data into two halves to avoid overﬁtting to parameters estimated from the data). The succeed/fail model is somewhat better, r (36) = .75, and the rational speaker model better still, r (36) = .81. In the Supplemental Materials (Kline, Schulz, & Gibson, 2017), we compare versions of the latter two models that also incorporate information about the base rate of each words; again, the rational speaker model is a closer ﬁt to human performance than the equivalent succeed/fail model. DISCUSSION As predicted by the RSA, when participants described events after seeing arrays of possible agents and patients, their two-word answers reﬂected the degree to which a given word could convey new information about the event. Participants were more likely to mention the agent of the event when the agent was more ambiguous, and more likely to mention the patient when the patient was more ambiguous. This was not limited to cases where an event could be uniquely identiﬁed: Even for the intermediate cases where there were multiple agents and multiple patients in the array, participants still chose the two-word sequence that reduced uncertainty the most. Quantitative comparison to the RSA reveals a close ﬁt to human data, with a baseline-adjusted version of the RSA performing best. While understanding language appears to involve assuming that we are listening to ra- tional speakers, our own speech also involves messy, sometime under- or overinformative utterances. Nevertheless, we mainly succeed in getting our meanings across, and it is clear that at least some aspects of adult speech are well designed for robustly transferring infor- mation. While there is a rich literature on how speakers accommodate nonlinguistic context when describing individual objects (cf. Brennan & Clark, 1996; Brown-Schmidt & Tanenhaus, 2008; Nadig & Sedivy, 2002; Pogue et al., 2016; Sedivy, 2003), this study provides the ﬁrst evidence that adults’ language production is affected, at the level of argument structure, by 4 This is also the case if both utterances are informative, though in this experiment it is never the case that both “AV” and “VP” would uniquely identify the event. 5 The RSA model also includes a prior probability term (i.e., how often each event is expected to occur); here we assume that all the events have equal prior probability of occurring. OPEN MIND: Discoveries in Cognitive Science 10 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 o p m i / l a r t i c e - p d f / / / / / 2 1 1 1 8 6 8 3 0 7 o p m _ a _ 0 0 0 1 3 p d . i 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 Partial Truths Kline, Schulz, Gibson the graded informativity of possible utterances in context. Although the two kinds of short- ened sentences (Agent-Verb, e.g., GIRL READ, and Verb-Patient, e.g., READ BOOK) are on average equal in length and express the same amount of information, participants recognize that informativity depends on the set of possible alternative events. This holds even when either utterance will leave some ambiguity, suggesting that RSA-type listeners are correct: Their speaker partners are choosing what to say and what to omit in a way that can maximally reduce their uncertainty. Understanding how listeners and speakers represent contexts and possible messages for verbs and events is a puzzling problem. In noun-referent studies, participants (listeners or speakers) need simply note how many possible referents there are and what features differ be- tween them (e.g., Stiller, Goodman, & Frank, 2015). For sentence-level meanings, the set of possible messages is much larger than the number of visible referents. When there are three potential agents and four patients, there are 12 possible combinations, and there may often be multiple verbs under consideration. Beyond this, the listener might have to guess at likely events, as well as multiple ways of referring to that event: Beyond just relations between a girl and an apple, speakers and listeners must consider the many different propositions or perspec- tives that can be used to refer to the same event (e.g., a girl swinging a bat and hitting a ball toward the outﬁelder can describe the very same event; cf. Gleitman, 1990; Kline, Snedeker, & Schulz, 2017). These perspectives might differ in argument structure, so that a listener might need to consider multiple argument sets: an agent and patient, an agent, theme, and recipi- ent, and so on. Furthermore, in the real world many referents, especially humans, can play many roles (e.g., agent and patient of hugging), and some possible referent pairs will permit different interactions due to either selectional