人物专栏 | David Poeppel博士访谈（下）
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17期，总第316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神经语言学的相关问题。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。David Poeppel博士简介可参考《理论语言学五道口站》“人物专栏”2024年第12期，总第311期。
访谈内容
06.
主持人：（接上期）这也刚好引出了我的下一个问题。神经学家似乎对人脑的计算模型持有争议。Fodor和Pylyshyn等学者认为图灵式计算模型（Turing style computation）才是对认知科学而言正确的抽象层次，而联结主义（connectionism）只是在执行层面上提供了一种可能的描述。最近，Randy Gallistel等学者则进一步发展了图灵式计算模型，并认为此后这一模型也会继续占据主导地位。请问您对此是怎么看的呢？对于神经语言学家来说，采用哪一种模型更有利于开展研究呢？
David Poeppel博士：在这个问题上，我想给出的建议与我个人的研究兴趣并不相同。对于初步接触这一领域的待业研究者来说，我会建议他们首先采用联结主义的理论框架，学会使用Python并深入探索神经网络模型，因为这种研究方法是比较实用且有效的。但在我看来，这一模型其实与人类心智、大脑的计算方式相去甚远。我个人的研究兴趣与Randy Gallistel十分相似，但我目前还处在寻找切入点的阶段，具体的研究方式也还不明晰。计算方面的疑点会更少一些，我们对这一块已经了解得比较清楚了。我估计再过二十年我们就能完全把握其机制。问题主要在于存储（storage）方面，具体而言，是在于人脑对语素等信息的存储位置及机制。目前我们对这一问题毫无头绪，而这也正是我们需要依赖于图灵式计算模型的地方。Randy Gallistel针对细胞间的存储和计算提出了大胆且有趣的设想，我们很快就会对这一设想的可验证性进行讨论。这个理论的耐人寻味之处就在于此：我们应该如何实践并验证这一猜想？怎样设计实验才能合理地得出结论？就现阶段而言，这无疑是一番苦战。
因此，我认为现在的研究者应该学会熟练使用Python并时刻关注当下研究发展的新动态。在我的实验室中，许多研究者都热衷于从计算机工程的视角来进行研究，因为这一方法不仅新颖而且高效。但问题是我们并不能确认这就是研究人类思维的最佳手段，这一点大家也都心知肚明。计算机的语音识别可以借助机器学习、深度学习来实现，但人类的语音识别却有着不同的机制。要研究人类的识别过程，我们需要更深入地研究耳蜗、听觉通路和皮层神经元等构造以及这些构造所允准的表征方式。
总的来说，我会推荐研究者们学会使用像Python这样的先进工具，但我个人感兴趣的是相对来说更为小众的领域。这种选择是大胆的，尽管这是我个人的偏好和追求，但我必须承认这样做风险很大，很有可能会失败。
07.
主持人：神经科学领域的发展总是伴随着风险。我记得您曾提到，人脑语言加工的经典模型在很多方面都存在不足，这也使得最近的研究进展都仅仅只是在修正其缺陷。您可以为我们详细介绍一下人脑语言加工的经典模型吗？它的不足又是什么呢？
David Poeppel博士：在每本语言学、心理学或神经学相关书籍中，都会有一章介绍语言系统的内容，其中便会有对于所谓“经典模型”的介绍，而这一模型通常是建立在单语的前提下的。我们会看到一张大脑左半球的照片，左半球前部有一个被称为布洛卡区（Broca’s area）的区域，而后部也有一个被称为韦尼克区（Wernicke’s area）的区域，这两个脑区由一束神经连接着。这一模型源于一位名为Broca的内科医生在1861年的一项发现，他的患者Leborgne先生由于脑损伤而产生行为异常，只能发出“Tan”这一种声音，这也使得他在之后的文献中被称为“Tan”。在他去世后，Broca对其尸体进行了解剖，发现他的大脑有一块特定区域受损。因此，Broca推断患者的异常行为与大脑结构受损相关。在那个学者们对脑功能的定位争论不休的时期，Broca的发现无疑是至关重要的。1865年，Broca发表了第二篇相关论文，文章进一步丰富了“偏侧化”（lateralization）这一概念。Broca的理论作为一种有趣的预测而深入人心，人们通常都会把脑损伤造成的诸如电报式言语（telegraphic speech）等语言不流利症状与与Broca所划定区域的损伤联系起来。1874年，德国神经学家Wernicke发现了大脑后部的另一个特定脑区，如果这个位于大脑颞叶的区域受损，尽管患者说话很流畅，不会出现电报式言语等症状，但所说的话语毫无意义。这对于当时的心理学和神经学研究来说是个非常了不起的发现。由此，我们发现了大脑有特定的脑区分别负责语言的产出和理解。这两个脑区通过神经纤维相互连接，各自又分别与发声系统和听觉系统相连。
