English

The relationship between computer science and linguistics begins with a shared object — formal grammar — and a shared founder, Noam Chomsky, whose work in the 1950s belongs equally to both fields. Chomsky was trying to understand human language, and in the course of it he developed a mathematical theory of grammars and the languages they generate that turned out to be foundational for computer science: the Chomsky hierarchy classifies formal languages by the computational power needed to recognize them, and it underlies the theory of parsing, the design of programming languages, and the construction of compilers. The same mathematics that Chomsky developed to characterize human language characterizes the languages that computers process, and this shared foundation is the deepest root of the relationship between the fields.

The relationship has since followed a remarkable trajectory, driven by the changing methods of computational linguistics — the attempt to get computers to process human language. It began with the Chomskyan, rule-based approach: if language is governed by formal grammar, then processing it is a matter of implementing the grammar. It shifted, in the 1990s, to statistical methods that largely abandoned explicit grammar in favor of learning patterns from data. And it has culminated, with large language models, in systems that process language with extraordinary facility while containing nothing that looks like the grammars linguistics describes — which has produced a genuine crisis, or at least a genuine confrontation, between the engineering success of the neural approach and the theoretical claims of linguistics about how language works. The boundary between the fields is now a site of active and consequential dispute about what language is and what these systems reveal about it.

This section examines the relationship through three lenses: the shared foundation of formal language theory, the methodological evolution of computational linguistics, and the confrontation between large language models and linguistic theory. It connects to the theory of computation (§3.2), where the Chomsky hierarchy classifies the computational power of language recognition, and to large language models (§5.5), which are the current and most provocative chapter of the relationship.

Background: Theory of computation (§3.2) provides formal language theory, compilers (§4.7) provide parsing and language implementation, and large language models (§5.5) provide the contemporary point of contact with linguistics: syntax, semantics, and language acquisition.

The Relationship: From Shared Grammar to Contested Language

The Shared Foundation: Formal Language Theory

The deepest root of the relationship is formal language theory, which Chomsky developed to characterize human language and which became foundational for computer science. The theory is a single mathematical edifice that both fields use, and its dual citizenship is the clearest case of a foundation shared between them.

Chomsky’s insight, in Syntactic Structures (1957) and the surrounding work, was that a grammar could be treated as a formal system — a set of rules that generate the well-formed sentences of a language — and that grammars could be classified by their generative power. The Chomsky hierarchy (§3.2) classifies grammars into four types of increasing power: regular grammars (recognized by finite automata), context-free grammars (recognized by pushdown automata), context-sensitive grammars (recognized by linear-bounded automata), and unrestricted grammars (recognized by Turing machines). Each level corresponds to a class of formal languages and to a class of machines that recognize them, connecting grammar (a linguistic notion) to computation (a computer science notion) in a single hierarchy.

For computer science, this hierarchy is foundational. The syntax of programming languages is specified by context-free grammars; the parsing of programming languages (§4.7) is the algorithmic recognition of context-free languages; the lexical analysis that precedes parsing recognizes regular languages with finite automata. The entire front end of a compiler is an application of formal language theory, and the theory tells the compiler writer exactly what kind of machine is needed to recognize each kind of language construct. The regular expressions used throughout computing are the practical face of the regular languages at the bottom of the hierarchy. Formal language theory is not a borrowing from linguistics that computer science happens to use; it is a foundation that the two fields share, developed by a linguist and built upon by both.

For linguistics, the hierarchy raised the question of where human language sits. Chomsky argued that human language is more powerful than context-free (that natural language syntax requires at least context-sensitive power, or more precisely “mildly context-sensitive” power, a refinement developed later), and the question of the formal complexity of human language — exactly how much computational power is needed to characterize it — became a research program in mathematical linguistics. The question matters because it bears on what kind of cognitive machinery language requires, which connects to the deeper Chomskyan claims about language and mind below.

The Methodological Evolution: Rules, Statistics, Neural Networks

Computational linguistics — getting computers to process human language — has undergone two methodological revolutions, and the trajectory from rules to statistics to neural networks is the story of the relationship’s recent history and of a deep shift in what counts as understanding language computationally.

The early approach, dominant from the 1950s through the 1980s, was rule-based and grounded in linguistic theory. If language is governed by grammar, then processing it means implementing grammars: hand-written rules for syntax, parsers that applied the rules, and systems that aspired to represent meaning symbolically. This approach embodied the Chomskyan view that language is a rule-governed system, and it sought to process language by encoding the rules that linguistics had discovered. It achieved limited success: rule-based systems were brittle, struggled with the ambiguity and variability of real language, and required enormous hand-crafted effort that did not generalize. The failures paralleled those of symbolic AI generally (§5.1, §9.3) — the rules could not capture the open-ended messiness of real language.

