Keywords 关键词

The social world requires people to make highly consequential predictions. People need to predict whether new acquaintances will become a friend or foe (Cuddy, Fiske, & Glick, 2008), how old friends will respond to constructive criticism, or how long the boss will be angry before approaching her for a favor. Whether in cooperation or competition, successful social interaction requires people to anticipate others’ future thoughts, feelings, and actions, and prepare their own actions accordingly. Social predictions are among the most common predictions a person must make because people spend so much time with other people (United States Bureau of Labor Statistics, 2003). Yet despite the importance of social prediction, researchers have only just scratched the surface of the predictive social mind. Here we consider recent research that is starting to reveal how people glimpse the social future.
社会世界要求人们做出非常重要的预测。人们需要预测新结识的人是会成为朋友还是敌人(Cuddy, Fiske, & Glick, 2008),老朋友会如何对待建设性的批评,或者老板生气多久才会在请求帮助之前平复情绪。无论是合作还是竞争,成功的社交互动都需要人们预测他人未来的想法、感受和行为,并相应地准备自己的行动。社会预测是人们必须做出的最常见的预测之一,因为人们花费了大量时间与他人相处(美国劳工统计局,2003 年)。然而,尽管社会预测的重要性,研究人员只是刚刚触及了预测社交心智的表面。在这里,我们考虑最近的研究,它开始揭示人们如何窥视社会未来。

Research on nonsocial prediction suggests that the brain is built to make predictions. It does not passively perceive the world around it and then react accordingly. Instead, people make reflexive predictions across multiple domains (Hohwy, Roepstorff, & Friston, 2008; Rao & Ballard, 1999; Vuust, Ostergaard, Pallesen, Bailey, & Roepstorff, 2009). When processing language, for example, people use the beginning of a sentence to predict the end of that sandwich. When watching a ball thrown into the air, people reflexively make a prediction about its eventual downward trajectory. However, the social world poses unique challenges to people’s well-honed predictive capacities. Humans are not billiard balls: people are probabilistic beings, moved to action by the unseen forces of thoughts and feelings. How do people represent these invisible mental states, and use them to make predictions?
研究非社交预测表明,大脑是建立在预测基础上的。它并不是被动地感知周围的世界,然后做出相应的反应。相反,人们在多个领域进行反射性预测(Hohwy, Roepstorff, & Friston, 2008; Rao & Ballard, 1999; Vuust, Ostergaard, Pallesen, Bailey, & Roepstorff, 2009)。例如,在处理语言时,人们利用句子的开头来预测句子的结尾。当观看一个球被扔到空中时,人们会本能地预测它最终的下行轨迹。然而,社交世界对人们熟练的预测能力提出了独特的挑战。人类不是台球:人是概率性的存在,受到思想和感情这些看不见的力量的驱使而行动。人们如何代表这些看不见的心理状态,并利用它们进行预测?

A recent theoretical framework for social cognition (Tamir & Thornton, 2018) proposes a simple answer to this question (Fig. 1). This multilayered framework of social cognition helps to explain how people predict others’ future states and behaviors in two steps: First, it suggests that the mind organizes social knowledge using conceptual “maps” of social stimuli. These maps allow people to easily track other people’s current thoughts, feeling, and actions. Second, it suggests that people track distances and trajectories through these maps to make efficient, automatic social predictions. This framework advances prediction as the central goal of representing social knowledge.
最近的社会认知理论框架(Tamir&Thornton,2018)对这个问题提出了一个简单的答案(图 1)。这个多层次的社会认知框架有助于解释人们如何通过两个步骤预测他人未来的状态和行为:首先,它表明大脑使用社会刺激的概念“地图”来组织社会知识。这些地图使人们能够轻松追踪其他人当前的想法、感受和行为。其次,它表明人们通过这些地图追踪距离和轨迹,以便进行高效、自动的社会预测。这个框架将预测作为代表社会知识的中心目标。

Fig. 1 图 1
figure 1

The multilayer model of predictive social cognition. Three layers of social knowledge—observable actions and hidden mental states and traits—are each organized into low-dimensional maps by psychological dimensions such as valence and power. Transitions between or within layers (arrows) decrease in probability with distance. Short hops between adjacent points (e.g., happiness and gratitude) are more likely than long treks between distant points (e.g., sleeping and hiking). This organization of social knowledge provides both for parsimonious representation of the complexities of the social world, and for accurate, automatic prediction of the social future. (Reproduced with permission from Trends in Cognitive Science)

In this chapter, we first describe how people simplify the complexity of the social world using these low-dimensional maps. Next, we discuss how people leverage these maps to make accurate, automatic social predictions. Finally, we offer suggestions for how future research can use this framework productively to model real-world social predictions, and constructively to enhance its explanatory power and scope.

The Organization of Social Knowledge

Humans enjoy extraordinarily rich social lives. Most people know hundreds or even thousands of unique individuals. In a given moment, each one of these individuals may be thinking widely differing thoughts, experiencing different emotions, or performing different actions. This richness makes life interesting and exciting, but it also poses a significant challenge: people must understand others’ traits, mental states, and action in order to successfully interact with them. How can the social mind come to grips with the complexity of social life?

The right organizational scheme can bring order to chaos. What kind of scheme does the human mind use to organize social information? We propose that people use a social map. Or rather, that people employ multiple maps, one for each type of social information. Maps organize information by localizing it to particular coordinates on a small set of continuous dimensions. In a geographic map of the United States, the physical location of any city can be described in terms of its dimensions of longitude and latitude. Simply knowing the north-south and east-west coordinates of a city allows you to extract important information about its location from single pair of numbers. For instance, you might learn that a city is located in the Pacific Northwest, and thus that it is nearby to Seattle, WA, but far away from Miami, FL.

