中文导读

我们说话写字时，为什么积极的词总比消极的词多？研究者早就发现了这种倾向，称之为“波丽安娜原则”（Pollyanna principle），过去通常把它解释为人们喜欢表达好心情。据 ScienceDaily 10 月 24 日报道，美国佛蒙特大学的一项研究给出了另一种解释。文章先介绍自 20 世纪 50 年代起被广泛使用的词义模型：它根据心理学家查尔斯·奥斯古德的研究，用“积极或消极”“兴奋或平静”“控制感强弱”三项情绪指标衡量词义，并被用于心理学和人工智能的文本情绪判断。接着介绍研究团队的做法：他们分析了两万多个单词在书籍、新闻、社交媒体和口语中的数十亿次使用，发现旧模型的三项指标并非彼此独立，而“强或弱、安全或危险、有序或混乱”这三组维度能解释九成以上的词义差异，旧模型约为 72%。然后，研究者把每个词实际出现的频率计算在内，发现与安全有关的词用得远多于与危险有关的词，因此认为“偏爱积极词”很大程度上只是更深层“安全偏向”的影子。最后，文章谈到这种偏向可能与人类求生的基本需要有关，但研究并不能证明它是由进化直接造成的；研究者还认为，新框架或许能帮助 AI 更准确地理解人们的话。阅读时可留意 instead of、rather than 如何表达对比，以及 may、suggest、believe 等词如何为推测划定边界。

正文

Researchers have long known that human language leans toward positive words. People use pleasant words more often than unpleasant ones, a pattern called the Pollyanna principle. For decades, this was seen as proof that we like to express good feelings. A new study from the University of Vermont in the United States offers a different explanation: our words may lean toward safety rather than simply happiness. The research appeared in the journal Science Advances and was reported by ScienceDaily on 24 October.

本段译文

研究者早就知道，人类语言偏向积极的词。人们使用令人愉快的词多于令人不快的词，这种规律被称为“波丽安娜原则”。几十年来，这被视为我们喜欢表达美好感受的证据。美国佛蒙特大学的一项新研究给出了不同的解释：我们的词语可能偏向的是安全，而不只是快乐。这项研究发表在《科学进展》期刊上，ScienceDaily 于 10 月 24 日进行了报道。

Since the 1950s, many scientists have measured word meaning with a model based on work by the psychologist Charles Osgood. It uses three emotional measures: how positive or negative a word is, how exciting or calm it feels, and how much control it suggests. The model has been widely used in psychology and artificial intelligence, especially in programs that judge the emotional tone of a text.

本段译文

自 20 世纪 50 年代以来，许多科学家一直用一个基于心理学家查尔斯·奥斯古德研究的模型来衡量词义。它使用三项情绪指标：一个词是积极还是消极，让人感到兴奋还是平静，以及它暗示了多大程度的控制。这个模型在心理学和人工智能领域被广泛使用，尤其是在判断文本情绪色彩的程序中。

The Vermont team, led by Peter Dodds and Chris Danforth, tested this idea on a huge scale. They examined billions of uses of more than 20,000 words in books, news, social media and spoken language. They found that the three old measures were not really separate from one another. Instead, meaning was better described by three other pairs: weak or powerful, safe or dangerous, and ordered or chaotic. This new framework explained more than 90% of the differences in word meaning, compared with about 72% for the old one.

本段译文

由彼得·多兹和克里斯·丹福思带领的佛蒙特大学团队，在极大的规模上检验了这一观点。他们考察了两万多个单词在书籍、新闻、社交媒体和口语中的数十亿次使用，发现旧模型的三项指标其实并不彼此独立。相反，用另外三组对立概念来描述词义效果更好：弱或强、安全或危险、有序或混乱。这个新框架解释了词义差异的 90% 以上，而旧框架约为 72%。

The team also counted how often each word was actually used, instead of treating every word as equally important. When this was included, a clear pattern appeared: words linked to safety were used much more often than words linked to danger. In the researchers‘ view, the familiar preference for positive words is largely a shadow of this deeper safety bias.

本段译文

研究团队还统计了每个词实际被使用的频率，而不是把每个词看得同等重要。把这一点计算在内后，一个清晰的规律出现了：与安全相关的词，使用频率远高于与危险相关的词。在研究者看来，人们熟知的“偏爱积极词”现象，在很大程度上是这种更深层的安全偏向的影子。

Why might this be? The researchers suggest that language may reflect the basic need to stay alive. Every day, people share information about whether a person, a place or a situation is safe. “Power, danger, and structure,” said Julia Zimmerman, a co-author, “are relevant to every person that’s ever lived.” However, the results do not prove that evolution directly caused the bias.

本段译文

为什么会这样？研究者认为，语言可能反映了求生这一基本需要。每天，人们都在交流某个人、某个地方或某种情况是否安全的信息。“力量、危险和结构，”合著者朱莉娅·齐默尔曼说，“与有史以来的每一个人都息息相关。”不过，这些结果并不能证明这种偏向是由进化直接造成的。

The findings may matter beyond language research. Many AI systems sort text into positive, negative or neutral feelings. If such systems also considered power, danger and structure, the researchers believe, they might understand people more accurately, especially when trust and risk are involved.

本段译文

这些发现的意义可能超出语言研究本身。许多 AI 系统会把文本分为积极、消极或中性情绪。研究者认为，如果这类系统也考虑力量、危险和结构，它们或许能更准确地理解人们，尤其是在涉及信任和风险的时候。

新闻来源

• ScienceDaily · A hidden bias in human language challenges 70 years of science (University of Vermont, 24 Oct 2026)
• Science Advances · Ousiometrics: The essence of meaning aligns with a power-danger-structure framework instead of valence-arousal-dominance (Dodds et al., DOI: 10.1126/sciadv.adr4039)

✳更多文章 →