几年之后，一位名为Lichtheim的神经学家综合了这些发现并提出语言加工还需要概念性的信息。于是，他的团队提出了一个模型，文献中通常称其为“房屋模型”（House Model）。在该模型中，运动图像（motor images）部分、声学图像（acoustic images）和概念（concepts）等部分呈单元式分布，使得整个模型的结构形似房屋。重点是，这个概念化的处理方式在临床医学上取得了巨大的成功。直至今天，由于神经受损而导致的行为异常仍然能够遵照这一19世纪的模型进行诊断，足见其含金量之高，而经典模型正是起源于此。除了其临床效用之外，这个模型也是第一个对大脑功能定位做出正式且有力的阐述的模型，它的提出极大地为神经科学的发展注入了动力。
那么这个模型有什么问题呢？从生物学角度来看，其问题就在于它对脑区的划分不够具体。通过对这些“脑区”的进一步了解，我们发现它们其实由更多不同的区域组成。实际上，大脑不存在单一的“布洛卡区”，而是可以进一步划分为10~20个更小的部分。也就是说，我们不能将这一脑区与某个单一的功能对应起来，而应认识到其中大脑活动的多样性以及各个子区域之间联结的复杂性。这就是我们从生物学角度得出的不同结论。根据各项研究中得到的证据，我们自然能够对细胞及其功能有更为细致的了解。那么从语言学的角度来看，“产出-理解”的二元化划分是否正确呢？对应到语言的“输入-输出”系统，这个划分看似合理，但实则未能捕捉到语言系统的复杂性——它对于语言系统中的表征与计算没有做出任何表述。就当前研究来看，仍有大量问题亟待解决：学会一门语言的时候，我们究竟掌握了什么内容？词汇是如何储存和组合的？准确的语言理解和语言产出意味着什么？基于这些问题，语言学或心理语言学又该建构怎样的模型？经典模型对于这些问题显然一筹莫展。
终于，在70年代早期，研究迈入了一个新的发展阶段。通过对患者数据进行更为详细的分析，人们意识到大脑功能的分布并不是前部负责产出、后部负责理解这么简单。因此，又有学者提出应就语法与语义两个层次对大脑结构进行划分：语法操作更多地与大脑前部皮层相联系；而语义操作则更多地发生在后部皮层。这一划分标志着认知心理学与语言学的交互，但就我们对该领域的了解，这种划分仍是极为不准确的。在过去20年中，包括我在内的一批研究者一直在探索语言加工机制的模型，随着研究的深入，语言学问题也开始与现代神经科学研究接轨。因此这个课题也成为了神经功能解剖学领域中许多学者的研究方向之一。
08.
主持人：关于您提到的功能定位，许多语言学家和心理语言学家都强调思维的模块化。正如您所说，大脑各区域并不能简单地与语言层次一一对应。如果语言功能在大脑中是分布式的，那是否意味着“语言官能”（the faculty of language）这一概念本身就是不准确的呢？
David Poeppel博士：并不是“不准确”，而是我认为这个观点有些缺乏想象力。语言系统对于目前的研究而言过于复杂，牵涉到太多方面，因此我们无法把握其机制。但如果换做是其他系统，我们会做出模块化的假设吗？比如说我们更为熟悉的呼吸、心跳或体温调节，这些功能对应的结构大部分都位于脑干，而这些结构正是一群组织严密的神经核（nuclei）。这些神经核是高度模块化的，各自都实现着非常特定的功能。大脑组织的这种高度模块化有众多证据支持，在其他哺乳动物的大脑中同样可见。所以模块化假设并非不切实际。同样，包括肝脏在内的其他器官的工作原理也是如此。因此，我们可以推断像语言或是心智这种涉及诸多方面的系统并非由某个单一的结构承担。语言系统最突出的特点在于它可以与多种官能无缝对接，比如我们现在交谈所使用的听觉、阅读或手语交流时所使用的视觉以及阅读盲文时所依赖的触觉。这说明语言的核心表征系统必须具备一种可以识别多感觉模态信息的输入-输出结构。视觉信息从枕叶传导而来，听觉信息由颞叶传导而来，触觉信息从顶叶传导而来……这些不同信息来源所生成的结构，与负责加工级联关系、结构构建、远距离关系等离散无限性（discrete infinity）信息的系统相连。由于信息来源的多样性，由这个系统输出的数据在结构上也必须相当灵活。同时，输出的数据在内容上也要保证连贯，因为我们可以通过不同的方式，无论是文字、手势还是盲文，来传达相同的信息。这就对语言模型的建构提出了限制条件。这种限制在整个神经系统中是普遍存在的，因此这些来自于不同感觉-运动系统（Sensory-Motor Systems）的信息是不太可能汇聚于同一处的。这些基本运算更有可能分布在大脑的不同位置。因此，模块化这个概念是功能性的，而非解剖学意义上的。Fodor也针对这个问题提出了“认知可渗透性”（cognitive penetrability）的概念。
所以，我们为特定信息所构建的表征会进行何种程度的浅表性加工呢？显然，月相或者祖母的生日这种外在因素是不会影响表征加工的。该系统就是以相对明确的内容为对象进行加工并实现产出的。所以，语言功能在大脑中是否是分布式的与我们所说的“功能分离”（functional isolation）抑或是Fodor理论中的“认知可渗透性”、“信息封闭性”（informational encapsulation）等问题无关。在我看来，大家对这个问题的广泛关注有些舍本逐末了。大脑各功能区的重叠与否并不是问题的关键。更何况，如果Gallistel关于大量细胞内计算的观点是正确的，我们目前所知的一切都可能会被颠覆。这样一来，神经语言学也将成为分子生物学一个分支。
09.