Conceptual maps of the social world act in a similar way, by reducing the complexity of social stimuli down to a few essential values. Social neuroscience research has begun to chart the maps that the brain uses to make sense of the social world. This work has revealed the cardinal dimensions that describe three key types of social information: actions, mental states, and traits. These three layers of social information form the core of what one might want to know about a person: what are they doing, how are they feeling, and what kind of person are they? Each layer captures the world at a different timescale. Actions occur at the shortest timescales, from less than a second up to a few hours, depending on their complexity. States might unfold over just a few minutes or persist for several days, depending on if they are more like emotions or moods. Traits are more lasting, or even permanent, ways to describe individual differences among people. To understand and predict other people at any time scale requires us to map out the content of each of these layers of social knowledge.

Mapping the Action Layer 映射行动层

Humans have an incredibly diverse behavioral repertoire. People are capable of engaging in thousands of different actions and activities, ranging from simple motor actions such as reaching and grasping, to complex extended activities such as conducting research or governing a nation. A successful social agent must have a keen understanding of these actions. Recent research (Thornton & Tamir, 2019b) has identified six psychological dimensions (Table 1) that scaffold people’s action concepts: the Abstraction, Creation, Tradition, Food, Animacy, and Spiritualism Taxonomy or ACT-FAST. These dimensions originated from a data-driven principal component analysis of verb use in large text corpora and were validated against behavioral judgments.
人类拥有极其多样化的行为表现。人们能够进行成千上万种不同的行动和活动,从简单的运动动作,比如伸手和抓取,到复杂的延续活动,比如进行研究或治理一个国家。一个成功的社会行为者必须对这些行动有深刻的理解。最近的研究(Thornton & Tamir, 2019b)确定了六个心理维度(表 1),这些维度构成了人们的行动概念:抽象、创造、传统、食物、生命力和灵性主义分类,即 ACT-FAST。这些维度源自对大型文本语料库中动词使用的数据驱动主成分分析,并且已经通过行为判断进行了验证。

Table 1 Dimensions of the FAACTS action taxonomy
表 1 FAACTS 行动分类法的维度

Like the other social maps we will explore, each dimension of the ACT-FAST carries intrinsic meaning. For example, if one knows that an action is high on the animacy dimension, then one knows that it is an action that tends to be performed by living agents such as humans and animals, rather than an act of nature or a machine. In this way, this conceptual map of the social world both offers the functionality of a physical map—representing the distances between difference locations—and also implies rich knowledge about each of those locations.
与我们将要探讨的其他社会地图一样,ACT-FAST 的每个维度都具有固有的含义。例如,如果一个人知道某个行为在生命活动维度上很高,那么他就知道这是一种往往由生物体,如人类和动物执行的行为,而不是自然或机器的行为。通过这种方式,社会世界的这个概念地图既提供了物理地图的功能——代表了不同位置之间的距离——又暗示了关于每个位置的丰富知识。

Together, the six ACT-FAST dimensions explain much of how people think about actions. Knowing an action’s coordinates on these dimensions can robustly predict: (1) who does an action, in terms of traits, (2) why one does an action, in terms of approach and avoidance motivations, (3) when one does an action, in terms of time of day, (4) where one does an action, in terms of outdoor versus indoor, and if indoor, public versus private, and (5) how one performs an action, in terms of body parts involved and mental or physical effort required. ACT-FAST can also explain patterns of natural language use, such as which verbs tend to co-occur with which nouns. In addition to answering these important psychological and linguistic questions, ACT-FAST can explain brain activity across a wide set of cortical regions implicated in action representation. Together, these findings suggest that ACT-FAST provide a useful, and biologically plausible map of how people organize knowledge about other people’s actions.
六个 ACT-FAST 维度共同解释了人们如何思考行为。了解行为在这些维度上的坐标可以强有力地预测:(1)谁执行行为,即特质方面,(2)为什么执行行为,即接近和回避动机方面,(3)何时执行行为,即一天中的时间,(4)在哪里执行行为,即室内还是室外,如果是室内,则是公共的还是私人的,以及(5)如何执行行为,即涉及的身体部位和所需的心理或身体努力。ACT-FAST 还可以解释自然语言使用的模式,比如哪些动词倾向于与哪些名词同时出现。除了回答这些重要的心理和语言学问题,ACT-FAST 还可以解释涉及行为表征的一系列大脑皮层区域的脑活动。这些发现共同表明,ACT-FAST 提供了一个有用且生物学上合理的人们如何组织关于他人行为的知识的地图。

Mapping the State Layer 映射国家层级

In addition to paying attention to others’ observable actions, successful social agents also attend to the hidden drivers of actions: mental states. These states must be inferred from indirect cues, such as facial expression and tone of voice, but once known, others’ thoughts and feelings can serve as powerful predictors of their behavior. People often share the same intuitions about the predictive power of mental states: angry people aggress, tired people rest, and happy people celebrate. Failing to attend to others’ mental states could lead to embarrassing faux pas at best, or to serious danger at worst. To avoid such pitfalls, we need a map of the state layer.