restrictions or real-world knowledge. We may be able to use the current paradigm to address features of argument structure communication like these: If a speaker learns that wugging can be performed by animals but not people, will he or she take this information into account when designing utterances for a partner who does or doesn’t know this restriction? How far do parallels between messages about object identity and propositions about the world (event descriptions) extend? Which of the complexities of sentence-level predictability do speakers and listeners fold into their models of communica- tive context? Understanding the dynamics of utterance production in these contexts will further our understanding of how adults calculate and use informativity to accomplish our commu- nicative goals. ACKNOWLEDGMENTS We would like to thank the members of the Schulz and Gibson labs for their helpful feedback; Audra Podany, Olivia Murton, and Dmetri Hayes for assistance in creating stimuli and data an- notation; and all of the participating AMT workers for their involvement in the study. This work was funded by grants from the National Science Foundation to Edward Gibson and Melissa Kline. AUTHOR CONTRIBUTIONS MK, LS, and EG conceived of and planned the experiment. MK implemented and carried out the experiment and performed all analyses with input from LS and EG. MK planned and carried out the computational modeling, with EG providing feedback on implementation and interpretation of the models. MK took the lead in writing the manuscript, and all authors provided critical feedback and input to the interpretation of the results and revision of the manuscript. OPEN MIND: Discoveries in Cognitive Science 11 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 o p m i / l a r t i c e - p d f / / / / / 2 1 1 1 8 6 8 3 0 7 o p m _ a _ 0 0 0 1 3 p d . i 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 Partial Truths Kline, Schulz, Gibson REFERENCES Altmann, G. T., & Kamide, Y. (1999). Incremental interpretation at verbs: Restricting the domain of subsequent reference. Cogni- tion, 71, 247–263. Aylett, M., & Turk, A. (2004). The smooth signal redundancy hypothesis: A functional explanation for relationships between redundancy, prosodic prominence, and duration in sponta- neous speech. Language and Speech, 47, 31–56. doi:10.1177/ 00238309040470010201 Brennan, S. E., & Clark, H. H. (1996). Conceptual pacts and lexi- cal choice in conversation. Journal of Experimental Psychology: Learning, Memory, and Cognition, 22(6), 1482–1493. Brennan, S. E., & Hanna, J. E. (2009). Partner-speciﬁc adaptation in dialog. Topics in Cognitive Science, 1, 274–291. doi:10.1111/j. 1756-8765.2009.01019.x Brooks, J., Groening, M., Jean, A., Scully, M., Sakal, R., & (2007). The Simpsons 20th Century Fox Home Entertainment. Movie [Website]. Retrieved from http://web.archive.org/web/ 20120317090828/http://www.simpsonsmovie.com Brown-Schmidt, S., & Tanenhaus, M. (2008). Real-time investiga- tion of referential domains in unscripted conversation: A targeted language game approach. Cognitive Science: A Multidisciplinary Journal, 32, 643–684. doi:10.1080/03640210802066816 Frank, M. C., & Goodman, N. D. (2012). Predicting pragmatic rea- soning in language games. Science, 336(6084), 998. doi:10. 1126/science.1218633 Galati, A., & Brennan, S. E. (2010). Attenuating information in spoken communication: For the speaker, or for the addressee? Journal of Memory and Language, 62, 35–51. doi:10.1016/j.jml. 2009.09.002 Ginzburg, J. (1996). Dynamics and the semantics of dialogue. Logic, Language and Computation, 1, 1–16. Gleitman, L. R. (1990). The structural sources of verb meanings. Language Acquisition, 1, 3–55. Goodman, N., & Frank, M. (2016). Pragmatic language interpre- tation as probabilistic inference. Trends in Cognitive Sciences, 20, 818–829. doi:10.1016/j.tics.2016.08.005 Goodman, N., & Stuhlmüller, A. (2013). Knowledge and implica- ture: Modeling language understanding as social cognition. Top- ics in Cognitive Science, 5, 173–184. doi:10.1111/tops.12007 Grice, H. (1975). Logic and conversation. Syntax and Semantics, 3, 41–58. Horton, W. S. (2005). Conversational common ground and mem- ory processes in language production. Discourse Processes, 40, 1–35. doi:10.1207/s15326950dp4001_1 Jaeger, T. F. (2010). Redundancy and reduction: Speakers manage syntactic information density. Cognitive Psychology, 61, 23–62. doi:10.1016/j.cogpsych.2010.02.002 Kao, J. T., Levy, R., & Goodman, N. D. (2013). The funny thing about incongruity: A computational model of humor in puns. Proceed- ings of the 35th Conference of the Cognitive Science Society, 