人物专栏 | Heidi Harley博士访谈（上）

点击上方蓝字关注我们

编者按

《理论语言学五道口站》（2024年第25期，总第324期）人物专栏与大家分享牛津大学语言学学会就分布式形态学相关问题对美国亚利桑那大学语言学系Heidi Harley教授所做访谈。访谈内容转自网站：https://www.youtube.com/watch?v=cqAtZrC-gOo，由本站成员黄静雯、安安翻译。后续内容将在人物专栏下一期的推送中继续与大家分享，敬请期待。

采访人物简介

Heidi Harley博士

Heidi Harley博士，美国亚利桑那大学语言学系教授，研究涉及形式句法学、形态学及词汇语义学，主要研究领域为分布式形态学，相关研究成果发表于Language、Linguistic Inquiry、Lingua等知名期刊。著作包括English Words: An Introduction一书，已于2006年5月在Wiley-Blackwell出版。由于她在语言学领域做出的杰出贡献，Heidi Harley于2018年被美国语言学协会（The Linguistic Society of America）任命为协会成员。

Brief Introduction of Interviewee

Heidi Harley, professor of Linguistics at the University of Arizona. Her research focuses on the syntax, morphology, and lexical semantics of language. Her work has appeared in various venues including Language, Linguistic Inquiry, Lingua, Journal of Linguistics, and Studia Linguistica. She is the author of the textbook English Words: An Introduction, which has been published by Wiley-Blackwell in May, 2006. For her tremendous contribution to the linguistic field, Heidi Harley was nominated by The Linguistic Society of America as one of the LSA membership in 2018.

访谈内容

01.

主持人：请问什么是分布式形态学（distributed morphology）？

Heidi Harley博士：分布式形态学是一种自上而下的语法模型。句法学家通常利用其他理论解释形态的具体实现，形态学家也不太深入考虑自己所提出的研究方法能否和句法框架兼容。分布式形态学基于最简方案的理论，形成了从词项选择到句法构建、再到音系实现的理论架构。因此，该理论能让我们更好地理解句法、语义和形态或是构建一个涉及句法、语义和形态学的完整模型 。

02. 

主持人：目前我们在分布式形态学里遇到比较争议的问题是词汇主义。请问您能就此进一步阐述吗？另外，您认为这些问题是概念上、方法上还是经验上的呢？

Heidi Harley博士：好问题。词汇功能语法（lexical functional grammar，LFG）和分布式形态学有许多相似之处，因为二者都是自上而下的理论。当我们在考察语言的亚结构特征时，两个理论能为你提供不同的角度。分布式形态学的核心观点是造词和造句是相同的过程，都为句法操作。最简方案理论认为我们将基元以合并的方式建构了句子的层级结构。而在分布式形态学的研究框架中，这些成分是抽象语素（subword abstract），即特征丛集（feature bundles）。利用这些成分，我们也可以构建一个层级结构，该结构能满足把语言成分联系起来的相关句法规则。之后，该结构会进行进一步的形态处理。只有在本阶段才会引入那些与词汇构建相关的成分，也只有在该阶段才能决定所生成的语言是由单个语音成分负载许多词素的粘着语或综合语，还是单个词基本上都是单形态成分的孤立语。因此，分布式形态学秉承两条原则：第一，句法推导自上而下，即结构是通过合并形成的。第二，迟后填音（late insertion），也就是说只有当结构已经基本形成后才会进一步引入音系成分。

以上操作完成后，从列表一（即终端表）出来的成分会进入列表二（即词汇表），此时会引入音系内容，而且句法结构会以“波你尼”式（Paninian）竞争方式¹得到最佳的音系实现。在倾向于表征和倾向于优选论的视角下，这是结构语音实现的优化过程。但实际上并非如此，其实 DM是以一种基于规则的形式化方法实现了该过程。这是一种典型的、具有生成音系学风格的形式化方法。现在，我们有了一个结构和一个音系成分，但至于该音系成分能否很好地实现该结构，这需要他们以“波你尼”这种典型方式来竞争。分布式形态学之所以强烈反对词汇主义与其所构建的整体模型有关。例如，对于词项是什么，典型的定义认为其包括句法信息、语音信息以及心理信息，而这三类信息构成了一个词项。

