人物专栏 | David Poeppel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12期，总第311期）人物专栏与大家分享牛津大学语言学学会对David Poeppel博士的访谈。David Poeppel博士，美国纽约大学心理学家、神经科学家。

本期访谈节选自牛津大学语言学学会对David Poeppel博士进行的专访。在访谈中，David Poeppel博士就自己当前的研究回答了语言与神经科学的相关问题。访谈内容转自网站：https://www.youtube.com/watch?v=5uOrTzL8dnI&t=1478s ，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

David Poeppel博士

David Poeppel博士，美国纽约大学心理学系教授，福恩斯特·斯特朗格曼神经科学研究所（the Ernst Struengmann Institute for Neuroscience）科学主任兼执行总裁。2014-2021年，曾任马克斯·普朗克研究所神经科学部门主任。研究涉及人类听皮层生理学、言语感知的神经基础、听觉/言语心理物理学、词汇层面的心理语言学和神经语言学。

Brief Introduction of Interviewee

David Poeppel is a professor of the department of psychology at New York University, the Scientific Director and CEO of the Ernst Struengmann Institute for Neuroscience in Frankfurt, Germany. He was the Director of the Department of Neuroscience at the Max-Planck-Institute (MPIEA) in Frankfurt, Germany from 2014-2021. He is working on four fronts: human auditory cortex physiology, neural basis of speech perception, auditory/speech psychophysics, lexical level psycholinguistics and neurolinguistics.

访谈内容

01.

主持人：您在一些文章里谈到了神经科学和语言学理论有交叉的内容，不过二者也有不同点，您能谈谈二者在哪些领域有相同点以及您是如何看待这种学科交叉的现象吗？

David Poeppel博士：我认为，从长远来看神经科学和语言学理论会有更多的交叉领域。但就目前而言，二者的研究内容仍有较大不同，这与它们实际上存在的差异和彼此的学科性质有关。自21世纪以来，语言学领域取得了瞩目的成就。我们都知道，在过去六十到七十年间，语言学和认知科学以及计算机科学之间的联系越来越密切，其发展有显著提升。一般来说，语言学主要探索我们是谁以及语言是如何构成的，这个领域的研究非常有趣且深入。在这样的研究背景下，语言学家基于许多跨语言的现象提出了很多极具洞察力的观点。因此，语言学主要探究人的语言知识并试着回答不同语言是如何组成的，语言是如何习得且加工的，“语言库存”（inventory）是什么等问题。无论你在研究语言的哪个方面，语言内部或语言间都存在大量的语料。这是所谓的“零件清单”（parts list），这个清单主要包括语言的基元（primitives）或本体（ontology）。目前这个领域有许多研究，虽然我们对语言学的基础结构是什么仍无定论，不过已经存在大量证据表明它们是具有良好动机的假设。同样，在新兴学科（如神经科学）的研究中，也取得了很好的进展。我们其实并不了解大脑的运作机制，但我们已经越来越了解脑中“零件清单”的相关性质，包括其中不同类别的细胞和生理活动。

假如我们认可语言学和神经科学的相关研究，实际上它们的可信度也很高，那我们就可以进一步假设语言是大脑的属性。而基于这个假设，我们也就可以认为两个领域之间的确存在某种非常明确的联系。这就是语言和大脑之间的主要关系，而这也是问题之所在。多年来，我和我的同事一直在担心这个问题，也经常思考如何才能更好地协调这两个领域之间的关系。目前大家得出的共识是，只要我们更深入地探索各自的领域，就能更好地解决这个问题。这其实是由“粒度不匹配问题” （Granulairty Mismatch Problem）导致的，简单来说就是语言学领域在处理语言事实时非常细致，而神经生物学家对语言的研究则相对笼统。在研究问题时解析度的不匹配会带来各种问题，二者之间的相关性只是表面现象。对两个学科而言，这都是一个问题，因此我们必须要有一套相同的术语和共同的概念理解。而且这些理论必须建立在正确的分析水平上，但这其实是非常困难的。不过我认为研究者们会逐步解决这个问题。但还有一个更棘手的问题需要解答，即心智（语言）和大脑的关系是什么，这是心身问题（Mind-body Problems）的例子之一，但目前对此没有较好的解释。我们可以认为语言是大脑的特质之一，当然也可以不再深入探索这个问题。但它的答案可以深入反映类似的关系假说是如何形成的。例如，关于“语素”（morpheme） 这个概念，我们并不能指出大脑对标语素的结构区域和有关功能，但可以提出相关概念来解释这种关联，这常常意味着在研究大脑和语言时，我们的预期是能学到关于大脑或语言的知识，或者理想情况下两者兼而有之，但通常的结果是一无所获。这种情况常被称为“跨学科1+1=0” （interdisciplinary cross sterilization）。这意味着两个学科不会为彼此的研究注入新的活力，反而会因为二者的关联没有特别紧密，导致对两个领域的理解都不够深刻。可是我们如何才能加深这两个学科的联系并深入理解思维和大脑是如何协调、工作的呢，这个问题相当复杂。