A number of established theories propose dimensions that organize mental states. For example, the circumplex model of affect offers a map with two dimensions, valence and arousal (Russell, 1980); or the distinction between emotional and rational states, prominent in many modern dual-process theories (Evans & Stanovich, 2013; Kahneman, 2003). Recent neuroimaging work (Tamir, Thornton, Contreras, & Mitchell, 2016) used PCA to synthesize the 16 dimensions that comprised these existing theories, and then validated the resulting components in terms of their ability to predict neural representations of mental states. This work identified three dimensions that structure the brain’s map of others’ mental states (Table 2). The first dimension, valence, captures whether others are feeling good or bad. Knowing the valence of a person’s mental state could help people avoid harm from those in negative states like rage, and enjoy pleasant, constructive social interactions with those in more positive moods. The second dimension on this map, social impact, captures which mental states would dispose others to engage in social interactions. Highly impactful states, whether good or bad, are more likely to affect one’s life. The final dimension, rationality, captures whether others are likely to act in a calm, deliberate, well-thought-out way, or react instinctively or rashly.
一些已建立的理论提出了组织心理状态的维度。例如,情感的圆环模型提供了一个具有两个维度的地图,即价值和唤醒(Russell,1980);或者在许多现代双过程理论中突出的情绪和理性状态之间的区别(Evans&Stanovich,2013;Kahneman,2003)。最近的神经影像工作(Tamir,Thornton,Contreras 和 Mitchell,2016)使用主成分分析来综合构成这些现有理论的 16 个维度,然后验证了所得到的组分,以确定它们对预测心理状态的神经表征的能力。这项工作确定了构成他人心理状态的大脑地图的三个维度(表 2)。第一个维度,即价值,捕捉了他人是感觉良好还是感觉不好。了解一个人的心理状态的价值可能有助于人们避免受到那些处于愤怒等负面状态的人的伤害,并与那些处于更积极情绪的人享受愉快、建设性的社交互动。这张地图上的第二个维度,社会影响,捕捉了哪些心理状态会使他人参与社交互动。 高度影响力的状态,无论是好是坏,更有可能影响一个人的生活。最后一个维度,理性,捕捉到其他人是否倾向于以冷静、深思熟虑的方式行事,还是本能地或冲动地反应。

Table 2 Dimensions of the mental state representation
表 2 精神状态表征的维度

Valence, social impact, and rationality together comprise the 3d Mind Model. This model can explain over 80% of the variance in neural representations of mental states. That is, it provides a near-complete map of the mental state layer (Thornton & Tamir, 2020a). Moreover, this map remains robust across different ways of perceiving mental states: similar dimensions emerge across modalities, regardless of whether people reflect on mental state-related scenarios presented as images or as text (Weaverdyck, Thornton, & Tamir, 2020). People also apply similar maps to thinking about their own minds, as opposed to the minds of others. While the structure of the map doesn’t change across targets, the resolution does: when people think about their own mental states, they pore over a highly detailed, richly annotated map (Thornton, Weaverdyck, Mildner, & Tamir, 2019). In contrast, when people think about the states of others, the resolution is much less fine-grained. This difference likely results from both the quality and quantity of information one has about their own minds, in contrast to the minds of others.
价值、社会影响和理性共同构成了三维心智模型。该模型可以解释超过 80%的神经表征心理状态的差异。换句话说,它提供了心理状态层的近乎完整的地图(Thornton & Tamir, 2020a)。此外,这一地图在感知心理状态的不同方式下仍然稳健:无论人们是通过图像还是文本来思考与心理状态相关的情景,类似的维度都会出现(Weaverdyck, Thornton, & Tamir, 2020)。人们在思考自己的心智时也会应用类似的地图,而不是他人的心智。虽然地图的结构在不同目标下并不会改变,但分辨率会有所不同:当人们思考自己的心理状态时,他们会仔细研究一个高度详细、丰富注释的地图(Thornton, Weaverdyck, Mildner, & Tamir, 2019)。相比之下,当人们思考他人的状态时,分辨率就会低得多。这种差异很可能是由于一个人对自己的心智和对他人的心智所拥有的信息的质量和数量不同造成的。

Mapping the Trait Layer 映射特质层

Although people may apply similar maps to the mental states of different people, individuals do differ in socially important ways. Enduring individual differences between people are known as traits. Compared with actions and mental states, traits are relatively permanent fixtures of an individual, changing slowly across a lifetime, if at all (Roberts & Mroczek, 2008; Srivastava, John, Gosling, & Potter, 2003). Knowing where a person places on trait dimensions can help people to make predictions about their likely states or actions. For example, if you know that someone is highly trustworthy, you can predict that they will not steal from you; if you know someone is highly social, you can predict that they might feel excited at a party. Traits thus help people to make nuanced predictions about people, across situations.
尽管人们可能将类似的地图应用于不同人的心理状态,但个体在社会上的重要方面确实存在差异。人与人之间的持久个体差异被称为特质。与行为和心理状态相比,特质是个体相对永久的特征,如果有变化,也是在整个一生中变化缓慢(Roberts & Mroczek, 2008; Srivastava, John, Gosling, & Potter, 2003)。了解一个人在特质维度上的位置可以帮助人们对其可能的状态或行为做出预测。例如,如果你知道某人非常值得信赖,你可以预测他们不会从你那里偷东西;如果你知道某人非常善于社交,你可以预测他们在派对上可能会感到兴奋。因此,特质有助于人们在不同情境下对人们进行细致的预测。