主持人：Fodor的九条准则之一指出，不同的认知功能由特定的神经结构行使。那将其解读为由“单独的”的神经结构行使就是不准确的。
David Poeppel博士：是的，我认为这种解读并不是Fodor的本意。这个“特定”的意思是：特定表征结构及其相应的计算实现于特定的神经回路，简言之就是加工某种特定的输入而不加工另一种。但问题是这个假设正确吗？这就是当前研究要解决的问题。假设某一操作作用于两个对象，将它们结合并贴上标签，我们简称其为合并。这种合并操作究竟是普遍存在于语言系统中，还是只作用于计算的终端元素呢？要回答这个问题，重要的不是具体的某对A和B，而是A和B两个位置允许放入的对象类型。实际上，A和B的最终身份还是取决于操作所调用的对象清单。清单里的对象也许与语言无关，比如两个音符之间可能也碰巧需要这种“结合+贴标签”的操作。这样一来，问题就变得有些复杂了。我们需要在理论中明确地划定这种操作的使用范围。在我看来，我们在清单中所做的大量规定有些过于精细了。
计算这一活动具有极高的普遍性，在针对运动控制系统或其他系统的研究中，甚至在动物实验中，我们都能找到相关的证据，因此我有理由相信，在未来的20年里，我们对计算的研究会取得更大的进展。而那份为我们提供词汇和百科知识的清单，或者说“心理词典”（the mental lexicon），才是更棘手的问题。清单中的大量特殊化规定也是难题之一。当然，这只是我个人的预测，不保证准确。
10.
主持人：在过去的六十年中，语言学研究见证了许多理论的诞生。除了从经典模型到当前研究框架的转变之外，您认为心理语言学和神经语言学领域中最重大的发现是什么呢？
David Poeppel博士：这是一个很好的问题。语言的生物学研究是否有值得一书的发现呢？答案是肯定的。但更重要的是：这些研究有没有增进我们对语言本身或是对其他系统的了解？换言之，心理语言学和神经语言学是否为我们揭示了语言系统的本质？抑或是我们只是借用了语言学这样一个理论清晰且形式规范的体系来解释大脑所进行的复杂计算？我认为答案是后者。我们对大脑结构已经有了相当多的了解。比如，Greg Hickok和我几年前提出的认知“双通路模型”（dual-stream model）在认知加工和神经结构的假设中被广泛采用。我们之所以投身于这个研究方向，是因为它与大脑中其他系统的组织方式极为相似，比如视觉系统。我们希望找到一个更普适的原则来统一描述这些不同的系统。当面对一个宽泛的研究问题时，我们不妨将其切分为一个个小问题，就像把程序分割为一个个子程序一样，再利用一切可用的、有用的模型对其进行解释。这也是我们采用“双通路”式推论的原因之一。现在这个模型有了很多类似的变体，Hickok和我也专注于为这些模型作初步的筛分（fractionation）。这项工作对神经科学研究来说十分有用，但对语言学研究可能并无帮助。
但另一方面，如果是从功能解剖学的角度解释执行层面的现象时，这个模型还是十分有用的。这方面的研究不乏真知灼见，但其中很多观点在语言学家们看来却似乎不言自明。可是，即便像“成分性”（constituency）这样简单的概念，实际上也远非不言自明。人们可能会认为，成分的构成不值一提，无非就是X向Y合并后获得X-Y的标签而已。但这其中却暗藏玄机：通过上述表述，我们能够知道表征具备结构性，语言加工也并非是像串珠这样简单的线性加工，而是具有结构依存性（structure dependence）和成分性的特点。对于语言学家来说，这只不过是入门知识；但对于心理学家、认知科学家或神经学家来说，这却是个重要概念，它让我们能够有意识地去审视那些已经习以为常的术语。文献里对“成分性”这样的简单概念进行的研究也少得出奇，很难想象这样一个古老的术语仅仅只受到了二三十项实验的关注。由此可以看出，认知科学的洞见是可以为其他领域的研究带来启发的。
另外，我们在语音与音系领域也颇有所获。我们发现，像“区别性特征”（distinctive feature）这样的概念对于任何语言的语音和音系研究来说都至关重要，但像“音节”这样的其他层级或表征却存在争议。对于实验研究而言，这是一个需要重点关注的问题。如果能确定区别性特征就是语音加工中最基本的信息组块单位，其重要性自然不言而喻，毕竟这是我们理论的立足之本，那么我此后的工作中心就是全身心与音节打好交道，将其作为核心概念来研究，即便它只是个大一级的单位而已。显然，在研究过程中，音段音系学（segmental phonology）的知识是不可或缺的。那么这些单位又是怎样构成我们识别、产出和存储的语音现象的呢？例如，最小对比对（minimal pair）的定义是什么？什么样的成分能充当最小对比对？这些都是需要一一论证的。
11.