02. 

主持人：在语言学领域中，心理语言学常被视作研究语言系统是如何运用的（performance），而理论语言学则被视为对语言能力（competence）的探索，您认为这种学科的区分是准确或者有益的吗？

David Poeppel博士：这种学科的分工和细化肯定是有益的。这个有趣的区分源于二十世纪五十年代至六十年代兴起的生成语法，当时它对语言的理想化提出了许多大胆的假设。当然，一些人认同生成语法，另一些人则谈之色变。不过对我而言，这种理想化十分寻常且毫无争议。提出语言能力这个仅存在于理想说话者社区中的概念是很正常的，这是理想化的过程，就像一个没有摩擦力的斜面也是理想化的概念一样。两者都是在试图作出归纳和找出底层规则。不过当Jerry Folder、Tom Beaverton和Marrow Garret的著作《语言心理学》（The Psychology of Language）问世时，情况就有了转变。深受理论语言学影响的学者认为这种分化是有益的，这可能与当时学界反对复杂性派生理论（The Derivational Theory of Complexity）有关。该理论认为越是复杂的事物，产生并处理对应表征的成本就越大。虽然这似乎是个合理的假说，但是大脑并不是这样运作的。因此，理论语言学和心理语言学在这一点上产生了深刻的分歧。但从长远来看，两个领域分化的益处并不多，因为彼此的研究不能再互相约束，从而让人们产生一种在单个领域内有许多进步和新见解的幻觉。因此，大家都希望理论语言学和心理语言学的联系在接下来的三十至四十年间会更紧密。

03.

主持人：您刚才说到，您认为当前生成语法研究对语言的实时计算过程欠缺关注。您能再展开讲讲吗？以及，您认为实验是否有助于我们对不同的句法理论进行选择呢？

 David Poeppel博士：以实验来实现对理论的比较是一种理想情况。设想有这样两种同样为句法提升或控制现象提供了有力解释的不同理论假设；而跨语言研究也不能对此提供进一步的证据。此时，理想情况是我们可以通过实验结果的差异来对两种假设的解释力进行比较。比如，在实验中我们希望发现不同的空语类假设会产生截然不同的结果，由此便可以利用实证研究来判断假设的正确性。然而这种理想的结果往往不易获得，这也是研究的困难所在。我们说，理论是研究者对语言加工过程做出的猜测。以“合并”（merge）这一当下热门的概念为例，如果没有证据证明该操作的话，这一概念的存在还具有合理性吗？“合并”概念的重要性在于它与语言输入-产出的界面相关。当我们试图理解我们是如何与其他人进行交流时，语言的输入-产出界面是解释我们交流能力的关键。正是由于我们可以获取交流中的信息并建立正确的表征才使得我们得以相互理解。

当然，我的意思并不是说所有研究者在作为句法学家、音系学家或语义学家的同时也应成为一名心理语言学家，这显然是不切实际的。研究者应各自在其领域进行学习并开展研究。但我们需意识到对语言的这些研究都属于认知科学，甚至是属于生物科学的一部分，即我们的研究目的都是试图理解人类的语言能力。而对于研究来说，理论证据的具体来源则并不那么重要。这些就是我们测试假设和某些理论而得出的证据。我们不会因为来源的不同而认为证据之间有高低之分。因此，问题的关键就在于语言学家和实验生物学家所构思的理论立场，能否借助实验结果对其进行验证。

04.