There are multiple existing dimensional maps of traits, including the Five Factor model (Openness, Conscientiousness, Extraversion, Agreeableness, and Neuroticism) of personality (Goldberg, 1990; McCrae & Costa, 1987), and the stereotype content model (Fiske, Cuddy, Glick, & Xu, 2002) consisting of warmth and competence. Recent neuroimaging research synthesized several of the most prominent trait theories in the literature (Thornton & Mitchell, 2018). In this work, a three-dimensional model—consisting of power, valence, and sociality—provided the best explanation for patterns of brain activity elicited by thinking about a large set of public figures. Knowing where a person resides on these dimensions can inform social judgments about them. Power indicates whether another person is dominant and competent, and thus, capable of enacting their will. Valence indicates whether that person is warm and trustworthy, and thus, likely to help or harm. Finally, sociality reflects whether a person is extraverted, and thus, likely to engage in the first place. Together, these dimensions provide a near-complete map of the trait space: Power, valence, and sociality explain more than two-thirds of reliable neural activity associated with making inferences about other people. This model outperforms even the Big 5 personality traits in explaining neural representations of other people. However, it is important to note that the Big 5 are ostensibly a model of the reality of traits, whereas the three-dimensional model instead aims to explain the perception of traits. Moreover, it is an open question whether people continue to apply any map of trait space for personally familiar others (Thornton & Mitchell, 2017). When people interact regularly, they may instead draw on the character of their relationship itself.
存在多个现有的特质维度图,包括五因素模型(人格的开放性、责任心、外向性、宜人性和神经质)(Goldberg, 1990; McCrae & Costa, 1987),以及由温暖和能力组成的刻板印象内容模型(Fiske, Cuddy, Glick, & Xu, 2002)。最近的神经影像研究综合了文献中最突出的几种特质理论(Thornton & Mitchell, 2018)。在这项工作中,一个三维模型——由权力、价值和社交性组成——为大量公众人物思考所引发的大脑活动模式提供了最佳解释。了解一个人在这些维度上的位置可以为社会对他们的判断提供信息。权力指示另一个人是否占主导地位和能力,因此,能够实施他们的意愿。价值指示那个人是否温暖和值得信赖,因此,可能会帮助或伤害。最后,社交性反映一个人是否外向,因此,可能会首先参与。 这些维度共同构成了特质空间的几乎完整地图:权力、价值和社交性解释了与推断他人相关的可靠神经活动的三分之二以上。这个模型甚至在解释对他人的神经表征方面胜过了大五人格特质。然而,重要的是要注意,大五人格特质表面上是特质现实的模型,而三维模型的目标是解释对特质的感知。此外,一个悬而未决的问题是人们是否会继续将特质空间的任何地图应用于熟悉的其他人(Thornton & Mitchell, 2017)。当人们经常互动时,他们可能会转而依赖于他们关系本身的特点。

Overlapping Social Maps 重叠的社会地图

The maps our mind makes of the social world rely on a similar set of brain regions. Both mental state and trait representation engage regions such as the medial prefrontal and parietal cortices, and the superior temporal sulcus extending from the temporoparietal junction forward to the anterior temporal lobe (Tamir et al., 2016; Thornton & Mitchell, 2018). These regions form the social brain network, a set of brain regions reliably activated by a wide range of social stimuli (Mitchell, 2008; Van Overwalle & Baetens, 2009). However, mental states and traits share more than just a gross anatomical similarity—they share a common neural code. Both maps include a dimension reflecting valence; both include a dimension reflecting sociality; and both include a dimension reflecting competence, dominance, and agency. This allows each dimension to be decoded from brain activity across domains (Thornton & Mitchell, 2018). This shared code hints at a deep connection between the way people think about others’ momentary and enduring mental properties—potentially bridging the traditional divide between traits and states.
我们心智对社会世界的认知地图依赖于一组类似的大脑区域。无论是心理状态还是特质表征都涉及到类似的大脑区域,比如额叶中央区和顶叶皮层,以及从颞顶联合向前延伸至前颞叶的上颞沟(Tamir 等,2016 年;Thornton & Mitchell,2018 年)。这些区域构成了社会大脑网络,一组被各种社会刺激可靠激活的大脑区域(Mitchell,2008 年;Van Overwalle & Baetens,2009 年)。然而,心理状态和特质不仅仅共享粗略的解剖相似性,它们还共享一个共同的神经编码。两种地图都包括反映情感价值的维度;都包括反映社交性的维度;以及都包括反映能力、支配力和代理性的维度。这使得每个维度都可以从跨领域的大脑活动中解码出来(Thornton & Mitchell,2018 年)。这种共享编码暗示着人们思考他人瞬时和持久心理特性的方式之间存在着深刻的联系,可能弥合了特质和状态之间的传统分歧。

Indeed, in recent research, we found that when our brain represents a person, it seems to do so by keeping track of the mental states that person habitually experiences (Thornton, Weaverdyck, & Tamir, 2019a). For example, if a politician is habitually in bad moods—grouchy, short-tempered, stubborn—one may form the impression that the politician has a negative disposition, at the trait level. Correspondingly, our results indicate that the pattern elicited by thinking about such a person could be reconstructed by adding together generic representations of grouchiness, short-temper, and stubbornness. This finding suggests a simple mechanism for impression formation—counting perceived mental states. Moreover, this process could later be reversed to make predictions about states based on people’s dispositions. Thus, although traits and states are typically thought of as separate, the trait and state dimensions described above reflect parallel concepts.
事实上,在最近的研究中,我们发现当我们的大脑代表一个人时,似乎是通过跟踪这个人习惯性经历的心理状态来做到这一点(Thornton, Weaverdyck, & Tamir, 2019a)。例如,如果一个政治家习惯性情绪低落——脾气暴躁、易怒、固执——人们可能会形成政治家性格消极的印象。相应地,我们的研究结果表明,思考这样一个人所引发的模式可以通过将易怒、脾气暴躁和固执的一般性代表相加来重建。这一发现表明了一个简单的印象形成机制——计算感知的心理状态。此外,这个过程后来可以被逆转,以基于人们的性格特征来预测状态。因此,尽管特质和状态通常被认为是分开的,上述的特质和状态维度反映了平行的概念。

However, not all socially relevant information occupies the same neural territory. In recent research, we and others have found that actions are represented in quite different portions of the brain than those involved in theory of mind (Tarhan & Konkle, 2019; Thornton & Tamir, 2020b). High-level visual regions within the dorsal and ventral paths—areas rarely implicated in social cognition per se—appear to play an important role in representing what others are doing. Understanding more about how maps of the social world relate—both in conceptual space and the physical territory of the brain—is a high priority for future research.
然而,并非所有社会相关信息都占据相同的神经领域。在最近的研究中,我们和其他人发现,行为在大脑中的代表与涉及心灵理论的部分大脑区域完全不同(Tarhan & Konkle, 2019; Thornton & Tamir, 2020b)。在背侧和腹侧路径内的高级视觉区域——这些区域很少涉及社会认知本身——似乎在代表他人的行为中发挥重要作用。了解社会世界地图在概念空间和大脑的物理领域中的关系,是未来研究的重要优先事项。

The Predictive Social Brain

Conceptual maps of actions, mental states, and traits allow people to organize their knowledge of the social world. However, the social world is not static; it is constantly in flux. At one moment, a person may be having fun dancing with their friends, but after a few hours of this energetic state, they may find themselves feeling exhausted. This mental state of exhaustion may in turn lead to a new activity, like resting, and so forth. Navigating the social world requires people to anticipate such transitions. Fortunately, the dimensional maps of social knowledge described above offer insight into these dynamics: knowing “where” a person is on a social map can tell you a great deal about where that person will “go” in the future.