在分布式形态学的理论框架中，语法成分不包含这三类信息。其中，句法和语义信息仅为词项的子部件，这部分信息位于列表一，即终端表。该列表中有各类句法和语义构形成分并且负责语言表征的句法和语义部分。这些部分在构建基于合并的成分中充当输入部分。除此之外，我们还有列表二，即词汇表。主要负责匹配句法和音系的信息。将该过程对应到词汇功能语法的表征中，我们既有包含句法、语义、音系特征的矩阵，这些信息位于列表一，也有包含句法、音系特征矩阵，这些成分位于列表二，同时它们会通过竞争的方式来实现对应的音系结构。另外，我们还有列表三，该列表主要负责表征结构的真值。也就是说，这个阶段与在列表二的音系实现类似。在列表三，已有的句法语义成分会在特定语境下得到解读并得到最终的意义诠释。

列表二和列表三里的音系和语义环境可大可小。我们可以有音系成分和句法信息一一对应的单形态成分，这也是列表中常规的成分之一。我们也可以有一个对周围结构特性很敏感的小音系片段，这可能是某类复杂的语素变体，但仍是在特殊形态音系境中的具体实现。也就是说，在列表三中，我们既有句法成分，也有与之对应的真值条件解读，从而从理论层面实现该结构的意义诠释。而且该成分可以是纯粹的组合性成分，即其总是提供相同的组合性意义，例如复数义的成分。

同样的，也可以有一个真值条件高度依赖其周围成分的句法组块，因此在特定的语境中，该成分会产生特定的解读。不过，当我们谈及习语时，是否应该将语言切分成小块仍具有一定争议。这些习语有特殊的含义，也就是说其意义并不完全由其构建组合而来，不过它们其实仍是组合性的，只不过这种组合性具有具体且受限的特点。以上便是列表三的全部内容。对于反词汇主义而言，句法-语义特征部分、句法-音系特征部分以及对应的语义信息被分别放置于三个不同的列表中，并且这些信息会与结构以不同的方式互动。因此，词汇主义并不是分布式形态学模型或词汇功能语法的一部分，因为词已经失去其首要地位。

03.

主持人：谢谢您。接下来这个问题您也在之前的回答中稍有提及，请问在分布式形态学的理论框架中，语言变异应该从推导的哪一部分进行解释呢？在其他的理论中，语言变异常被解释为词库内部的不同。DM对此的看法是否相同呢？

Heidi Harley博士：你可以就你所说的“语言变异”给出具体的例子吗？

主持人：我指的是不同语言之间的变异。如果是同样的生成性操作负责生成所有的词或短语，那么为什么生成结果在不同语言中存在差异呢？

Heidi Harley博士：我们知道，分布式形态学中关于句法机制的看法沿袭强势的“最简方案”理论观点，因此认为特定的跨语言句法差异与列表一中的特定语素密切相关。这么来看，我们既可以说这种语言差异是“词汇性”的也可以说它不是“词汇性”的：说它是词汇性的理由是，列表一中的语素是组合系统的输入项，决定了该系统的组合结果；而说它不是则是因为这些语素并不具有任何的音系内容，也就是说它们并非通常意义上所定义的“词”。然而我们之所以认为这是一种强势的“最简方案”观点就是因为该理论把那些导致差异产生的因素都归入列表一之中。

接下来便涉及句法操作。在“最简方案”理论框架下，句法所做的事情并不复杂。简单来说，它所做的只是判断“某两个成分是否能够结合？其中一个成分是否选择了另一个？而这种选择性要求在该理论中被诠释为特征的匹配。也就是说，一个成分是否要经过移位来对功能中心语的特征进行核查？”就我对此的理解来看，我认为在这一点上中心语驱动短语句法（Head-driven Phrase Structure Grammar，HPSG）与词汇-功能语法的看法相近，共同点就在于它们都认为列表一中元素的特征驱动句法操作。我推荐你们阅读Ash Asudeh，Dan Siddiqi和Paul Melchin近期的研究。他们尝试在LFG模型中实现迟后填音的操作。他们所呈现的也正是成功纳入迟后填音这项分析技术的LFG修正模型。

04.