主持人：与脑科学研究相比，最简方案（Minimalist Program）这样的理论具有极强的抽象性。您认为这种抽象性是否会导致语言学理论和脑科学研究之间无法结合呢？

David Poeppel博士：我认为这取决于研究者自身。最简方案的研究者对实验的态度不尽相同，有人热衷于进行实验研究，有人则不然。我认为这种差异并不是由研究中的理论差异造成的，而是反映了研究者本身的倾向。当面对一个合理的假设时，研究者既可以选择对该假设本身进行深入的理论探究，也可以选择采取脑科学的视角，比如说结合人的记忆结构来思考该假设的可行性。因此我们可以说这取决于研究者的研究兴趣而非理论框架。而采取实验手段也说明了语言学与认知科学的其他学科一样，都需要充分利用学科发展所带来的优势。

05.

主持人：您之前也提到过计算（computation）和表征（representation）。这两个概念经常为心理语言学家所提及，他们也将计算视为认知科学中心理加工的基本活动。那您是如何看待这些所谓在人脑中进行的计算活动的呢？这对首次接触到这一领域的研究者来说可能是晦涩难懂的。

David Poeppel博士：对于在这一领域的学者来说，这也不是一个容易回答的问题。对于大脑，我们似乎知道得越多就了解得越少。包括我在内的很大一部分人都在基于心灵的计算理论（the Computational Theory of Mind） 进行研究，而具体的研究课题则各不相同，可能包括语言、记忆或注意力等等。这一理论认为人类思维是计算的一种而且是对心理表征所进行的计算。这就意味着我们首先必须承认在我们的心智与大脑中是有这种表征存在的。但我们直至目前也不能对此给出任何确切定义。我们做出了一个简单的基础假设，即思维中存在着一个表征库存。以音系学为例，在音系学中，一些学者将区别特征（distinctive features）视为基元，认为就神经与计算来说区别特征是不可再分的基础单位。这一观点则受到了其他一部分学者的反驳。从某些角度来说这个问题具有争议，我与后者看法相似，我认为音系的基元应为音节。原因是：音节是语言输入系统的基本单位。当我们对语言信息进行加工时，便是依赖于音节提取信息。同时，音节也是产出系统的基本单位。这也使得两个系统得以相互连接。因此，基于心理语言与神经语言实验研究，我认为音节是一个基础性概念，是表征库存的一部分。换句话说，音节组成了某种零件清单。

计算活动就是基于这样的零件清单进行的。以最简单的计算而言，我们有X和Y这样两个零件并要在它们之间建立起联系。这便是合并操作的前提：无序列的零件清单以及将零件组合起来的操作以最终实现信息的输入或产出。我们所关注的问题则是：我们如何将这两种零件进行组合？我们又是如何使这些零件成为能被识别的普遍表征？举例来说，形容词“black”和名词“table”的组合所得到的结果为名词“black table”，这是由中心语限制决定的。这一简单的过程便是我们试图所理解的一种计算活动。同样，细胞的增殖也可以被视为一种简单的计算。一些神经生物学研究在这方面取得了一定的进展，实现了对细胞增殖的描述；也有研究证明神经元的树突也可以进行某种异或计算（XOR computation）。

因此，我们做的便是寻找最为基础的零件。我们可以将此与电子工程师的工作进行类比。他们手头有芯片、电阻器、电容器之类的基础部件；而他们要做的便是利用这些部件搭建所需的最基础电路。这也说明了语言学、认知科学和神经生物学研究应有所区别：语言学家提出假设，研究计算活动的学者则需要对假设中的操作机制提供相应的解释，即说明人类的神经组织中是如何实现计算的。系统生物学就是以神经环路为研究对象，试图解释为何我们的神经组织中发生着这些可以以某一公式来进行描述的计算活动。这一领域的代表性成果为对昆虫导航系统的研究。通过研究，我们得以了解，在一只箭蚁体内存在大约50,000个细胞来进行这类计算活动。计算公式以某一数值和变量为输入信息且根据昼夜节律进行规律性变化，并最终产生输出结果。因此，昆虫便可知晓其行动方向。可见，这样看似复杂的生物现象是可以用计算公式进行描述的。接下来，我们要做的就是在大脑中寻找进行该计算活动的相关生理结构。由此，我们也可以看出不同学科研究之间的联系：语言学研究确定计算活动的基元，神经生物学研究便在此基础上解释计算活动的实现。

English Version

01. 