Social Dimensions Scaffold Social Predictions

How can the static maps described above be used to predict the social future? The key assumption that licenses such predictions is that the proximity between points on the map reflects the likelihood of moving between those points. This assumption holds for geographic maps: on average, people tend to travel locally—e.g., to the store, work, school, or gym—with relatively high frequency. People travel to more distant locales—another city, country, or continent—much more rarely. Consequently, if you know another person’s current location on a map, you can accurately predict that their next destination is going to be close by.

We propose that people apply an analogous algorithm to make social predictions using conceptual maps. The coordinates on these maps represent actions, mental states, and personality traits rather than physical locations, but the logic of distance-based prediction still applies. Simply put, people are more likely to transition from one location within a layer to a nearby location than they are to transition to faraway locations. For example, if a person is currently feeling pride, a positive emotion, one might guess that she is much more likely to feel happy next than to feel sad. In this way, one could predict the likely trajectory of a person through a layer simply by knowing that person’s current coordinates. If so, then the multilayer structure mapped above would provide the foundation for understanding the dynamics of the social world.

Behavioral research suggests that people do predict the likelihood of transitions between states based on the proximity between those states on the dimensions of the 3d Mind Model—valence, social impact, and rationality. To demonstrate this, we first elicited people’s intuitions for how states transition from one to the next (Thornton & Tamir, 2017). For example, a participant might be told that a person is currently feeling “excitement” and be asked to rate the likelihood that that person will next experience “sleepiness” from 0% to 100%. Each state was mapped according to its location on the dimensions of rationality, social impact, and valence so that we could calculate the “distance” between each pair of states. Across two studies, proximity on the dimensions was positively associated with participants’ transitional probability ratings. The closer states were to each other on any of these mental state dimensions, the more likely participants judged the transitions between them (Fig. 2).
行为研究表明,人们根据 3D 心智模型的维度——价值、社会影响和理性之间的接近程度来预测状态之间的转变可能性。为了证明这一点,我们首先引发人们对状态如何从一个转变到另一个的直觉(Thornton & Tamir, 2017)。例如,参与者可能会被告知一个人目前感到“兴奋”,然后被要求评价这个人下一次经历“困倦”的可能性,评分从 0%到 100%不等。每个状态都根据其在理性、社会影响和价值维度上的位置进行了映射,以便我们可以计算每对状态之间的“距离”。在两项研究中,维度上的接近程度与参与者的转换概率评分呈正相关。在任何这些心理状态维度上,状态之间越接近,参与者就越可能判断它们之间的转变(图 2)。

Fig. 2 图 2
figure 2

Dimensional proximity predicts transitional probabilities between states. Each point on the scatter plots represents a transition from one mental state to another. The x-axis indicates the absolute distance between those states on each mental state dimension. The y-axis indicates the predicted transitional probability from one state to the other. The further two states are on each dimension, the less people expect a transition from one to the next. People are also less likely to actually transition between distant states. These state dimensions explain much of the accuracy of predicted transitional probabilities (Thornton & Tamir, 2017)
维度接近程度预测了状态之间的过渡概率。散点图上的每个点代表了从一个心理状态到另一个状态的过渡。横轴表示在每个心理状态维度上这些状态之间的绝对距离。纵轴表示从一个状态到另一个状态的预测过渡概率。在每个维度上,两个状态之间的距离越远,人们就越不太可能预期从一个状态过渡到另一个状态。人们也更不太可能实际上在远距离状态之间过渡。这些状态维度很大程度上解释了预测过渡概率的准确性(Thornton & Tamir, 2017)。

In the same way, people use conceptual proximity on the ACT-FAST dimensions to predict others’ actions (Thornton & Tamir, 2019a). Across five preregistered behavioral experiments, participants rated the likelihood that a person currently engaged in one action to next engage in another. For instance, how likely would it be for someone currently “dancing” to next “rest”? Proximities between actions on the ACT-FAST action dimensions reliably predict people’s transitional probability ratings. As with states, this result suggests that people may draw upon their map of the action layer to make predictions about others’ likely future actions, based on their current actions.
以同样的方式,人们利用 ACT-FAST 维度上的概念接近性来预测他人的行为(Thornton & Tamir, 2019a)。在五项预先注册的行为实验中,参与者评定了一个人目前从事一项行动后接下来从事另一项行动的可能性。例如,某人目前“跳舞”,下一步“休息”的可能性有多大?在 ACT-FAST 行动维度上的行动之间的接近性可靠地预测了人们的过渡概率评分。与状态一样,这一结果表明人们可能利用他们对行动层的地图来预测他人当前行动基础上可能的未来行动。