主持人：您认为未来最有前途的研究领域是什么？目前您觉得哪些研究特别令人兴奋和感兴趣？
David Poeppel博士：从我的研究经历来说，我深受电生理学方法的影响。语言是感知的产物，而且这一过程非常快，所以在研究语言加工与大脑结构的关系时，我们需要用高时间分辨率的技术以发现和描述其中潜在的关联。因此，我对脑电图、脑磁图、脑皮质电图和其他一些脑成像技术的发展非常着迷。目前，我了解到有一批非常重要的研究利用了功能性神经成像技术来进行定位推理，这些研究很有趣也很关键。对我来说，电生理技术并不能直接解决认知科学所关注的问题，例如表征的定义、基本运算的定义及其对象等等。要回答这些问题，我们需要能在时间轴上呈现动态变化的模型，因此神经动力学非常重要。总而言之，我对这个领域和相关技术都很感兴趣，而且借助这个领域日新月异的技术发展，我们对细节的探索也就越来越深入。这些技术中的一部分也与数据分析有关，比如解码技术以及一些基于机器学习的技术。如果你有一个不错的研究问题，借助这些技术便可以从数据中获知更多的相关信息。心理语言学和神经语言学中很多实验的条件都受到了严格控制，比方说，我们通过呈现最小对比对的方式获得了一批零散的数据，但最终需要解释的却是我们在实际对话时大脑的工作原理。所幸目前已经有了越来越多的技术能让我们在更自然的环境中研究语言现象，但这并不意味着我们要放弃对实验条件的控制，而是要将其与实验设计更好地结合起来。现在我们已经能在被试听故事或者做其他事情时记录其丰富的语言加工数据，而这些才是我们真正感兴趣的。所以分辨最小对比对不是重点，重点在于我们在处理这些事情时大脑是如何工作的，这正是实时研究的乐趣所在。机器学习和人工智能领域的一些技术对分析实时数据非常有用，但我们需要注意这并不意味着这些技术能够充当思维和大脑的模型。不使用最先进的技术处理数据并不可取，但请记住这只是为了数据分析，只是为了理解和解决问题。放眼未来，电生理技术将会越来越精细，在最细微的尺度上，我们已经能在临床实验中记录人脑单个细胞的活动。而在最粗略的尺度上，我们也有像fMRI这种能看到整个大脑的成像技术，它具备毫米级别的高空间分辨率和较低的时间分辨率。在这个尺度上，我们通常希望获取一些数据来探究大脑信息加工的网状结构，但更重要的是大脑是如何通过这个网络连接起不同功能的。我们也可以从最细微的尺度开始，先观察单个细胞的活动。但通过研究860亿个脑细胞之一的活动来管窥整个大脑，视角未免太过狭隘。不过如果能弄清楚有多少信息被存储或是加入了从细胞到回路再到系统的整个传输过程，那就很让人振奋了。这样动机成熟、理论可行、计算明确的课题是很适合做量化研究的。当然，建模所需要的数据一般都来源于病重的患者，但其实很多病人都愿意参与研究，因为他们在医院里也并没有其他事情可做，而且许多患者也知道他们的参与对研究的开展而言意义重大。不过，我还是希望能开辟更多新技术，比如，我认为用创新性手段研究动物大脑建模会是一个很有前景的方向，但这需要极高的理论技巧和相当谨慎的态度支撑。因此，我不主张用动物模型来研究诸如约束原则C这样明显与之不匹配的问题。另一方面，如果我们能在计算层面上对一个问题做普遍性描述，并将其拆解为一个个必要的步骤，那么我们将其代入动物模型中研究又何尝不可呢？当然，这并不是说鸟鸣具备人类句法的特点。这是一个执行层面的问题，我们关注的是一个加工特定信息（比如单词识别）的脑回路及其后续所有的传导。我们可以把已有的研究对象分解成一些具有普遍性的子步骤，然后在其他生物身上观察类似现象，因为这些步骤很可能在生物保守的进化过程中保留下来。所以，尽管我们与其他生物有不同的加工方式和大脑架构，但从结构和功能相适应的层面上，我们仍能通过观察动物来取得一些进展，就像弄明白螺母和螺栓的铰接一样。在我看来，构建动物大脑模型来辅助研究是无可厚非的。
English Version
06.