主持人：太好了，感谢您的推荐。我还想要更加深入地和您探讨一下列表三。在之前的回答中，您提到列表三与真值条件有关。我的问题是，在DM理论的推导过程中，语用因素是否作用于词语的生成；如果是的话，又是在哪一阶段的推导中发挥作用呢？同时，我们也提到列表一和列表三都涉及一定的语义内容，那么这又是怎样影响词的推导的呢？

Heidi Harley博士：实际上，列表一究竟涉及何种程度的语义内容是一个非常有趣的问题。其实，列表一的内容也有可能只与句法相关。人们认为列表一包含语义内容是因为他们很难想象这一系统何以在毫无相关信息的情况下运行并得到预期的语义结构的。但显然在分布式形态学中我们不需要做出这一假设。列表三中的信息被称为“百科信息”，即心理语言学家或心理学家所定义的“语义”信息。正是因为有这样的“百科信息”，我们才得以判断为什么“Colorless green ideas”这样的组合是不可说的：虽然在结构上没有任何问题，但它存在语义冲突。因此，我们说列表三是包含了现实世界的真值条件的。这也非常像是与某种概念层面的接口，在这个层面上我们以这种百科全书式的世界知识方式来评估事物。

分布式形态学着重考虑了句法、语义与形态三个模块之间的接口问题，但却较少涉及这些与语用接口的内容。就你提出的这个问题而言，我认为我们可以从近年来“最简方案”视角下的句法、语义、语用相关研究来入手进行分析。目前来说，我发现我还很难就一些问题给出充分的回答，比如说在当前的模型中，我们应该怎样分析不同表征之间的竞争情况以及话语含义呢？让我们来看这样一个例子。我们知道，单词“bad”的比较级“worse”和“badder”之间存在差异。这两个词是呈现“同素异干”语素变体的常见例子，即“bad”的异干语素变体为“worse”；而规则的比较级“badder”是“bad”的非异干语素变体。我们可以用 “worse”来表达更高程度的“糟糕”；而“badder”这一更为特别的语素变体则更接近于“sick”、“cool”这样的习语所表达的意思，比如说它在“the niftiest guy is the baddest guy”一句中的使用。我想在某些学者所持的语用视角中，例如在格赖斯会话含义理论（Gricean’s view of implicature）中，“badder”一词的使用可能与话语含义有关。也就是说，两个形式相互竞争来实现特定的话语含义。如果说话人想要表达通常意义上的“更加糟糕”，他们会选择“worse”。而如果是相反情况的话，我们则可以根据其话语含义排除使“worse”成立的真值条件。

现有DM理论尚未对此进行解释。结合一些相关文献来看，我目前能想到的唯一解释是列表一中同时存在两个“bad”相关抽象语素，这两个“bad”分别对应两种不同的含义。这两个语素都不具有任何音系内容，它们仅有与其相应的索引。根据我们对这两个抽象语素的选择，在推导的最终阶段便分别得到“worse”或“badder”。虽然我仍不清楚我们应怎样将这种基于话语含义的真值条件实现纳入到现有的模型中，但我们可以确定的是这应发生于推导最初的词项选择阶段。

05.

