Here is a picture of the relationship between natural-language semantics and pragmatics that many theorists would accept: semantics seeks to reverse-engineer the database of word meanings and composition rules by means of which we encode and decode the meanings of sentences. This is going pretty well, because our encoding-decoding algorithms are sufficiently discrete and well behaved that they can be treated as an autonomous grammar in this way. But this approach doesn’t work in pragmatics, which is fueled not by a proprietary database of rules but by a mess of domain-general reasoning that resists computational tractability. Recent debates about the semantics-pragmatics boundary have tended to be about which parts of language-use can be treated as part of the grammatical algorithm and which must be consigned to the much blacker box of pragmatic inference.In Context and Coherence, Una Stojnić argues that far more of our explanation of language-use should be relocated into the neatly rule-governed grammar than nearly anyone would have guessed. The implications of this view are far reaching, but the book revolves around a formal model of discourse that treats the process of assigning specific contents to pronouns and modals as case studies. If someone says, “He should run for office,” what makes it the case that “He’ refers to one person rather than another and that ‘should’ is being used to say something about this person’s obligations as opposed to something about what is likely? Stojnić’s formal model includes representations of contexts that evolve over the course of conversations and that determine the contents of the pronouns and modals that are uttered along the way. She takes the crucial components of contexts to be rankings of entities and possibilities according to their “prominence.” The content of an occurrence of ‘he’ is the most prominent male at that moment in the conversation, and whether ‘should’ is being used as a deontic or epistemic modal depends on what sort of possibility is most prominent.Stojnić innovates by grammaticalizing phenomena that others have taken to be part of the extralinguistic background to speech. Paralinguistic gestures, topic situations, intonational contours, and discourse-coherence relations are all inserted into logical form, which for Stojnić is a grammatical representation of a whole discourse rather than of an individual sentence. Each of these phenomena plays a distinctive role in manipulating the context’s prominence rankings. Some of the empirical predictions of this model are also innovative, many of them organized around Stojnić’s attempts to defend classical logic and propositional semantic contents from recent threats posed by dynamic semanticists and expressivists.The most surprising aspect of Stojnić’s view is her conventionalist interpretation of her model of discourse context. She argues that the context “is a running record of linguistic information that is contributed by discourse-internal, linguistic cues” and that it, “in turn, fully determines the interpretation by fixing the content expressed by the discourse” (10). As Stojnić puts it, “My conception of context is thus linguistic, rather than extra-linguistic” (10).This is an iconoclastic idea. Nearly all recent work takes context to be a body