Host: That leads perfectly onto my next question. So people like Fodor and Pylyshyn suggested that one way of looking at neural architecture was, if connectionism is true, it would be true at the sort of implementational level, but Turing style computation would be the correct level of abstraction for cognitive science. But more recently, people like Randy Gallistel still have taken that further and said, exactly like what you mentioned, Turing style computation all the way down. If you had a stance on that, and whether you see one tack is more productive for neurolinguists?
Dr. David Poeppel: I have the personal taste, and I have a recommendation. If you are starting out and you still need a job, learn Python, get into your deep neural nets and work with a connectionist framework, because it’s sort of practical, and it works. Whether it has any relationship to what human minds and brains do, I think it’s remote. And my own interest is actually very much in the stuff that Randy is doing and the most interesting problem I’m trying to figure out is how to even approach it at all. It’s not clear how to do it. It’s less about the computation part, which I think we have a grip on. In fact, as I suspect in 20 years, we have a really good grip on it. It’s about storage, and basically, where do you store? And how do you store the inventory? If you have a bag of morphemes, how is that written down? And there you might need Turing style computation. And how to do that, and where in a body to do that is completely unclear. I mean, Randy Gallistel has a very bold suggestion about intercellular storage and computation is a cool idea. We’re going to talk about that soon to see if we can even figure out how to ask the question the right way. It’s another one of those examples where the theory makes them interesting. It’s an interesting story. How do you actualize? How do you then test? If it were so, what kind of experiment would lead you to that conclusion plausibly. And that’s extremely difficult in this context.
But I think that the young generation should be aggressive Python users and get your head around and see what everybody’s doing. Everybody in my lab is wanting to do that because it’s cool and it’s fun. It’s sort of elegant that it works from an engineering point of view. It’s a different question, as we all acknowledge, whether that’s actually something that is an approach to studying the human mind brain. For instance, the machine learning bases, deeply training things to a really, pretty impressive job at recognition right now, speech recognition. That’s super cool. I’m totally into it. But that’s different than studying how people do it. I have to account for, like, what’s a cochlea? What’s the auditory pathway? What do cortical neurons care about? I mean, that stuff, I have to worry about what representations to such an architectural license, and so on.
So my gut level recommendation is to learn the modern stuff, and my intellectual passion goes for the really weird stuff. I think it’s cool to be bold, and I want to pursue that kind of thing, but it’s extremely risky and likely to fail.
07.
Host: I suppose, in terms of taking risks and the progress of neurosciences, you’ve mentioned that the classical model has been incorrect along too many lines for recent developments to merely count as revision. Could you briefly spell out what the classical model of language in the brain is, and what’s wrong with it?
Dr. David Poeppel: In every linguistics book, or psychology book, or neurology book, there is a kind of chapter on language, usually there’s one language in the brain, and there is the more or less formally called the classical model. There is a picture of a left hemisphere of the brain. And there is a region in the front of the brain, which is typically called Broca’s area, and in the back of the brain just typically called Wernicke’s area. And then there’s a string of cables between the two. So this is a model that goes back to 1861 with the discovery that a particular brain injury led to performance deficits in a patient of Broca’s, Mr. Leborgne. So this guy could only say one thing “Tan”. He was subsequently called “Tan” in the literature. And when he died and Broca performed an autopsy, he noticed that there was a particular chunk of brain compromised. And he made the kind of totally reasonable inference there is a relationship between the behavioral deficit, and the brain structure. And this, by the way, in its own historical context, is very important, because people were really arguing about the notion of functional localization a lot. So this is a very important insight by Broca. Then he wrote a second paper a few years later in 1865, where he basically kind of enriched the notion of lateralization. So that really stuck, because it was an interesting prediction. So in fact, if you have a lesion and you don’t speak fluently, like telegraphic speeches, it is often correlated with that kind of lesion. Some years later, in 1874, Wernicke, a German neurologist, found the one in the back. So if you have a lesion in the back, the temporal lobe and then you can still speak, it’s not disfluent in the sense of telegraphic, but it didn’t seem to make any sense. So that was a very cool idea. In the psychology and neurology of its time, actually quite brilliant. So there’s a part of the brain that’s responsible for output, for production, there’s a part of the brain that’s responsible for comprehension. Then you need a cable between the two, because you have a comprehension area and a production area, and then those are linked to your mouth, in your ear.
So then a few years later, the neurologist, Lichtheim, integrated all this information and said we also need that sort of conceptual information. So they came up with what in the literature called the House model. So there, you know, the motor images, acoustic images, concepts, and then the whole thing, so that the kind of boxology of the model looks like a house. Now, what’s really important to know is that this particular way of conceptualizing processing was really successful, clinically, as to say, if you have some kind of behavioral deficit as a consequence of a neural injury, you will get diagnosed on the basis of that 19th century model today as well. So that’s sort of what’s right with this. That’s its historical origin. Other than its clinical utility, it’s animated the field neuroscience more broadly, by the way, because it was the first, really formal and sort of robust statement about functional localization in the brain.
