That's what she (could have) said: How alternative utterances affect language use
That's what she (could have) said: How alternative utterances affect language use
复制标题
这就是她（本可以）说的：替代话语如何影响语言使用
DOI：
--
复制
发表时间：
2012
期刊：
影响因子：
--
通讯作者：
R. Levy
中科院分区：
文献类型：
--
作者：
Leon Bergen;Noah D. Goodman;R. Levy
这就是她（本可以）说的：替代话语如何影响语言使用 Leon Bergen (bergen@mit.edu) 1 , Noah D. Goodman (ngoodman@stanford.edu) 2 , Roger Levy (rlevy@ucsd.edu) 3 1 Department of Brain and Cognitive Sciences, MIT, Cambridge MA 02139, of Psychology, Stanley University, Stanley CA 94305, 3 Department 语言学系，加州大学圣地亚哥分校，拉荷亚 CA 92093 2 系抽象语言。另一种选择是，这种解释仅仅提供了一种总结语用现象的简洁方式，而语用意义是作为语法的一部分来学习的，即作为语言的一部分而被约定俗成的。研究人员认为，某些类型的语用意义是根据语法计算出来的，而不是根据合作的社会推理（Chierchia，2004）。然而，直觉上，一些实用的推论可以推广到以前无法学习的环境。考虑上面描述的沙拉/龙虾推论。这种推理高度依赖于上下文，并且必须需要超出语法中所学知识的推理过程。但是，约定俗成的含义和社会衍生的含义之间的界限在哪里呢？在这里，我们使用实验和计算模型来研究这些问题。尽管 Q 和 R 原则的解释表面上很简单，但开发一个体现这些原则（或对话准则）的正式框架却是出了名的困难。此外，很少有实证研究来调查语用推论是否依赖于约定俗成的含义。我们将研究不同话语之间对比的最小方法，以推动语用解释。具体来说，我们将研究没有任何语言约定的情况。如果人们的语用解释对这些环境中的对比表现出同样的敏感性，这将提供证据表明社会推理更多地而不是更少地解释了他们的语用能力。我们重点关注语用学中反事实推理的两个传统例子。第一个，标量含义，是由于信息量不断增加的单词之间的对比而出现的。信息较少的含义通常会被强化为信息较多的含义的补充，就像上面“一些”与“全部”的情况一样。因为术语“标量含义”有时被保留来指代词汇项落在信息量的规范尺度上的情况，所以我们将使用术语特异性含义来描述即使在没有这样的规范尺度的情况下信息量较少的含义的强化。第二个，我们称之为霍恩含义，指导对复杂性不同的话语的解释（Horn，1984）。通常，更复杂的结构会得到有标记的（或不太可能的）解释。上面的汽车案例是喇叭含义的一个例子，假设（这是合理的）两个表达式具有相同的文字内容。这两种含义将使我们能够沿着两个不同的维度探索实用的对比效果：信息性和成本。虽然我们将对这两种效果进行建模，但我们研究了替代性话语对语言的语用解释的影响。我们关注两种具体情况：特异性含义（不太具体的话语意味着对更具体话语的否定）和霍恩含义（更复杂的话语被分配给不太可能的含义）。我们根据递归社会推理提出这些现象的模型。我们最复杂的模型不仅能够处理特异性含义，而且也是对霍恩含义的第一个正式解释，它可以在没有先验约定的情况下正确预测人类在信号博弈中的行为，而无需诉诸专门的均衡选择标准。两个实验提供的证据表明，这些含义是在缺乏先前的语言惯例或语言进化的情况下产生的。综上所述，我们的模型和实验结果表明，替代性话语的语用效果可以通过合作社会推理来驱动。关键词：语用学；沟通;贝叶斯建模 简介 语用学领域的一个中心观察是替代话语会影响我们对语言的解释。如果老师说“有些学生通过了考试”，那么这意味着并非所有人都通过了，因为如果他们通过了，老师就会这么说。如果有人被问到他们在餐厅吃了什么，他们说“沙拉”，那么这意味着他们没有得到龙虾；否则他们就会这么说。如果有人说，“我让汽车启动”，那么这意味着启动汽车涉及到比转动钥匙更不寻常的事情。如果没有，他们可能会说：“我打开了汽车。”格赖斯（Grice）（1975）的经典著作中描述了许多其他类似的案例。 Horn (1984) 根据他的 Q 和 R 原则对这些不同的情况提出了统一的解释。这些原则描述了对话伙伴应该如何相互沟通。 Q 原则指出：尽可能多地说。 R 原则规定：不得多说。这些原则解释了像上面这样的反事实话语如何对意义产生影响。如果说话者按照 Q 原则行事，那么当她说一些学生通过了测试时，这必定意味着她已经说了她能说的一切。尤其是她根本没资格说所有的学生都通过了。同样，如果说话者遵循 R 原则，那么当她报告说她打开了汽车时，这意味着更简单的话语不足以表达她的意思。尤其是单纯的说“打开车子”根本无法表达她的意思。关于这些原则（或格赖斯的相关对话格言）的一个基本问题是，当人们务实地解释时，它们在多大程度上捕捉到了人们的在线推理
That’s what she (could have) said: How alternative utterances affect language use Leon Bergen (bergen@mit.edu) 1 , Noah D. Goodman (ngoodman@stanford.edu) 2 , Roger Levy (rlevy@ucsd.edu) 3 1 Department of Brain and Cognitive Sciences, MIT, Cambridge MA 02139, of Psychology, Stanford University, Stanford CA 94305, 3 Department of Linguistics, UC San Diego, La Jolla CA 92093 2 Department Abstract language. The alternative is that such explanations merely provide a succinct way of summarizing pragmatic phenom- ena, and that pragmatic meanings are learned as part of the grammar, i.e. conventionalized as part of the language. Re- searchers have argued that some types of pragmatic mean- ings are computed from the grammar, and not from coopera- tive social reasoning (Chierchia, 2004). Intuitively, however, some pragmatic inferences generalize to settings in which they could not have been previously learned. Consider the salad/lobster inference described above. This inference is highly context-dependent, and must require a reasoning pro- cess that extends beyond what has been learned in the gram- mar. But then where is the boundary between conventional- ized and socially-derived implicatures? Here we investigate these questions using experiments and computational modeling. Despite the apparent simplicity of explanations in terms of the Q and R-principles, it has been notoriously difficult to develop a formal framework that cap- tures these principles (or the maxims of conversation). In addition, there has been little empirical work investigating whether pragmatic inferences rely on conventionalized mean- ings. We will be looking at the minimal ways for contrasts between alternative utterances to drive pragmatic interpreta- tion. Specifically, we will be looking at cases in which there are no linguistic conventions whatsoever. If people’s prag- matic interpretations show the same sensitivity to contrast in these settings, it will provide evidence