主持人：那么DM理论又是怎么处理屈折形态与派生形态之间的界限的呢？

Heidi Harley博士：这一界限在DM中是不存在的。在推导中，语素在语义与音系方面的实现都是有条件的。就我们所讨论的屈折、派生界限来说，我们可以重点关注音系方面语素的实现。最简方案中存在“语段”（phase）这一概念，即认为句法的推导发生于某种语块中。因此，据此假设，我们只需关注某些语块而非整个句子中的所有成分。我们可以将语段概念与实现条件局域性观点结合来进行分析。

考虑到语素是有条件地实现的，在句法结构上距离更近的成分会首先影响其实现，其次才是距离较远的那些成分。因此对于特定成分的合并来说，我们可以根据它距离词根的距离而分析这究竟是偶然情况下发生的合并或是由DM模型所决定的合并。而我们所说的屈折/派生区分、能产性/非能产性区分和组合性/非组合性区分也与成分与词根的距离有关。在DM理论中，可以决定或改变所黏着成分词类的语素被称为定类语素，它们通常被归类为派生形态；而那些贡献一致、数、语气、时态等“语法信息”的成分则被归类为屈折形态。观察它们与词根的距离，我们可以发现前者距离要远近于后者，这也就是为什么前者具有那些使它们被归类为派生形式的性质。这些定类成分通常更不规律，其分布也更加不规律；它们在语义上也更为特殊。而两类成分分布位置距词根的远近则是由形式层级以及DM模型组合意义的方式所决定的。因此，这种位置分布的差异也便可单纯地从语义或句法出发进行解释：除了某些南美语言或其他语言中可能的例外情况，已经具有时态标记的成分是不可以重新定类的；同样，也不能再对已经携带了名词化语素的成分进行时态的屈折变化。所以，为了保证正确的解读，这些屈折语素必须分布于远离词根的位置，也更具有规律性。

但这些被用来区分派生/屈折形式的性质其实是非固定的，即它们并非其中一种成分的专有属性。派生形式也可能是极为规律的，以英语中动词的ing分词形式为例；而像时制或数的屈折形式也可能是极为不规律的，这种不规律性也体现在解读上。以英语中只以复数形式出现的名词为例（pluralia tantum）为例，虽然它们在形态、句法上是规律的，但从它们的解读情况来看，这类屈折形态也体现出了一定的不规律性。因此在DM中，派生、屈折形态之间并不存在根本性的界限。不论哪种形态都是以自下而上的方式生成的。只不过两种形态成分会倾向于表现出不同特质，这也是其他理论中将两者进行区分的原因。

我们接下来来讨论一下之前所说“语段”的作用。部分学者认为语素实现的限制只作用于语段内部成分。也就是说，如果列表一或列表三中成分的实现受到了一定的条件限制，那么这种限制也是相对于一定结构而言的，即其只作用于语段中的成分。在此基础上，一些学者认为在推导过程中存在多层级的边界，也就是Ramchand提出的“第一语段句法”（first phase syntax），即第一语段中包含那些在形态上不具有规律性的成分。而一旦这一语段被送往接口，下一语段内部成分的推导则不对其造成任何影响。

总体来看，我们既可以从成分与词根之间的相对距离来解释派生形态与屈折形态各自的性质；我们也可以依靠“语段”这一推导发生的封闭区域的边界来对其内部的独特性、不规律性成分与其外部的规律性成分进行区分，即“语段”概念也与派生/屈折区分有一定的相关性。其实，“语段”这一概念仍待进一步的研究，与之前一段时间之内较为集中的研究相比，当下人们对这一概念的研究热情已不比从前。Maya Arad在她2002年关于语义词根的研究中提到成分的规律性以其是否处于定类语段中的情况进行区分：其中的成分具有语义灵活性；其外的成分则表现出绝对的规律性。然而David Pesetsky, Ezer Rasin以及Omer Preminger提出了相反意见：由于首个定类语素环境外的成分可以获得独特解读且该环境内的成分也展现出一定的规律性，因此实际上并不存在语段边界来对派生/屈折形态进行区分。目前，学界尚未定夺这两种分析的正误，我们既可以通过成分与词根之间的距离来将派生/屈折形态之间的差异分析为一种梯度差异；也可以以语段作为区分两者的界限。总之，派生/屈折形态的区分在DM理论中存在多种分析思路，并无定论。而在LFG理论中，两者的差异则是巨大的——派生形态会产生同时具有句法、语义、音系信息的词汇；而屈折形态则依赖句法推导的完成。而在DM理论中，则不存在这种生成过程的差异。

编者注

1. 波你尼，约公元前4世纪梵语研究学家，著有《八章书》，曾采取形式主义模式系统地描写过梵语的语法结构。波你尼将语词进一步划分为更小的意义单位并从四个阶段对表征的实现进行分析。更多可参考Kiparsky2002年相关研究On the Architecture of Panini’s Grammar。

English Version

01. 