And so what’s wrong with it? Well, what’s wrong with it is that it’s sort of an underspecified with respect to the biology. We know a lot about these brain areas, and they turn out to be comprised of multiple brain areas. There’s no Broca’s area. There’s like a Broca’s region that’s made up of between 10 and 20 different parts. So you can’t just say, you know, this is the function of that. There’s lots of stuff going on there. The connectivity is much more sophisticated. So the biology is different. We know more about cells and what they’re doing. So that’s not that surprising, simply accumulation of evidence. We know way more about how stuff works. But of course, from a linguistic point of view, you have to decide, well, is that the right cut? Right? So is the production versus comprehension the right way to cut nature at its seams? And there is something to be said for it in terms of the input, output systems. But it really doesn’t capture the richness of the system, right? So there’s nothing about the representations or the computations that are part of the language system. So what is it that you actually know when you know a language? How does the vocabulary get stored? How does it get put together? What does it mean to have successful comprehension or production? What are the underlying potential models, linguistically, psycholinguistically? So, of course, none of that was available.
So its next development was in the early 1970s. In the early 1970s, it became clear that, you know, very careful analysis of patient data, it couldn’t be simply production in the front, comprehension in the back. So a subsequent claim was, well, maybe the cut is actually syntax versus semantics. This more syntactic type processing, or operations are more associated with frontal cortex. In semantic type, things are more in the posterior cortex. So that was an interesting, sort of next step, sort of at the level of cognitive psychology, intersects with linguistics. But that too, of course, is dramatically underspecified with the respect to how we study the domain. Right? So for the last, let’s say, 20 years, a bunch of us have been trying to figure out, well, what would it look like? What would an organization actually look like? That’s a little bit more closely. That allows you to ask questions about linguistics, but also is plausibly embedded in modern neuroscience. So that’s sort of one of the lines of functional neural anatomic research that many people are doing.
08.
Host: So, something you touched upon there was sort of functional localization. And linguists, psycholinguists have been keen to stress modularity of mind. Because, as you said, we can’t find one bit of the brain that corresponds to doing the language stuff. Do you see that as problematic at all for the notion of faculty of language, that if it’s distributed across various parts of the brain?
Dr. David Poeppel: No, I think I’ve no bearing on it. I mean, think that’s a poverty of the imagination argument. I mean, stuff is complicated. We have no idea how it could work. It’s just too complicated at the moment, right? Because there’s too many parts involved. But suppose you were talking about a different part of the mind and brain. Would you make the same arguments? I suppose it’s something that’s really near and dear to you, something like, um, breathing, right? Or your heartbeat, or the regulation of your temperature. So all the structures that are largely in the brain stem, uh, you will find it not surprising that those are extremely, highly organized nuclei that are extremely modular and do very, very particular kind of things that you need to function. So there are plenty of examples in brains, including in mammalian brains, of highly, highly organized tissue that’s super modular, right? So it’s not like “oh my God, what a bad idea”. I mean, similarly about the liver or something. You know, liver does lots of cool stuff. So psychological or mental systems, or organs like language, are unlikely to be a single spot because they draw on lots of different stuff. So, take one of the most interesting things about the linguistic system is that it can interface seamlessly with sound, like we’re using now, or vision, like when you’re reading, or vision when you’re signing, or touch when you’re reading braille. So whatever happens in the so called the central representational systems has to be able to take as input and output data structures, stuff from any sensory modality. So it means stuff coming from the back of the head, for vision, from temporal lobe for hearing, from up here, for touch, these very different pathways are generating structures that then align with the system that does, let’s say, concatenation, structure, building, long distance relations, everything that’s part of, let’s say, discreet infinity. And so those things are, since everything is coming from somewhere else, it has to be relatively flexible in terms of what data structures it outputs. At the same time, it has to be coherent, because we can have the same conversation with whether we’re reading, signing, brailing and so on. So it places extremely interesting constraints on how you want to do it. And because those constraints are sort of all over the place, it means that it’s unlikely to be just everything converges to one spot from all the different Sensory-Motor Systems. So it’s a bit more likely that the kinds of elementary operations sit in different places in the head. The notion of modularity is a functional one, not an anatomic one. This is the notion what Fodor called cognitive penetrability.
So to what extent do you build a representation that really has to be shallowly processed for a certain kind of thing? And it doesn’t care about the phase of the moon, or, like, you know, your grandmother’s birthday. It just does what it does because it works on relatively specific things and generates outputs. So the question whether that’s localized or distributed doesn’t speak to that issue of functional isolation or penetrability or informational encapsulation, which is another way to talk about it in Fodorian terms. So for me, that issue that people get pretty exercised about is a bit of a red herring. Showing spatial overlap or spatial non-overlap are sort of coarse hints, but they’re not even closely good enough for this. Also, if Gallistel is right, and there’s a lot of intracellular computation, then we don’t know anything. I mean, then we have to start from scratch, more or less. And then neural linguistics also ultimately do become a species of molecular biology.