Host: You’ve written a lot about what neuroscience and linguistic theory can say to one another, but also, quite importantly, about what they can’t say to one another. Could you tell us a little more about how you see them sort of fitting together, and where they come upon?

Dr. David Poeppel: I would very much like to think that in the long game, they will increasingly fit together. But for now, the neurobiology and linguistics are still a little bit far apart. And so in part that’s an issue of just for practical reasons and one, and then there are sort of principal reasons. I’m very impressed with what linguistics have achieved over 2,000 or more. It’s not new, and it’s accelerated significantly in the last and, say, 60 or 70 years since it's become more closely aligned with the cognitive sciences, or computational methodology. But in general linguistics, the study of language is fascinating, fun and deep, and it is something about who we are and how we are organized. And so in the context of doing that research, linguistics has come up with just an enormous amount of insight at a very fine grain about many different phenomena across languages. So it’s about what it is that you know as a speaker of a different language.So how are different languages organized? How are they acquired? How are they processed? What is the inventory? And so regardless of what aspect of language you’re studying, there has been really a wonderful amount of data inside and cross-linguistically. So what that has resulted in is what I informally often call the parts list. It is the parts list of theprimitives, or theontology of the field for which there is really quite considerable evidence. That doesn’t mean it’s the final word, but it means that they’re extremely well motivated hypotheses about the infrastructure of linguistics. And the likewise, neuroscience, which is a little bit newer the field, has made unbelievable progress. We don’t know how the brain works, but we have a pretty good idea of what sort of the parts list is we’re getting better and better at well, the different cell types and the different types of physiological activities.

Now, if we believe both areas’ investigation, and we have no reason to doubt them. Assuming that language is something that the property of brains, then you assume that there’s some kind of very clear link between these domains. So that is a kind of principal relation between language and brains. And that’s where things are not so good. So my colleague and I have worried about this problem for many years, and we discuss it about how can we have a kind of better alignment across the disciplines? And one thing we’ve argued over the years is we can do better simply by attending to each other's fields more. So this is what we call the Granularity Mismatch Problem, which is simply that linguistics knows a lot, or the language sciences have very nuanced, detail-oriented insights when it addresses language, and when neurobiologists deal with language, the questions are pretty coarse, and so that mismatch and the grain size of the question can lead to all kinds of unsatisfactory stuffer, which just remains superficially correlational. And so that’s a problem from both directions. We have to have a kind of shared vocabulary and shared conceptual understanding, but it has to be at the right level of analysis, and that’s actually tricky. And I think step by step, it’s getting solved more and more by researchers. But then there is a more difficult problem, which is how is the mind, in this case, linguistics, related to the brain? And that’s kind of just an instance of the Mind-body Problems. We don’t have a particularly good solution. We assume that language is a property of brains, or we should probably stop. But it would be good to reflect very deeply on what such linking hypotheses need to look like. So if you’re committed to say the notion of morpheme, can you even point to the brain structure and function that would underwrite this kind of concept, or be related to it? And you simply have no idea. So we have some correlational notions about this, but in a deep explanatory sense. It means that when you study brain and language, like people, then your assumption is either you’re going to learn something about the brain, or you’re going to learn something about language, or ideally both.But what happens often is you learn about nothing. So this is what I have often called interdisciplinary cross sterilization. So instead of animating each other’s insights, it’s sort of a little bit of mediocre biology and a little bit of mediocre linguistics. They’re not particularly well connected, and the questions, how can we take deep next steps and understanding this part of how human minds and brains are organized, how they work, and so on. So the very long-winded answer to the question is complicated.

02. 

Host：If cutting up the disciplines of psycholinguistics and theoretical linguistics has been one's dealing with our application of our knowledge performance systems, and the others dealing with competence on knowledge. Do you think this way of dividing the fields is accurate or even helpful?