Social Dimensions Describe Real Social Dynamics

In the previous section, we described evidence that people use proximity within a social map to make predictions about others’ mental states and actions. However, it would only make sense to use social maps to make predictions if the social dimensions describe actual social dynamics. That is, one should use proximity on the dimension of valence to predict mental state transitions if and only if valence in fact describes regularities in the mental state transitions that others actually experience. As part of the investigations described in the previous section, we used experience sampling data and other real-world data to measure actual state and action transitions. This allowed for an estimation of the actual transitional probabilities between pairs of mental states (Thornton & Tamir, 2017) and pairs of actions (Thornton & Tamir, 2019a).
在前一节中,我们描述了证据表明人们利用社会地图中的接近性来预测他人的心理状态和行为。然而,只有在社会维度描述实际社会动态时,才有意义利用社会地图进行预测。也就是说,只有当 valence 实际上描述了他人实际经历的心理状态转变的规律性时,才应该利用 valence 维度来预测心理状态的转变。作为前一节中描述的调查的一部分,我们使用体验抽样数据和其他真实世界数据来测量实际状态和行为的转变。这使得我们能够估计不同心理状态之间(Thornton & Tamir, 2017)和不同行为之间(Thornton & Tamir, 2019a)的实际转换概率。

Distance on the mental state map predicted actual mental state transitions. The further away two states were on any dimension, the less likely people were to transition between them. Thus, these dimensions likely serve as a scaffolding for social prediction because they describe experienced social dynamics. Not only were perceived and actual emotion transitions correlated, but these associations could be partially explained by how close the emotions were on the dimensions of mental state representation described above. This indicates that people use their maps of mental state space to accurately predict others’ emotion dynamics.

Subsequent research found similar results in the action layer (Thornton & Tamir, 2019a). Across five studies, distance on the action map predicted actual action transitions. The further two actions were on the ACT-FAST dimensions, the less likely people were to transition between them. Moreover, as in the case of mental state representation, people were highly accurate about actual action dynamics, and distance within the action map statistically mediated much of the association between perceived and actual action transitions. This finding adds further weight to the contention that people use their maps of the social world to make accurate social predictions.
随后的研究发现在行动层面上有类似的结果(Thornton & Tamir, 2019a)。在五项研究中,行动地图上的距离预测了实际的行动转变。在 ACT-FAST 维度上,两个行动之间的距离越远,人们转变的可能性就越小。此外,与心理状态表征的情况类似,人们对实际行动动态非常准确,行动地图内的距离在感知和实际行动转变之间的关联中起到了统计学上的中介作用。这一发现进一步加强了人们利用社会世界地图进行准确社会预测的观点。

The ground truth in all the studies above reflects transitional probabilities aggregated across many people or datasets. These data demonstrate that people make accurate predictions about a “generic” other. More recent data suggest that people can likewise make accurate predictions about both specific people and relevant social groups (Zhao, Thornton, & Tamir, 2018). For example, in one set of studies, we asked undergraduates to make predictions about a specific other—either a close friend, or their current roommate—as well as their undergraduate community in general. In all cases, people were able to accurately and specifically predict their friend, their roommate, and their community. This indicates that people may have highly accurate models of state dynamics in general, and that they can also tailor these models to make predictions about the individuals in their lives.

People Make Social Predictions Automatically

People can predict the social future with such high fidelity because prediction is built into the way that people represent social knowledge. In recent research from the action domain, we found that while people watch a movie, their brains spontaneously encode the actions they perceive on the ACT-FAST dimensions, and these ACT-FAST coordinates predict actions later in the movie (Thornton & Tamir, 2020b). That is, a participant’s brain activity at a given moment in time automatically predicts the actions which are actually likely to occur later on. For example, if participants saw a person “running,” then they would encode this action on ACT-FAST—attributing to this act a high degree of animacy. This would, in turn, accurately predict that they were likely to see other high animacy actions in the near future. Merely by encoding perceived actions on an appropriate set of dimensions—dimensions which through real actions flow smoothly—the brain thus automatically predicts likely future actions.
人们之所以能够如此准确地预测社会未来,是因为预测已经融入了人们对社会知识的表征方式。在最近的行动领域研究中,我们发现,当人们观看电影时,他们的大脑会自发地在 ACT-FAST 维度上编码他们所感知到的行动,而这些 ACT-FAST 坐标会预测电影后期的行动(Thornton & Tamir, 2020b)。也就是说,参与者在某一时刻的大脑活动会自动预测实际上可能在稍后发生的行动。例如,如果参与者看到一个人“跑”,那么他们会在 ACT-FAST 上对这个行动进行编码,将这个行为归因为高度的生命力。这反过来准确地预测了他们很可能在不久的将来看到其他高生命力的行动。仅仅通过在适当的一组维度上编码感知到的行动,大脑就能自动预测可能发生的未来行动,这些维度通过真实的行动流畅地传递。

Similar evidence for the automaticity of social prediction comes from the domain of mental states. Whenever someone thinks about a mental state, they do not think about that mental state in isolation; they also spontaneously think about likely future states. For example, when one observes a friend experiencing pride, they can accurately predict that the friend will soon feel happy because the representation of pride incorporates the representation of happiness. Neuroimaging work has provided a unique source of evidence that this is the case: neural patterns associated with a mental state currently under consideration literally resembled patterns of likely future states (Thornton, Weaverdyck, & Tamir, 2019b). The more likely one state is to transition to another state, the more similar the neural patterns that represent them. Importantly, this work also showed that transition predictions, and not simply similarity, drove this neural finding. Even though similarity and transition likelihood are highly intertwined concepts, multiple lines of evidence suggest that transitions, and not similarity, may be primary in defining the conceptual space of mental states.
社会预测的自动性的类似证据还来自于心理状态领域。每当有人思考一种心理状态时,他们并不是孤立地思考这种心理状态;他们也会自发地思考可能的未来状态。例如,当一个人观察到朋友正在经历自豪时,他们可以准确地预测到朋友很快会感到快乐,因为自豪的表征包含了快乐的表征。神经影像学的研究提供了一种独特的证据,证明了这一点:与当前正在考虑的心理状态相关的神经模式实际上类似于可能的未来状态的模式(Thornton, Weaverdyck, & Tamir, 2019b)。一个状态更有可能过渡到另一个状态,它们的神经模式就越相似。重要的是,这项研究还表明,过渡预测而不仅仅是相似性推动了这一神经发现。尽管相似性和过渡可能性是高度交织在一起的概念,但多条证据表明,过渡而不是相似性可能是定义心理状态概念空间的主要因素。