09.
Host: So, to go back to Fodorian terms, one of the nine criteria is dedicated neural architecture. To read that as a localized neural architecture would be a misreading.
Dr. David Poeppel: Yeah, I don’t think it was what Fodor meant either. It’s about, is the circuitry that supports particular kind of representational infrastructure, and the computations over those representations dedicated in the sense of, let’s say… “this circuit works on this kind of input, but not this other kind of input”. Now, the question is, is that right? That’s where things in active research are going. So suppose you have an operation that takes two items, puts them together, and then labels them. So some form of what you might call merge in shorthand.So is that operation, which could be very generic, specific to the language system, or does the specificity merely arise in virtue of the terminal elements that get inserted into the computation? So if you have, you know, an A and a B kind of thing, does it matter? Uh. So maybe all that matters is that the right thing fits under this A slot, and the right thing fits under the B slot. But what really determines what it is iactually the part of the inventory that comes up. So maybe it’s a musical note and a musical note, in which case it’s nothing to do with language. It just happens to take that kind of operation, namely, you take two things, you stick them together, you put a hat on them, right? So this is where things get a little bit more tricky. You have to make commitments to wherein lies it. So I think that the huge amount of specialization in what we have stored as the inventory, that’s very unique and weird.
For the next 20 years, I’d say we can make good progress on computation, because it’s simply because it’s generic enough that we might find examples in animal perhaps in other experiments, in other systems, perhaps would say motor control or something like that. That the inventory, let’s say, the vocabulary, the mental lexicon, or whatever you want to call that, that has served lexical and encyclopedic knowledge that you point to. That is a hard problem. And that, and a lot of specificity, lies in that. That’s my prediction. I’m willing to be wrong.
10.
Host: So besides the move from the classical model, rather than looking forward to it, look back, in the last 60 years or so, where as you mentioned, a lot of well linguistic theories we know started. What do you think the most important discoveries in the psycholinguistic and neural linguistic research are?
Dr. David Poeppel: It’s a very good question. So, are there insights and, let’s say biology of language that have been particularly noteworthy? There have been. But the question is, have they been noteworthy for our understanding of language, or have they been noteworthy for our understanding of some other system, right? So is there something that we’ve learned about how the language system works by doing, let’s say, psycholinguistics or neural linguistics, or have we simply used linguistics, because it’s a very well delineated and formalized domain to learn something about how brains do complex computations. So the latter, I think, is for sure true. And I think we’ve learned a lot about how brains are organized. For example, Greg Hickok and I carried on about some years ago, what we called the dual-stream model of recognition, which has become quite popular in terms of the sort of hypothesis about processing and the neural architecture. And the reason we were moved by that direction of research is because it leans very closely on how other systems are organized in the brain, like the visual system for instance. And it seems so we’re trying to see are there underlying broader principles that are simply recycled. Do you take some broad domain and you basically carve up the problem into sub problems, like this subroutine and that subroutine, and you capitalize on whatever structures that are available, that are really good for that? And so that was one reason to do this kind of dual stream reasoning. Now, there’s many species of models like that. Hickok and I spend all the time working at the initial fractionation of them. That’s turned out to be pretty useful for neuroscience. I don’t think it’s been useful for linguistics.
But on the other hand, it is very useful if you’re trying to build functional anatomic insights into how you would implement this kind of stuff. So at the implementation level. I think there have been some extremely interesting insights about that. But They are sort of confirmatory for linguists. For example the notion of constituency. I mean, that’s not an innocent notion. People are like, what’s the big deal? You know, you take an X in a Y and you make an X-Y hat. But actually it is a big deal because it shows you something about representational structure and shows you that, for instance, language processing is not a string of beads, that it has actually notions of structure dependence and constituency. Now, again, for a linguist, you’re like, yeah, obviously that’s like lesson one, right? But for a psychologist or cognitive scientist or neuroscientist, that’s actually important, because it means it allows us to actually probe certain things that you take for granted in the cognitive sciences. So something as simple as constituency. There have been, you know, surprisingly few studies on that. You think it’s an old concept, but maybe, you know, I don’t know, 20, 30 experiments. But I think that’s a very important insight, where cognitive sciences tell us what may be looking for in other fields.