Dr. David Poeppel: It’s definitely helpful to think about many cuts and splits to illuminate the problem. And that particular distinction, which is really a result of thinking of the 1950s and 1960s in the generative context, is very interesting, and it makes very bold assumptions about idealization. Some people endorse that way of thinking about stuff. Some people find it just appalling, but in principle, it strikes me as pretty vanilla and uncontroversial in the sense that you’re just making some idealizations. So the notion of a competent system, isolated in an ideal speaker’s community, is trivially true that that’s an idealization, just like the notion of an inclined plane with no friction is an idealization. You’re just trying to find some kind of generalizations and principles that underly certain things.

But here's that kind of important twist when Jerry Folder, Tom Beaverton, Marrow Garret wrote this extremely influential book, Psychology of Language. People who are deeply influenced by linguistic theory, celebrated that split. And partly that had to do with rejections of what was then called the derivational theory of complexity. And so the more complicated something is in the context of a derivation, there is a representational thing, the more processing cost. It seems like a pretty reasonable theory. And that didn’t work out that way. And so there was a really deep split between and linguistics and psycholinguistics at that point. And I think that in the long game, has been actually a negative effect, because it then allowed both domains of research to go unconstrained from each other, which has led to important mythological advances and a lot of insights, because the uncoupling of that has been locally helpful. So one wishes for a tighter alignment over the last whatever of 30 or 40 years.

03. 

Host: Following on from that, you’ve expressed concern that a lot of the work in generative grammar has not really been concerned with real time computation in mind. Could you elaborate on that a little more? And do you think that experimental work could help us choose between competing theories of syntax or something like that?

Dr. David Poeppel: One would hope so. I mean, so you would like your fields to be integrated to the extent that, like, suppose you have two really good alternative theories of raising or control in syntax. And you have a very difficult time disentangling them with cross linguistic considerations. In the best of all possible worlds, you’d be able to kind of work those out, those alternative hypotheses as predictions that may have a clear processing consequence. I mean, you’d like to be able to say, “look, this kind of empty category, if I endorse it, then has certain kind of consequences. This other kind of empty category should have these other consequences, and I should be able to be sufficiently specific in my processing predictions”. So that would be a case where empirical research of a certain form could adjudicate between alternative hypotheses about the status of certain kinds of empty categories, or something like that. So that’s possible but it doesn't mean it’s easy because research is hard. Theories are always also a statement about processing. A popular notion now, like Merge, is there a notion like that, independent of the operation that makes it happen? And it has particular consequences at the interfaces. So you can build structural representation all day long, but at the end of the day, we’re having this conversation because something about the input-output interfaces that takes that and builds the right kind of representation to allow you to make sense of what I'm saying.

It sounds a little bit like every syntactician, or every phonologist or semanticist should also be de facto a psycholinguist, or something like this. I think that’s too strong. People are specialized to learn about certain generalizations to do scholarly work on it. But a mindfulness that one is part of the cognitive sciences, and even the biological sciences, would be good in the sense that our goals for all of us is presumably to figure out how the thing works. I don’t really care where my information comes from. My insight comes from your new work on French verb, inflection, or something like that. And mine happens to come from another imaging studying. It doesn’t matter. It’s evidence in light of some theory or hypothesis that we’re testing. And so there’s no hierarchy of evidence, right? Evidence is just evidence. So the question is that can linguists and experimental biologists formulate their theoretical stands in a way that permits experimental consequences.

04. 

Host: Should you worry that something like a Minimalist Program, which operates in an enormous level of abstraction, is quite far away from the brain science?

Dr. David Poeppel: No, it depends on the researcher. So there are people who work on this approach, who feel very near and dear to experimental research, and others who are less interested in it. So I think that’s it. That's a little bit independent of the particular theoretical commitments there. I endorse these, you know, fundamental assumptions, or something like that. It’s a little bit more about the researcher’s stance. It’s not intrinsically tied to the theoretical commitments you make. Assuming there is a plausible hypothesis, then you can either just proceed, or you can say, let me think about this and how would this actually work if I talk to someone who works on memory structures, how would you pull something like this out of a bag of words. So it's not framework dependent. It’s more sort of research interest dependent, and partly an acknowledgment that linguistics is not different from any other disciplines that’s part of the cognitive sciences, and therefore should draw on the advantages that that carries along with it.