of information that boils down to the interlocutors’ beliefs or other extralinguistic states of mind. These states inform how hearers interpret speech, and since speakers anticipate this, they also inform how speakers design speech. Some would say that these states themselves fix the contents of context-sensitive expressions. Others would insist that the speaker’s intentions do this but that the context is an important part of the evidence that a speaker expects their addressee to rely on when inferring their intentions. These mental states can be changed by linguistic utterances but also in other ways. Robert Stalnaker (1999: 86) illustrates this point by imagining a goat walking into the room where a conversation is happening. By way of perceptual rather than linguistic channels, this event allows interlocutors to felicitously refer to the goat with a pronoun or otherwise speak in a way that presupposes its presence and familiarity.This looks like a problem for the idea that the context is fully under linguistic control. One way that Stojnić responds is by positing a coherence relation called Summary, which connects an utterance to a “perceptually present” topic situation and “makes the central entity in the situation described prominent” (174). Stojnić illustrates this idea with an example in which we see a video of Julia Child uttering (1) while cooking an omelet: (1) That’s your omelet.Most theorists would think of this situation and its omelet as figuring in the extralinguistic beliefs or perceptual representations that lead interlocutors to triangulate on a referent. But Stojnić sticks the situation right into the logical form of (1) and has the Summary relation manipulate its “central entity” into prominence to serve as the referent of ‘that’.Stojnić uses a similar strategy to understand deictic pronouns that are accompanied by pointing gestures or other demonstrations. Most theorists think of demonstrations as extralinguistic evidence of the speaker’s intentions. But Stojnić treats them as further elements in logical form, whose function is to raise the prominence of their demonstrata, thus effectively serving as grammatical antecedents of the pronouns that they accompany.Consider also Stojnić’s defense of modus tollens against Seth Yalcin’s (2012) apparent counterexample, which revolves around a scenario in which the premises of (2) seem true but the conclusion false: (2) If the marble is big, then it is likely red.The marble is not likely red.So, the marble is not big.Stojnić argues that (2) is not a genuine instance of modus tollens because the two occurrences of the modal “likely” are interpreted relative to different prominent possibilities, which gives them different contents. Her explanation for this is that there are coherence relations in the logical form of (2) that make different possibilities prominent at different stages of the discourse. An Elaboration relation in the consequent of the first premise forces us to read the first instance of “likely” as saying what is likely in the initial context modified with the assumption that the marble is big, which has been raised by the first premise’s antecedent. But the two premises are linked by the Contrast relation, which forces them to be interpreted relative to the same context. This