Indirect neural evidence also supports the automaticity of social prediction. In the same study, repetition suppression (also known as fMRI adaptation) tested how the brain reacted to expected and unexpected sequences of mental states. The principle behind this analysis is that, if the brain is constantly making automatic predictions, perceiving information that violates these predictions should elicit more activity than perceiving prediction-consistent information, since the latter requires recalibration of subsequent predictions. In line with this hypothesis, the study found that seeing states in unexpected sequences elicited more activity in the precuneus than observing predictable state sequences. This finding suggests that this brain region might automatically track errors in mental state predictions and update subsequent predictions accordingly.
间接的神经证据也支持社会预测的自动性。在同一研究中,重复抑制(也称为 fMRI 适应性)测试了大脑对预期和意外心理状态序列的反应。这种分析的原则是,如果大脑不断地进行自动预测,那么感知违反这些预测的信息应该比感知与预测一致的信息引发更多的活动,因为后者需要重新校准随后的预测。与这一假设一致,研究发现,看到意外序列中的状态比观察可预测的状态序列在顶枕叶引发了更多的活动。这一发现表明,这个大脑区域可能会自动跟踪心理状态预测中的错误,并相应地更新随后的预测。

People’s maps of the social world play a key role in making the brain’s social predictions automatic. Since the proximity on dimensions such as rationality, social impact, and valence is associated with transitional probabilities, as described in the previous sections, then merely encoding a state using these dimensions implicitly makes a prediction. That is, simply specifying the location of a state on such dimensions provides an indication of which other states are more or less likely. Supporting this idea, this study also found that proximity on these three dimensions of mental state representation statistically mediated much of the relationship between transitional probabilities and neural pattern similarity (Thornton, Weaverdyck, & Tamir, 2019b). This suggests that part of the reason that neural representations of current states resemble neural representations of likely future states may be because all states are encoded as coordinates within the mental state map.
人们对社会世界的地图在使大脑的社会预测自动化方面发挥着关键作用。由于在诸如理性、社会影响和价值等维度上的接近与转换概率相关联,正如前几节所描述的那样,仅仅使用这些维度对状态进行编码就隐含地进行了预测。也就是说,仅仅指定状态在这些维度上的位置就能表明其他状态更可能或更不可能发生。支持这一观点的是,这项研究还发现,心理状态表征的这三个维度上的接近在转换概率和神经模式相似性之间的关系中起到了统计学中介作用(Thornton, Weaverdyck, & Tamir, 2019b)。这表明,当前状态的神经表征与可能的未来状态的神经表征相似的部分原因可能是因为所有状态都被编码为心理状态地图中的坐标。

Conclusion 结论

The brain makes sense of other people’s minds by charting conceptual maps of social stimuli, such as actions, mental states, and traits. These maps make the deluge of information from social world more tractable by reducing the complexity of these stimuli down to coordinates on a few essential psychological dimensions, such as valence or animacy. Moreover, these maps allow people to make accurate, automatic predictions about others’ actions and mental states. Just knowing where someone is on these maps can tell you a great deal about where they are going next because shorter journeys are more likely than lengthy ones. The short hop from joy to gratitude is far more likely than the arduous trek from delight to despair. The research outlined in this chapter offers insight into how people make sense of other people in order to navigate the choppy waters of everyday social life.

So far, research has focused on how the most basic information conveyed by social maps—the locations of traits, states, and actions—can inform social prediction. However, as with physical maps, social maps can also convey other forms of information. For example, if you look at oceans on many physical maps, you will see small arrows that indicate the direction of prevailing winds and currents. If you dropped a sealed bottle in the ocean at given location, you could use its location, along with knowledge about local water currents to make a precise—and directional—prediction about its future location. The space of mental states likewise has prevailing winds and currents: there are statistical regularities in the trajectories that states and actions follow on their respective maps. For example, people in high energy states are likely to gradually flow toward lower energy states as they tire themselves out. Future research must attempt to map these vector fields in the social domain to further refine the predictive framework we describe here.

The framework for predictive social cognition described in this chapter faces at least four additional challenges (Saxe, 2018). First, the types of social knowledge mapped so far all relate to the person—what they are doing, how they are feeling, and their character. However, one of the most potent drivers of real-world behavior does not dwell within any one person. Instead, the situation, or context, crucially shapes how one will think, feel, and act. The current framework must expand to incorporate the power of the situation and its role in social predictions. Fortunately, in recent years, behavioral research and text analysis have suggested potential maps of the situation layer (Parrigon, Woo, Tay, & Wang, 2017; Rauthmann et al., 2014). Future research may productively test how well such maps explain neural representations of situations, and whether these maps have the same predictive properties as the maps of action and state layers described above.
本章描述的预测性社会认知框架至少面临四个额外挑战(Saxe,2018)。首先,迄今为止绘制的社会知识类型都与个人有关——他们在做什么,他们的感受以及他们的性格。然而,真实世界行为最有力的驱动因素之一并不驻留在任何一个人身上。相反,情境或背景关键地塑造了一个人的思维、感受和行为。当前框架必须扩展以纳入情境的力量及其在社会预测中的作用。幸运的是,近年来,行为研究和文本分析已经提出了情境层的潜在绘图(Parrigon,Woo,Tay 和 Wang,2017;Rauthmann 等,2014)。未来的研究可以有益地测试这些绘图能否很好地解释情境的神经表征,并且这些绘图是否具有与上述行动和状态层绘图相同的预测性质。