Host: What is distributed morphology?

Dr. Heidi Harley: Distributed morphology is a sort of top to bottom model of grammar. If you are just a syntactician, you might often just hand off the nitty-gritty of morphological realization to some other theory. If you are a morphologist, you might not be thinking too hard about whether your proposals mesh well with any particular syntactic framework. Distributed morphology takes you from lexical selection and syntactic building all the way to realization and phonology within a minimalist sort of conception of how the grammar is organized. So, it enables you to do a better understanding of syntax, semantics and morphology, or a complete model of it.

02. 

Host: One of the areas that we come across, a sort of debate within the distributed morphology sphere, is the issue with lexicalism, and can you expand on that, and if you think these issues are conceptual, methodological or empirical?

Dr. Heidi Harley: This is a great question. LFG and Distributed Morphology have more in common than you might think, given that both of them are, as they say, kind of top to bottom theories. They give you angles to work when you’re looking at problems in any of those substructural subdues of language. Distributed Morphology takes at its heart the idea that the word building operations and the sentence building operations are the same set of operations, namely, syntactic operations. So just as a minimalist would say that you build a sentence by merging primitive elements together to build a hierarchical structure. Within a disturbed morphology framework, those elements are sub-word abstract, feature bundles. And so you build them, you make a structure that has hierarchical structure to it that meets all of the sort of syntactic rules that the language has for relating pieces of the structure to each other in various ways. And then you hand it over to the morphological component. And it’s only at that point that sort of elements that are relevant for words sized building come into the picture or not. So it’s only at that point that you know whether you will be a strongly agglutinating, probably synthetic language with many pieces to your individual phonological words, or whether you going to be in isolating language, where you just basically have word size units that are monomorphemic in the sort of classical sense. So, there are two slogans for distributed morphology. One is syntax all the way down, which refers to the first piece that I describe, or you build all your structure using merge. And the other is late insertion. So, you don’t bring phonological material to the table until you basically have all the structure, or sub-piece of the structure that you’re working on.

And then you start going to List Two where you go, and you get a bunch of phonological material, and you try to find the best representation for that syntactic structure in a kind of a competition, Paninian of a way. You could think of that piece as kind of an optimization of the phonological realization of the structure, if you are sort of representationally inclined, optimality theoretically inclined. But that’s not the way. The DM models that piece in the very rule-based formalism. That’s kind of classic SPE style formalism. Here’s a structure, here’s a phonological piece. Is this phonological  piece going to be a good competitor to realize that structure? So they kind of compete in a classic Pininian-Kiparsky kind of a way. The sort of strong anti-lexicalist piece of distributed morphology has to do with that overall architecture, such that, in a classical idea of what a lexical item is, you have a syntactic piece, a semantic piece and a psychological piece, which form a kind of a triple. That’s your lexical item.

In distributed morphology, there’s no single piece of the grammar that is that three-way piece. So the syntactic and semantic subparts of it, sub parts of the lexical item are linked in List One. So you have a whole bunch of syntactical, semantic formatives that have the syntactic and semantic parts of the representation.Those are the input to the structure-building, merge-based piece.Then you have another list, which basically matches syntactic parts and phonological parts. That’s a separate list called the Vocabulary. So if you think about your sort of LFG representation, where you have, like a SYN-SEM-PHON matrix, which is in List One, the SYN-PHON part of the matrix is in List Two. And that’s what competes to realize the structure, those little bits. And then you have this third list, called List Three, where there is basically a SYN-SEM part that includes the truth conditions for the pieces.So you think of the truth conditions as the analog to the phonological realization on List Two, andin List Three, you have a syntactic piece and a semantic piece that tells you, in this context, I will have the following truth conditions.