means that the second instance of “likely” must be read as saying what is likely in the initial context, unmodified.What are coherence relations, and why must we posit them in logical form just where Stojnić’s explanations need them to be? Some previous theorists have thought of coherence relations as assumptions that interpreters make for broadly extralinguistic reasons, as part of a strategy for understanding how a speaker’s communicative intentions fit with their broader plans. On this view, they flow from our tendency to attribute coherent plans to others as part of a general strategy for understanding behavior.But again, Stojnić finds linguistic convention where others have found pragmatic inference. She argues that coherence relations are part of the grammatical machinery that we must learn to use when acquiring a language. Speakers use subtle but rule-governed signals—sometimes involving intonation or gesture—to indicate which relations go where in logical form. In some cases there are multiple options, in which case interpreters have to guess which arrangement makes the most sense. For example, although Contrast isn’t grammatically required in (2), the two premises would turn out to be inconsistent without it.Stojnić defends the conventionality of both coherence relations and demonstrations, as well as her strategy of putting them in logical form, by citing cross-linguistic variation in how they work (48, 69–71). But presence in logical form does not follow from conventionality.1 And pure conventionality does not follow from a degree of cross-linguistic variation. Different languages sometimes include different grammatical affordances to serve universal communicative needs. For example, every language gives us clause-types that are specialized for making assertions, issuing directives, and asking questions, respectively, but the syntactic implementation details of these clause-types vary greatly (Portner, Pak, and Zanuttini 2019). It does not follow that, for example, the existence of assertion itself is a matter of convention. By contrast, our most deeply conventional activities are conventional not just in their implementation but in their very ontology. Different societies differ not only in how one gets married but also in the nature and purposes of marriage. Surely the fact that we converse in coherent ways and track entities’ prominence during conversation is conventional at most in its implementation details, and not in its raison d’être. To me, this suggests that our linguistic tools for signaling coherence relations may be tools for doing something that we would still be doing even if we didn’t have specialized grammatical tools with which to do it.It remains to be seen whether Stojnić’s technical innovations will be widely taken up. Even if we do possess a surprisingly rich grammatical system of the kind that Stojnić describes, I am still tempted to reject Stojnić’s view of it as an autonomous determiner of content that floats free of our extralinguistic psychology, and to instead think of it as a system for providing our addressees with finely crafted but still merely partial and defeasible evidence of our intentions.Linguistic evidence is partial in that our addressees still need some nonlinguistic sources of information in order to