Second, cultural differences may shape the way people construct maps of social knowledge. Societies differ greatly in the way they perceive emotions, the value they place on different traits, and in the actions that people typically perform (Ching et al., 2014; Gendron, Roberson, van der Vyver, & Barrett, 2014; Tsai, Knutson, & Fung, 2006; Watson-Jones & Legare, 2016). Measuring the generalizability and variability across cultures of the dimensions identified above must be another priority for this research program.
其次,文化差异可能塑造了人们构建社会知识地图的方式。不同社会在感情知觉、对不同特质的价值观以及人们通常采取的行为方面存在显著差异(Ching 等,2014 年;Gendron,Roberson,van der Vyver 和 Barrett,2014 年;Tsai,Knutson 和 Fung,2006 年;Watson-Jones 和 Legare,2016 年)。衡量上述维度在不同文化中的普适性和变异性必须成为该研究项目的另一个重点。

Third, although the model of predictive social cognition described in this chapter has demonstrated its ability to predict real social experience, it faces another major challenge in making these predictions precise. Specifically, this probabilistic framework must incorporate propositional information. For instance, the current framework can describe the properties of “desire” as an abstract mental state, but the meaning of desire can change depending on what one desires. Desiring a cheeseburger and desiring a job are both recognizable forms of desire, but each would predict dramatically different behaviors. Other models of theory of mind—such as Bayesian inverse planning models (Baker, Saxe, & Tenenbaum, 2009)—can deal well with these sorts of propositional problems. However, these types of models do not scale to real-world experience as easily as the current model. Finding ways to unite these models may prove challenging, but a variety of emerging methods, such as word vector embeddings to quantify semantics (Pennington, Socher, & Manning, 2014), may help address these challenges, as similar neural networks have shown to also implicitly represent propositional relations (McCoy, Linzen, Dunbar, & Smolensky, 2018).
第三,尽管本章描述的预测性社会认知模型已经证明了其预测真实社会经验的能力,但它在使这些预测变得精确方面面临着另一个重大挑战。具体来说,这种概率框架必须纳入命题信息。例如,当前的框架可以描述“欲望”作为一种抽象的心理状态,但欲望的含义可能会随着一个人的欲望而改变。渴望一份奶酪汉堡和渴望一份工作都是可以识别的欲望形式,但每种欲望都会预测出截然不同的行为。其他理论心智模型,比如贝叶斯逆向规划模型(Baker, Saxe, & Tenenbaum, 2009),可以很好地处理这些命题问题。然而,这些类型的模型不像当前模型那样容易扩展到现实世界的经验中。 寻找统一这些模型的方法可能会很具挑战性,但各种新兴方法,比如词向量嵌入来量化语义(Pennington, Socher, & Manning, 2014),可能有助于解决这些挑战,因为类似的神经网络也已经显示出隐式表示命题关系(McCoy, Linzen, Dunbar, & Smolensky, 2018)。

The fourth major challenge faced by the predictive model of social cognition is the question of how the mind learns to map the social world. The dimensions of mental state representation arise over the course of development—infants do not start off understanding mental states on all of the dimensions which adults do (Nook, Sasse, Lambert, McLaughlin, & Somerville, 2018). Do children learn new dimensions by observing statistical regularities in emotion dynamics? Perhaps—it is well known that children possess the ability to learn the transitional probabilities between components of speech (Saffran, Aslin, & Newport, 1996), so the same might well be true with respect to other social stimuli. However, it is also possible that children may have “built-in” core knowledge, or inductive biases that help them learn social maps more adeptly than they otherwise might. Future developmental and comparative research, as well as study of machine intelligence, may help to answer this question.
社会认知预测模型面临的第四个主要挑战是心智如何学习将社会世界映射出来的问题。心理状态表征的维度是在发展过程中出现的,婴儿并不是从一开始就理解成人所理解的所有心理状态的维度(Nook, Sasse, Lambert, McLaughlin, & Somerville, 2018)。孩子们是通过观察情绪动态中的统计规律来学习新的维度吗?也许是的——众所周知,孩子们具有学习语音组成部分之间的转换概率的能力(Saffran, Aslin, & Newport, 1996),因此对于其他社会刺激也可能是如此。然而,孩子们也可能具有“内置”的核心知识,或归纳偏见,这有助于他们比其他情况下更熟练地学习社会地图。未来的发展性和比较性研究,以及对机器智能的研究,可能有助于回答这个问题。

As the model of predictive social cognition described in this chapter becomes more comprehensive and more refined, it holds considerable promise for addressing problems of societal importance. For instance, it may provide a concrete way to quantify abnormal social cognition, such as in Autism Spectrum Disorder, or to track how children learn social concepts over development. Indeed, a theory of “Mind-space”—similar to the mental state and trait layers of our predictive model—has been proposed as a way to gain traction on individual differences in social cognition (Conway, Catmur, & Bird, 2019). Predictive social cognition may also provide a roadmap for enhancing artificial intelligence in the social domain, allowing smart devices to better anticipate people’s needs and to interact with people in more natural, human-like ways. Finally, the shortcuts the brain takes to social understanding could reveal the precise sources of harmful social biases and suggest potential approaches to mitigating them. Even optimistically, such development remains many years away, but nonetheless, understanding predictive social cognition holds much promise for the future.
随着本章描述的预测性社会认知模型变得更加全面和精细,它在解决社会重要问题方面具有相当大的潜力。例如,它可能提供了一种具体的方式来量化异常的社会认知,比如自闭症谱系障碍,或者追踪儿童在发展过程中如何学习社会概念。事实上,“心灵空间”的理论——类似于我们预测模型的心理状态和特质层——已被提出作为一种获得社会认知个体差异的方法(Conway, Catmur, & Bird, 2019)。预测性社会认知也可能为增强社会领域的人工智能提供一条路线,使智能设备更好地预测人们的需求,并以更自然、更类人的方式与人们互动。最后,大脑在社会理解中采取的捷径可能揭示有害社会偏见的确切来源,并提出减轻它们的潜在方法。即使乐观地看,这样的发展仍然需要很多年,但是理解预测性社会认知对未来有很大的希望。