And both the semantic context and the phonological context for List Two and Three can be kind of large or small. So you could have something that’s “monomorphemic”, where you have just one little phonological piece that corresponds to one little syntactic feature. That would be a normal element of list, too. But you could also have a little phonological piece that is sensitive to a whole bunch of other properties of the structure around it. That would be some complicated allomorph of something, but it’s a special realization in a special morphophonological context. Analogously, you could have, in List Three, you could have a little syntactic piece and a little, truth conditional interpretation of that piece model theoretic truth condition interpretation of that piece. That would be like a purely compositional element that just whenever you run into it, it contributes the same trick conditions over and over and over again, and you kind of recognize that piece contributes plurality or something like that.

Analogously, you could have a little syntactic chunk whose truth conditions are very dependent on the structure that’s around them. So they could yield a peculiar interpretation in some specific context. There’s debate about whether this is the right way to cut the part or not, but you could think of those kinds of things as idioms. So you have a special truth condition for those phrases. And so that brings in a sort of way, you do kind of these non-strictly compositional. They are strictly compositional, but they’re strictly compositional in a very specific conditioned way. That’s all part of List Three. So the anti-lexicalist part is the fact that you take the SYN-SEM parts, you take the SYN-PHON parts, and you take the other kinds of semantic parts, and you split them up into these three lists of items which interface with the structure in different ways. So, that’s the way in which lexicalism is not a part of the model, in that the sort of primacy of the word is denied or lived without.

03. 

Host: Thanks. I think you touched on it a little bit, but where’s linguistic variation to be explained? Those sorts of things are often described as all within the lexicon. Is that still held in DM?

Dr. Heidi Harley: Could you give me an example of the particular kind of variation that you’re thinking of?

Host: I meant between languages. So if the generative procedure to put stuff together is doing all the work, then why does it look different in different languages?

Dr. Heidi Harley: So since distributed morphology is couched within a strongly minimalist kind of view of how the syntax works, the way in which particularly synthetic variation is dealt with cross linguistically has a lot to do with the particular elements in List One. So in that sense, that kind of variation is “lexical”, in that it’s the elements that are input to the combinatory system that tell you how the combinatory system is going to build something.It’s not “lexical” in the sense that those elements don’t have any phonological content. So they’re not words in any normal idea of what a word is, but it’s strongly minimalist that you do try to put all the conditioners of variation into that List One batch.

And so then the syntax is going to come along. In a minimalist framework, syntax is not supposed to do anything super fancy. It’s just going to look and see, “can I put these two things together? Does one select for the other, which is modeled in terms of featural match of some kind. If I have this element, I have to move it to validate a feature against some higher functional head?” All of that piece of it actually, insofar as I understand, sort of, the HPSG view of the world is actually very compatible with the LFG view of the world, because it’s all driven by the features on the elements in List One. I want to commend to you the recent work of Ash Asudeh, Dan Siddiqi and Paul Melchin, who are actually trying to operationalize an LFG view of a late insertion model. So they they’ve been giving presentations on their revised LFG model that incorporates particularly the late insertion piece into LFG, and don't seem to be running into any fundamental conceptual roadblocks.

04. 

Host: Wonderful. Thank you. I’m going to poke a little bit more at the third list. So you’re saying something about the truth conditions. Could I ask how, and if it at all where pragmatics might come in? Then if we’ve sort of got List One and the List Three being both sort of semantic, how that would work?