05. 

Host: Carrying on from that, you’ve mentioned computation and representation, where psycholinguistics like to talk about representations and the fundamental motion in cognitive science of mental processing is computation. Could you speak about a little more about how you see the brain carrying out these computations? That can be quite difficult to picture for someone new to the field.

Dr. David Poeppel: Well, it’s even more difficult for someone who’s in the field, because the more I know, the less I know. So it’s better if I could just make a guess. Certainly a large fraction of those of us who work on this kind of stuff and adopt what has been called the computational theory of mind, and doesn’t matter whether you’re working on linguistics or memory or attention. So that means that thought is a species of computation, and that in turn means that it’s computation over representations. So it forces you to assume that there are such things as representations somewhere in our minds and brains. And we haven’t the vaguest idea what those are. We have the hypotheses, and that there are computations of certain forms. And we don’t know what they are. But the underlying assumption is simply that there is a kind of representational inventory. Let’s take again, the notion of distinctive feature in phonology is like a primitive. It’s an irreducible primitive in neural and computational terms. And some linguists would endorse that, others not. And similarly, I think that actually, for other reasons, that’s much more controversial linguistics, I happen to think and argue that the notion sticking with phonology of a syllable is actually fundamental, because it’s the fundamental unit for input process. That’s how you actually segregate information during processing. And it’s also one of the fundamental building blocks for the production system. So it’s how perception and production systems are actually coupled. So based on experimental research, both from psycholinguistics, but mostly neural linguistics, I would suggest that the notion of syllable has to be fundamental. It has to be one of the primitives. It means that’s something of the basic inventory representation.

Now you’re going to make a whole bunch of these, and that’s your parts list. And now you have to have a parts list of computation. So let’s take the simplest one. You want to make a sequence of things. You have an X and Y, and you have to chain them together. So that would be a prerequisite for something like merge. First, you have an unordered list of things, and you have to stick them together at some point, because they have to be spelt out, or read into your tape. So how do you take two items, an X, and Y, and put them in a chain? And even more tricky, how do you make them a common representation that has some kind of index? So I took an adjective and a noun, and they fit together, and then the whole thing becomes a species of noun, like “I’m sitting at a black table” where black table is a species of table, not a species of black because of headedness constraints on language speaking and so on. So, that’s the kind of computation you have to figure out. In other areas of neurobiology, there has been more progress on cell multiplication as a kind of computation. For instance, there's been interesting work on where we can characterize what cells do as multiplication. That’s very interesting. So you get some input here, you get some input here, and then you get some output from a group of cells, and it looks like what the output is a kind of multiplication. And there’s very recent evidence that dendrites on neurons can do something like an XOR computation.

So the way you want to think about this is you want to break it down into, like the simplest, most boring little things. Literally, it’s helpful to think about, like an engineer supposed you have to build it. So you have a bunch of chips, and you just have some resistors and capacitors, and your job is just, what’s the most elementary circuit that would give me what I need? So as I’ve argued a lot, cognitive scientists and linguists should take a more muscular stance. It shouldn’t be that linguistics reduces to neurobiology. On the contrary, linguists should confront computationalists and other scientists with particular, well-motivated and empirically, the solid insights, and say, “hey, we’re pretty confident that this has to work. Go figure out how it works. That’s your job”. And so the kind of computations you might think of are simple things like that, addition, subtraction, multiplication. And then the question is, well, how would you implement that kind of stuff in neural tissue? And that’s what people do in systems level neurobiology. They try to work on circuits and there are people who then literally try to figure out how would you actually have an equation in neural tissue. Because we know that you have to, actually. The most elegant work in this context is like insect navigation, and for insect navigation, people have a pretty clear idea what kind of equation has to be carried around in the 50,000 cells that a desert ant has. That equation takes a value and a variable as input and changes as a function of its diurnal rhythm, and then an output comes out. It tells you where to go. And so it’s pretty fancy stuff, but it's that level of math or computation. And then the trick is to find, well, does it match the kind of circuitry that brains would have and do naturally? So that’s the link from linguistics to spell it out into computational primitives, which in turn are kind of possibly linked to neurobiological implementation.

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