that social reasoning explains more, rather than less, of their pragmatic abilities. We focus on two traditional examples of counterfactual reasoning in pragmatics. The first, scalar implicatures, arise because of the contrast between words that fall on an increas- ing scale of informativeness. The less informative meaning is typically strengthened to the complement of the more infor- mative meaning, as in the case of “some” vs. “all” above. Be- cause the term “scalar implicature” is sometimes reserved to refer to cases where lexical items fall on a canonical scale of informativeness, we will use the term specificity implicatures to describe the strengthening of less informative meanings even in the absence of such a canonical scale. The second, which we will call Horn implicatures, guides the interpreta- tion of utterances that differ in their complexity (Horn, 1984). Typically, more complex constructions receive marked (or less probable) interpretations. The car case above is an ex- ample of a Horn implicature, assuming (as is plausible) that the two expressions have the same literal content. These two kinds of implicatures will allow us to explore pragmatic contrast effects along two distinct dimensions: in- formativeness and cost. While we will model both of these ef- We investigate the effects of alternative utterances on prag- matic interpretation of language. We focus on two specific cases: specificity implicatures (less specific utterances im- ply the negation of more specific utterances) and Horn impli- catures (more complex utterances are assigned to less likely meanings). We present models of these phenomena in terms of recursive social reasoning. Our most sophisticated model is not only able to handle specificity implicature but is also the first formal account of Horn implicatures that correctly pre- dicts human behavior in signaling games with no prior conven- tions, without appeal to specialized equilibrium selection crite- ria. Two experiments provide evidence that these implicatures are generated in the absence of prior linguistic conventions or language evolution. Taken together, our modeling and experi- mental results suggest that the pragmatic effects of alternative utterances can be driven by cooperative social reasoning. Keywords: Pragmatics; Communication; Bayesian modeling Introduction A central observation in the field of pragmatics is that alter- native utterances affect our interpretation of language. If the teacher says, “Some of the students passed the test”, then this means that not all of them passed, because the teacher would have said so if they did. If someone is asked what they ate at the restaurant, and they say “salad”, then this means that they did not also get the lobster; otherwise they would have said so. If someone says, “I got the car to turn on,” then this means that turning on the car involved something more unusual than just turning the key. If it hadn’t, they could have just said, “I turned on the car.” Many other cases like this were described in Grice’s (1975) classic. Horn (1984) proposed a unified account of these disparate cases in terms of his Q and R-Principles. These principles de- scribe how conversational partners are expected to communi- cate with each other. The Q-Principle states: Say as much as you can. The R-Principle states: Say no more than you must. These principles explain how counterfactual utterances like the ones above have their effect on meaning. If the speaker behaves according to the Q-Principle, then when she says that some of the students passed the test, this must mean that she said all that she could. In particular, she must not have been in a position to say that all of the students passed. Similarly, if the speaker is following the R-Principle, then when she re- ports that she got the car to turn on, this means that a simpler utterance would not have sufficed to convey her meaning. In particular, simply saying that she turned on the car would not have conveyed her meaning. A basic question about these principles (or Grice’s related maxims of conversation) is the extent to which they capture people’s online reasoning when they pragmatically interpret