Dr. Heidi Harley: So it’s actually interesting to wonder exactly how semantic List One is. List One might be just syntactic. And people thought of it as semantic because they couldn’t conceive of any way to make the system run without any idea of the semantic structure you’re going for at the beginning of it. But certainly you don’t need so in the classic distributed morphology literature. The kind of information that is present in list three is called encyclopedic and it corresponds to what psycholinguists or psychologists think of as semantics. You know, it’s all of the stuff that tells you that “colorless green ideas” is oxymoronic, even though there’s nothing sort of structurally wrong with it. So List Three, which contains these kind of real world truth conditions.This is very much an interface with some kind of conceptual level where you evaluate things in this kind of encyclopedic world knowledge, kind of a way.

I would say that distributed morphology thought a lot about the interface between syntax, semantics and morphology. We have not got a ton about the interface between all of that and pragmatics. So I would say that probably the kind of syntactic, semantic, pragmatic work that has really exploded in the last few years that people approach from a minimalist perspective, I think that would be the place I would start to kind of answer that question. In terms of the kinds of questions that I think are very hard to answer for me so far, are like, how these kind of ideas about competition and implicature work in this model? So, just to take a probably not appropriate example, we can look at the difference between “worse” and “badder”. Right? So this is sort of a famous case where you have a normal comparative form of “bad” that is suppletive. And then you have, like, the regular comparative form of “bad” that is not suppletive. So you have “bad-worse” and the normal run of things when you want to say “more bad”, and you have “bad-badder” when you have the special allomorph that means something like…all of the things I can think of to realize this meaning are idioms, but like “sick” or “cool” or “weird”, or, you know, just “the niftiest guy is the baddest guy.”  I think that maybe people who approach the world from some pragmatic, Gricean implicature point of view, might think of that failure of realization of the “worse” stem, as an implicature-based kind of a thing. Like you didn’t choose the normal form; instead, you chose the regular form. And then, by implicature, you could say, well, the normal form is competing with the regular form. If the speaker had wanted to express “normal bad”, they would have chosen “worse”. They didn’t choose worse, they chose “badder”. So that, by implicature, I can rule out all of those “worse” truth conditions.

None of that has any place in the theory at the moment. So the only way that I could think up to model that procedure, and there is some literature on these kinds of questions, it is to say, you’ve got two items in List One. You have the “bad” item, which means “sick” or “cool” or whatever. You have the other “bad” item, which means “bad”. Neither of them have any phonological material. They have some index; 25 is one and 683 is another. And if you pick one, you’re saying “badder”, and if you pick the other one, you’re saying “worse” at the end of the derivation. If there was some kind of way to model the implicature-based discovery of the appropriate set of truth conditions in that word, I don’t know what it is. It would have to happen at the sort of lexical selection stage, at the beginning of the derivation.

05. 

Host: So what happens to the distinction between inflectional morphology and derivational morphology in DM?

Dr. Heidi Harley: It doesn’t exist. So, you have conditioned realization of items, both on the semantic side and the phonological side. Sticking to the phonological sides, since that’s the place where you think about inflection versus derivation the most. Basically, there is a concept in minimalism called the phase, which is some kind of chunking idea about how this syntactic derivation takes place. So you don’t work with your whole set of sentential elements once you work with “chunks”. So the interaction between that kind of idea of building the structure in chunks, and the idea that conditioning is local.

If you sort of just think of it in terms of the likelihood that you are going to be conditioned by something. You’re going to be conditioned by elements that are close to you in the morphological hierarchy, in this syntactic structure, before you’re going to be conditioned by elements that are far away from you. So, the sort of likelihood that a special form is going to merge, it’s just sort of by happenstance or by the way the model is set up, dependent on how local and close to the root you are. And sort of that inflection/derivation divide, that productive/unproductive divide, the compositional/ non-compositional divide, all of those things tend to correlate with how close you are to the root. In the theory, you have this idea that there are little categorizing heads that can change the category of things, that would normally change the category of a form that they attached to, those would normally be classed as derivational morphology. And you have these kind of elements that contribute “grammatical information” like agreement, number, mood, tense, these kinds of things, you would normally classify them as inflectional morphology. Well guess what? The ones that recategorize stuff are typically a lot closer to the root than the ones that contribute the sort of tense information or number information. And so the recategorizors have these properties that attribute to the derivational morphology. They are irregular more often, they are gappier more often. They do this category changing work. And they also are very often semantically special. That’s because in the terms of the way the hierarchy is set up and the way the model puts meanings together, those category changing elements are closer to the root in the structure than a sort of more inflectional elements that are farther away from the root. So just in general, you expect more derivational properties to cluster on those kinds of morphemes that occur lower in the structure for probably purely semantic reasons or syntactic reasons: you can’t recategorize something that’s already got a tense marking on it, unless you’re a South American language or something like that. And similarly, you can’t inflect something for tense that’s got, like, a nominalizer on it. So just in terms of the way the interpretive component has to work, those kinds of elements have to be farther from the root. Those kinds of elements are less likely to be subject to idiosyncratic things.