figure out what we’re saying. It seems to me that Stojnić’s strategy of putting topic situations and demonstrations into logical form merely leaves us with a new version of this old problem. In virtue of what, after all, does a demonstration itself refer to one thing rather than another? What selects and individuates a given topic situation and its central entity? And what about cases in which there is no perceptually available topic situation? (For example, suppose that we witness a horrific car accident and, after several hours and some unrelated conversation, I see a disturbed look on your face and say, “Are you still thinking about that?” What determines the referent of “that” in this case?) Presumably, the answers to these questions aren’t narrowly grammatical but have something to do with the interlocutors’ nonlinguistic psychology.When faced with the inescapable partiality of linguistic evidence, Stojnić sometimes says that this shows only that interpreters must sometimes disambiguate between admissible logical forms. But this strikes me as relabeling the problem, since in these cases disambiguating is a matter of selecting the referent of a demonstration or a topic situation and central entity from among many options—the same kinds of tasks, involving the same extralinguistic psychology, that Stojnić’s opponents need to posit.Linguistic evidence is defeasible because successful interpretation sometimes involves recognizing that you have been given misleading linguistic evidence of a speaker’s intentions. Suppose that John intends to tell you that he and his husband practice monogamy, but utters, “My husband and I are monotonous.” Or suppose that Sue wants you to know that she loves a certain Fabergé egg but sloppily points to the wrong thing when uttering “I love that.” What would you have to take John or Sue to have said in order for successful communication to happen in these cases? A conventionalist like Stojnić predicts that if someone carelessly invokes the wrong conventions when they speak, that’s too bad, because it’s the conventions rather than the intentions that set the terms of successful communication. But this seems like the wrong prediction. The best communicative outcome would be for you to somehow realize that the speakers have provided misleading evidence of what they were trying to say, and infer what they intended instead.I am ultimately unconvinced by Stojnić’s defense of conventionalism, then. Nonetheless, I do think that it is the most ingenious such defense that has been so far articulated. And even beyond this foundational issue, this book’s innovative technical and empirical advancements make it required reading for anyone interested in the semantics and pragmatics of pronouns, modals, gesture, intonation, discourse coherence, and context sensitivity in general.

中文翻译：

---

语境和连贯性：突出的逻辑和语法