I think we’ve also learned a fair amount about phonetics and phonology. I’m pretty convinced that a notion, like distinctive feature, is really, you know, a critical notion for any kind of phonetics, and phonology in particular, there are other levels or representations that are very contentious, like the notion of a syllable. And there, I think, experimental work, and say, look, that those are actually important insights. And if it’s true that that’s sort of the chunk size that’s foundational for processing, well, that’s important to know, because your theories are accountable for that, right? So I’m going to push harder, I have pushed hard on that, to be very sensitive to syllabicity as a core notion for what you have to achieve, even though it’s, of course, a bigger unit, right? So you need segmental phonology. It’s not a substitute. But the question is, how are these things organized to get the range of phenomena that you actually get in recognition and production and storage. Like, what’s the notion of a minimal pair even like? What’s it defined over? And that’s not innocent.
11.
Host: What do you think the most promising areas of research going forward are? And what work do you find particularly exciting and interesting at the moment?
Dr. David Poeppel: As is clear from my own scientific biography, I’m really swayed by electrophysiological approaches, and that has to do with the fact that language is in general perception products, those are fast processes. And to really get your, as it were, wrap your head around these things, you want techniques that have a temporal resolution appropriate to the sort of underlying boxology that you’re trying to unravel and unveil and describe, and so on. So I’m most fascinated by advances in techniques of electroencephalography, magnetoencephalography, electrocorticography and some other images. I think there’s really important work being done using functional neural imaging that's about the localization-based reasoning. It’s fascinating, and a lot of it is important. For me, it’s less germane to addressing the questions that are closer to the cognitive science concerns, which is what’s the representational, what are the primitives and what are the primitive operations. And to get to that, you need models that are sort of dynamic in time. And so neural dynamics seems to me very important. And in general, I’m interested in that. There’s a lot of technical developments in that area. That allows you to go further and further into details. Partly, that has to do with data analysis techniques, like decoding techniques, or machine learning based techniques, so if you have good questions, you can get more and more into specific stuff. So for example, much of what we do in psycholinguistics and neurolinguistics, is extremely controlled experimentation. So let’s say you present a bunch of minimal pairs with something and so you get really granular. But of course, in the end, one of the things, what’s like our explanandum, is to try to figure out what we’re doing right now, we’re just having a conversation, and how does that actually work? So increasingly, there are a technique available to begin to study linguistic phenomena in a more naturalistic context. That’s pretty cool, because that doesn’t mean you’re giving up on the notion of control, but sort of paired with experimentation. But it’s very interesting how much sort of rich data about language processing per se you can get out of just now, data that are recorded while someone, say, listening to stories or something. And that’s nice, because in the end, that’s what we’re interested in. So you’re not interested in figuring out these minimal pairs, but you have to figure out how it works while you’re doing it. That’s the fun in real time. And so some of the techniques that people have developed in machine learning and AI are extremely useful for data analysis. It does not follow that using those techniques is a model for minds and brains. So I think you want to be careful about it. So you would be a fool not to use the most cutting-edge techniques to wrangle with your weird data it. But that’s for data analysis. That’s trying to understand how you actually parse the problem. So look, one thing that’s going to happen, for instance, is electrophysiological techniques will get finer and finer. At the other end of the spectrum, like single cell recordings in human brains, obviously in clinical contexts. So now we have the sort of at the coarsest grain, something like fMRI which sees the whole brain, a very high spatial resolution, better than the millimeter, if you want, at very low temporal resolution. And at the highest level, you see something, and you want that because brains are sort of network principles. It matters how things are wired up over the differences. And at the finest grain, you can start to look at single cells. So now you have a very narrow view. Your lens is just a single cell of 86 billion, and you can try to figure out what’s going on there. And the fun and the excitement will be to figure out how much information is stored or carried from cell to circuit to system. And hopefully motivated, kind of theoretically well motivated and computationally very explicit, so you can kind of quantify and study these things. Modulo that, of course, these kinds of data always come from patients that are very ill. It’s wonderful that a lot of patients are willing to participate in research, because they have to be there anyway. And it’s it can be very meaningful to people to participate in this kind of research. But we want increasing techniques to do in other ways. And so the next another way to do stuff that’s I think, promising, but requires great theoretical finesse and care, is to look in a more creative way at animal models. So I’m not going to look at an animal model to study, say, principal C. I mean, that doesn’t mean like the right way to break the problem down. The other hand, if there is a problem, I can characterize at a computational level, in a pretty generic way. To parse the problem apart and say this is the kind of thing that must happen, then what’s wrong with studying that in an animal model in which you would learn about that? So this is now going to the implementational level of analysis. Again, it doesn’t mean that you think bird songs are like syntax. This is an implementational level question, saying that we need this kind of circuit to point to all the different things that a recognized word connects with whatever it is. So, you can decompose the problems into kind of generic subroutines that you could then look at in other creatures, because evolution is conservative, and the stuff is retained. So we do stuff differently. Our brain’s kind of differently organized. But at the level of the brain structure and function, we can make some progress. Just some, as it were, nuts and bolts, how’s stuff wired together. There’s nothing wrong with doing models as I’m concerned.
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 董泽扬
排版：安安
审校：吴伟韬 时仲
英文编审负责人：黄静雯 安安