But as usual, that set of properties that people associate with derivational versus inflectional elements, those set of properties are fluid. So you can find perfectly regular derivational morphology. Participle “ing” forming elements in English are usually good examples; You can find highly irregular inflectional morphology, like tense or number. Even in terms of interpretation, those things can be irregular. So if you think about pluralia tantum in English, they are interpretatively very irregular, even though they are sort of, more or less morphosyntactically quite regular. So there isn’t any hardcore divide. The whole system works the same way from the bottom up.  But there’s this cluster of properties that seem to go with being close, and other cluster properties that seem to go with being farther, that have motivated this inflectional/derivational divide.

Now that said, I haven’t yet mentioned the role that this phases property plays. Several people have proposed that all conditioning can only pay attention to within phase elements. So if your syntax is being built up in chunks, and you have some kind of irregular conditioning on either List Two or List Three elements, that condition can only look at the structure that it’s in, and the structure that’s in is its phase. So, people have made a lot of “hey, including me”, out of this idea that there are these sort of boundary layers in the derivation, such that you’ve got what Rampchand calls the first phase syntax, within that you get any idiomaticity, morphological irregularity. And as soon as you pass that chunk off to the interfaces and go on to your next phase, nothing that happened in the first phase can be dependent on something that’s happened in the next phase.

So you have this kind of gradient idea of closer and farther, that gives you more a derivational and more inflectional kinds of properties. And then you also have this sort of “draw the veil” kind of moment in the derivation, which might give you a very hardcore line between conditioned, irregular and idiosyncratic stuff and perfectly regular stuff. That might also correspond to another idea about the derivation/inflection divide. Because the notion of phase is a moving target, and it had a long decade or so of intense investigation and I think people are not investigating phases quite as much as they used to do. So, for example there’s a famous proposal within distributed morphology from Maya Arad in 2002, particularly about semantic roots, that said, within a categorizing phase, you have all kinds of weird semantic flexibility. As soon as you get outside of it, it’s a hundred percent regular. Everybody sees the perfection of the crystal-building, combinatoric system once you’re outside the first categorizing head. But I see that David Pesetsky and Ezer Rasin and Omer Preminger have just jumped on the bandwagon and presented their re-evaluation of Arad’s argument that basically said no, she was wrong to draw such a hard line. You see idiosyncratic interpretations outside the first categorizing head. You see regularity inside the first categorizing head. There’s no hard line, there’s no sort of phase boundary there. So people have tried to sort of have it both ways, to have both a kind of gradient difference between inflection and derivation. Also, somewhere in the derivation, you get this hard line that you might call a categorical difference, that you could’ve called the inflection/derivation difference. But anyway, we’ve got a few different ways to slice the pie there, but there’s no sort of principle. In LFG, there is a big difference, I think, between sort of word forming stuff that gives you new triples, and then the inflectional stuff, which is, you don’t figure it out until after this syntax is over. There’s no such spot in distributed morphology that says these guys are built in a completely different way than these guys.

往期推荐

本文版权归“理论语言学五道口站”所有，转载请联系本平台。

编辑：黄静雯 安安 何姝颖 

排版：安安 

审校：何姝颖 吴伟韬 时仲

英文编审负责人：黄静雯 安安