这是许多理论家都会接受的自然语言语义和语用学之间关系的图景：语义试图对词义和构成规则的数据库进行逆向工程，通过这些数据库我们对句子的含义进行编码和解码。这一切进展顺利，因为我们的编码-解码算法足够离散且表现良好，可以通过这种方式将它们视为自治语法。但这种方法在语用学中不起作用，语用学不是由专有的规则数据库推动的，而是由一大堆阻碍计算可处理性的领域通用推理推动的。最近关于语义-语用学边界的争论往往是关于语言使用的哪些部分可以被视为语法算法的一部分，以及哪些部分必须被置于语用推理的黑盒子中。 在语境和连贯性中，Una Stojnić 认为我们对语言使用的更多解释应该被重新定位到严格规则管理的语法中，这几乎超出了任何人的想象。这种观点的影响是深远的，但本书围绕着一种正式的话语模型展开，该模型将向代词和情态词分配特定内容的过程视为案例研究。如果有人说“他应该竞选公职”，那么为什么“他”指的是一个人而不是另一个人，并且“应该”被用来谈论这个人的义务，而不是谈论这个人的义务可能吗？斯托伊尼奇的形式模型包括在对话过程中不断演变的语境表征，并决定了对话过程中说出的代词和情态动词的内容。 她将上下文的关键组成部分视为根据实体和可能性的“重要性”进行的排名。 “he”出现的内容是对话中当时最突出的男性，而“should”是否被用作道义模态或认知模态取决于哪种可能性是最突出的。Stojnić通过将现象语法化来创新其他人认为这是言语的语言外背景的一部分。副语言手势、主题情境、语调轮廓和话语连贯关系都被插入到逻辑形式中，对于斯托伊尼奇来说，这是整个话语而不是单个句子的语法表示。这些现象中的每一种都在操纵上下文的显着性排名方面发挥着独特的作用。该模型的一些经验预测也是创新的，其中许多预测是围绕斯托伊尼奇试图捍卫古典逻辑和命题语义内容免受动态语义学家和表现主义者最近构成的威胁而组织的。斯托伊尼奇观点最令人惊讶的方面是她对她的传统主义解释语境模型。她认为，语境“是由话语内部的语言线索贡献的语言信息的运行记录”，并且“反过来，通过固定话语所表达的内容来完全决定解释”（10）。正如斯托伊尼奇所说，“因此，我对语境的概念是语言性的，而不是语言外的”（10）。这是一种反传统的想法。几乎所有最近的工作都将语境视为一组信息，可归结为对话者的信念或其他语言外的心理状态。 这些状态告诉听者如何解释言语，并且由于说话者预见到这一点，它们也告诉说话者如何设计言语。有人会说这些状态本身修复了上下文相关表达式的内容。其他人会坚持认为说话者的意图是这样做的，但上下文是说话者希望收件人在推断其意图时依赖的证据的重要组成部分。这些心理状态可以通过语言表达来改变，但也可以通过其他方式改变。 Robert Stalnaker (1999: 86) 通过想象一只山羊走进正在谈话的房间来说明这一点。通过感知而非语言渠道，这一事件允许对话者用代词恰当地指代山羊，或者以预先假定其存在和熟悉的方式说话。对于上下文完全处于语言之下的观点来说，这看起来是一个问题控制。斯托伊尼奇的回应方式之一是提出一种称为“摘要”的连贯关系，它将话语与“感知上存在的”主题情境联系起来，并“使所描述情境中的中心实体变得突出”（174）。斯托伊尼奇用一个例子来说明这个想法，在这个例子中，我们看到朱莉娅·柴尔德（Julia Child）在煮煎蛋卷时说出（1）的视频：（1）那是你的煎蛋卷。大多数理论家会认为这种情况及其煎蛋卷体现在语言外的信念或感知中。引导对话者对所指对象进行三角测量的表征。但斯托伊尼奇将这种情况直接运用到（1）的逻辑形式中，并让概括关系将其“中心实体”突出起来，作为“那个”的指称。斯托伊尼奇使用类似的策略来理解伴随着指向手势或其他演示的指示代词。大多数理论家认为示威是说话者意图的语言外证据。但斯托伊尼奇将它们视为逻辑形式的进一步元素，其功能是提高其演示的突出性，从而有效地充当它们所伴随的代词的语法先行词。还可以考虑斯托伊尼奇针对 Seth Yalcin (2012) 明显反例对 modus tollens 的辩护，它围绕一个场景，其中（2）的前提似乎为真，但结论为假：（2）如果弹珠很大，那么它很可能是红色的。弹珠不太可能是红色的。所以，弹珠不大Stojnić 认为，(2) 并不是 modus tollens 的真正实例，因为情态词“likely”的两次出现是相对于不同的突出可能性来解释的，这赋予了它们不同的内容。她对此的解释是，(2)的逻辑形式中存在连贯关系，使得不同的可能性在话语的不同阶段凸显出来。第一个前提的后件中的阐述关系迫使我们将“可能”的第一个实例解读为在最初的上下文中可能发生的情况，并根据第一个前提的先行词提出的大理石很大的假设进行了修改。但这两个前提通过对比关系联系起来，这迫使它们相对于相同的上下文进行解释。这意味着“可能”的第二个实例必须被解读为在未修改的情况下在初始上下文中可能发生的情况。什么是连贯关系，为什么我们必须以逻辑形式将它们置于斯托伊尼奇的解释需要的地方？ 一些先前的理论家认为连贯关系是口译员出于广泛的语言外原因而做出的假设，作为理解说话者的交际意图如何与其更广泛的计划相契合的策略的一部分。从这个观点来看，它们源于我们倾向于将连贯的计划归因于他人，作为理解行为的一般策略的一部分。但斯托伊尼奇再次发现了其他人发现的实用推论的语言惯例。她认为连贯关系是我们在习得语言时必须学会使用的语法机制的一部分。说话者使用微妙但受规则控制的信号（有时涉及语调或手势）来指示逻辑形式中的关系。在某些情况下，有多种选择，在这种情况下，口译员必须猜测哪种安排最有意义。例如，尽管（2）中在语法上不需要对比，但如果没有它，两个前提就会不一致。斯托伊尼奇捍卫了连贯关系和论证的常规性，以及她将它们置于逻辑形式中的策略，通过引用其工作方式的跨语言差异 (48, 69–71)。但逻辑形式的存在并不源自约定俗成。1并且纯粹的约定俗成并不源自一定程度的跨语言变异。不同的语言有时包含不同的语法可供性来满足普遍的交流需求。例如，每种语言都为我们提供了专门用于做出断言、发出指令和提出问题的子句类型，但这些子句类型的语法实现细节差异很大（Portner、Pak 和 Zanuttini 2019）。 例如，这并不意味着断言本身的存在就是一个惯例问题。相比之下，我们最传统的活动不仅在实施上而且在本体论上都是传统的。不同的社会不仅在结婚方式上有所不同，而且在婚姻的性质和目的上也有所不同。当然，我们以连贯的方式进行交谈并在对话期间跟踪实体的显着性这一事实最多在其实现细节上是传统的，而不是其存在的理由。对我来说，这表明我们用于表示连贯关系的语言工具可能是我们仍然会做的事情的工具，即使我们没有专门的语法工具来做这件事。斯托伊尼奇的技术创新是否有效还有待观察。将被广泛采用。即使我们确实拥有斯托伊尼奇所描述的那种令人惊讶的丰富语法系统，我仍然很想拒绝斯托伊尼奇的观点，即它是不受我们的语言外心理影响的内容的自主决定者，而是将其视为一个系统为我们的收件人提供精心设计但仍然只是部分且可撤销的我们意图的证据。语言证据是部分的，因为我们的收件人仍然需要一些非语言信息来源才能弄清楚我们在说什么。在我看来，斯托伊尼奇将主题情境和论证逻辑化的策略只是给我们留下了这个老问题的新版本。毕竟，示威本身是凭借什么来指代一件事而不是另一件事呢？是什么选择并个性化了给定的主题情境及其中心实体？没有感知上可用的主题情况的情况又如何呢？ （例如，假设我们目睹了一场可怕的车祸，经过几个小时和一些无关的谈话后，我看到你脸上不安的表情，然后说：“你还在想那个吗？”什么决定了“那个”的所指对象在这种情况下？）大概，这些问题的答案不是狭隘的语法问题，而是与对话者的非语言心理有关。当面对语言证据不可避免的偏颇时，斯托伊尼奇有时会说，这仅表明口译员有时必须消除可接受的逻辑形式之间的歧义。但这让我觉得这是对问题的重新贴标签，因为在这些情况下，消歧就是从许多选项中选择演示或主题情境的所指对象和中心实体——相同类型的任务，涉及相同的语言外心理学，斯托伊尼奇的语言证据是可废止的，因为成功的解释有时需要认识到你已经获得了关于说话者意图的误导性语言证据。假设约翰打算告诉你，他和他的丈夫实行一夫一妻制，但他说：“我和我的丈夫都很单调。”或者假设苏想让你知道她喜欢某个法贝热彩蛋，但在说“我喜欢那个”时草率地指出了错误的东西。在这些情况下，为了成功地进行沟通，您需要约翰或苏说些什么？像斯托伊尼奇这样的传统主义者预测，如果有人在说话时不小心援引了错误的惯例，那就太糟糕了，因为是惯例而不是意图决定了成功沟通的条件。但这似乎是错误的预测。 最好的沟通结果是让你以某种方式认识到演讲者为他们想说的话提供了误导性证据，并推断出他们的意图。那么，我最终不相信斯托伊尼奇对传统主义的辩护。尽管如此，我确实认为这是迄今为止最巧妙的此类辩护。即使超越这个基本问题，本书的创新技术和经验进步也使得任何对代词、情态、手势、语调、话语连贯性和语境敏感性感兴趣的人都需要阅读这本书。