35, 738–733. Kao, J. T., Wu, J. Y., Bergen, L., & Goodman, N. (2014). Nonliteral understanding of number words. Proceedings of the National Academy of Sciences, 111, 12002–12007. doi:10.1073/pnas. 1407479111 Keysar, B., Barr, D. J., & Horton, W. S. (1998). The egocentric basis of language use: Insights from a processing approach. Current Directions in Psychological Science, 7(2), 46–50. doi:10.1111/ 1467-8721.ep13175613 Kline, M., Schulz, L., & Gibson, E. (2017). Supplemental material for “Partial truths: Adults choose to mention agents and patients in proportion to informativity, even if it doesn’t fully disambiguate the message.” Open Mind: Discoveries in Cognitive Science, 1(2). doi:10.17605/OSF.IO/AWGH6 Kline, M., Snedeker, J., & Schulz, L. (2017). Linking language and events: Spatiotemporal cues drive children’s expectations about the meanings of novel transitive verbs. Language Learning and Development, 13, 1–23. doi:10.1080/15475441.2016.1171771 (2016). What do we mean by Language, Cognition prediction in language comprehension? and Neuroscience, 31, 32–59. doi:10.1080/23273798.2015. 1102299 Kuperberg, G. R., & Jaeger, T. F. Kurumada, C., & Jaeger, T. F. (2015). Communicative efﬁciency in language production: Optional case-marking in Japanese. Journal of Memory and Language, 83, 152–178. doi:10.1016/ j.jml.2015.03.003 Levy, R. (2008). Expectation-based syntactic comprehension. Cog- nition, 106, 1126–1177. doi:10.1016/j.cognition.2007.05.006 Lockridge, C. B., & Brennan, S. E. (2002). Addressees’ needs inﬂu- ence speakers’ early syntactic choices. Psychonomic Bulletin & Review, 9, 550–557. doi:10.3758/BF03196312 MacDonald, M. C. (2013). How language production shapes lan- Frontiers in Psychology, 4, guage form and comprehension. 226–253. doi:10.3389/fpsyg.2013.00226 MacDonald, M. C., Pearlmutter, N. J., & Seidenberg, M. S. (1994). Lexical nature of syntactic ambiguity resolution. Psychological Review, 101, 676–703. Mahowald, K., Fedorenko, E., Piantadosi, S. T., & Gibson, E. (2013). Info/information theory: Speakers choose shorter words in predictive contexts. Cognition, 126, 313–318. doi:10.1016/j. cognition.2012.09.010 Nadig, A. S., & Sedivy, J. C. (2002). Evidence of perspective-taking constraints in children’s on-line reference resolution. Psycholog- ical Science, 13(4), 329–336. Pickering, M. J., & Garrod, S. (2004). Toward a mechanistic psychol- ogy of dialogue. Behavioral and Brain Sciences, 27, 169–190. doi:10.1017/S0140525X04000056 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 o p m i / l a r t i c e - p d f / / / / / 2 1 1 1 8 6 8 3 0 7 o p m _ a _ 0 0 0 1 3 p d . i 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 Pogue, A., Kurumada, C., & Tanenhaus, M. K. (2016). Talker-speciﬁc generalization of pragmatic inferences based on under- and over- informative prenominal adjective use. Frontiers in Psychology, 6, 2035–2036. doi:10.3389/fpsyg.2015.02035 (1960). Speaking of objects. Quine, W. V. O. In W. V. O. Quine (pp. 1–16). (Ed.), Ontological New York, NY: Columbia University Press. relativity and other essays Resnik, P. (1996). Selectional constraints: An information-theoretic model and its computational realization. Cognition, 61, 127– 159. Sedivy, J. (2003). Pragmatic versus form-based accounts of referen- tial contrast: Evidence for effects of informativity expectations. Journal of Psycholinguistic Research, 32, 3–23. OPEN MIND: Discoveries in Cognitive Science 12 Partial Truths Kline, Schulz, Gibson Shannon, C. E. (1949). Communication in the presence of noise. Proceedings of the IRE, 37, 10–21. doi:10.1109/JRPROC.1949. 232969 Stiller, A. J., Goodman, N., & Frank, M. C. (2015). Ad-hoc implica- ture in preschool children. Language Learning and Development, 11, 176–190. doi:10.1080/15475441.2014.927328 Tanenhaus, M. K., Spivey-Knowlton, M. J., Eberhard, K. M., & Sedivy, J. (1995). Integration of visual and linguistic information in spoken language comprehension. Science, 268, 1632–1634. van Son, R. J. J. H., & van Santen, J. P. H. (2005). Duration and spectral balance of intervocalic consonants: A case for efﬁcient communication. Speech Communication, 47(1–2), 100–123. doi:10.1016/j.specom.2005.06.005 Weel, J. (2008). Willow, a Python library for economics. San Francisco, CA: Palantir Technologies. Retrieved from http:// econwillow.sourceforge.net/ 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 o p m i / l a r t i c e - p d f / / / / / 2 1 1 1 8 6 8 3 0 7 o p m _ a _ 0 0 0 1 3 p d . i 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 OPEN MIND: Discoveries in Cognitive Science 13 Partial Truths: Adults Choose to Mention Agents image

下载pdf