The statistical revolution of the 1990s largely abandoned explicit grammar in favor of learning patterns from data. Instead of hand-writing rules, the statistical approach trained probabilistic models on large corpora of text, learning the regularities of language from examples rather than encoding them by hand. The famous provocation, attributed to Fred Jelinek of IBM’s speech recognition group, captured the shift: “every time I fire a linguist, the performance of the speech recognizer goes up.” The statistical methods worked better than the rule-based ones for many tasks, and they did so largely without the linguistic theory that the rule-based approach had been built on — which was the first sign that engineering success at language processing might not require, and might not validate, the theoretical account of language that linguistics offered.

The neural revolution, culminating in large language models (§5.5), completed the trajectory. Neural language models learn language from data with even less built-in linguistic structure than the statistical methods — no grammar, no parts of speech, no syntactic categories, just the task of predicting the next token, learned from enormous corpora. And they process language with a facility that dwarfs everything before: they generate fluent text, answer questions, translate, and exhibit apparent understanding, all without containing anything that resembles the grammars and rules that linguistics describes. The trajectory from rules to statistics to neural networks is a trajectory away from explicit linguistic theory and toward learning from data, and its endpoint — systems that process language superbly while containing no linguistic theory — is the source of the confrontation that defines the relationship’s current state.

The Confrontation: Large Language Models and Linguistic Theory

The success of large language models has produced a genuine confrontation with linguistic theory, particularly the Chomskyan tradition that has dominated theoretical linguistics, and the confrontation is among the most consequential at any disciplinary boundary in this chapter.

The Chomskyan tradition makes strong claims about language: that human language is governed by an innate universal grammar, a language-specific cognitive endowment that explains how children acquire language so quickly from limited and imperfect input (the “poverty of the stimulus” argument — that the input children receive is too impoverished to explain what they learn without an innate grammatical endowment). On this view, language is a specialized cognitive system, structured by innate principles, and learning a language is setting the parameters of an innately given grammar rather than learning patterns from data.

Large language models are, on the face of it, a challenge to this picture. They learn language — to a high degree of fluency and apparent competence — purely from data, with no innate grammar, no language-specific structure, just a general-purpose learning architecture and an enormous amount of text. If a system with no universal grammar can learn language this well from data, does that undermine the poverty-of-the-stimulus argument and the case for innate universal grammar? The proponents of the neural approach argue that it does, or at least that it shows learning from data is far more powerful than the Chomskyan tradition assumed. The defenders of the Chomskyan view respond that LLMs learn from vastly more data than any child receives (a child does not hear trillions of words), that LLMs may be learning something quite different from what children learn, that fluency is not the same as the specific competence linguistics studies, and that the success of LLMs at producing language does not show that they have the same underlying system as humans or that human language acquisition works the way LLMs do. Chomsky himself has dismissed LLMs as irrelevant to the science of language — as engineering that produces fluent output without illuminating the human capacity — a position that strikes many as too quick given how much LLMs achieve, but that points to a real question about whether engineering success illuminates the underlying science.

The confrontation is genuine and unresolved, and it is sharpened because both sides have serious points. The neural approach has the engineering success: these systems work, and they work without the apparatus linguistics said was necessary, which is at least a challenge to claims that the apparatus is necessary. The linguistic tradition has the scientific question: what is the human language capacity, how is it acquired, what is its structure — questions about humans that the engineering success of LLMs does not obviously answer, because a system that produces language is not thereby a model of how humans do. The deepest question at the boundary is whether large language models illuminate human language or merely imitate its output — whether they are telling us something about what language is, or just doing a remarkable trick that is silent about the human capacity. This question is live, consequential, and unresolved, and it is one of the most interesting disputes at any disciplinary boundary in contemporary intellectual life.

What This Perspective Changes

Seeing the relationship between computer science and linguistics changes how a practitioner understands both formal languages and the natural language that modern systems process.

The first change is recognition of the shared foundation. The practitioner who sees that formal language theory — the Chomsky hierarchy, the classification of grammars and the machines that recognize them — was developed to characterize human language and became foundational for computer science understands the theory’s origin and its dual nature. The parsing of programming languages, the design of regular expressions, the front end of every compiler, are applications of a theory that a linguist developed to understand human language. This shared root is one of the most direct connections between the fields, and seeing it deepens the understanding of formal language theory.

The second change is perspective on the methodological trajectory and what it means. The practitioner who understands the movement from rules to statistics to neural networks sees a deep shift in what counts as processing language computationally — from encoding linguistic theory to learning from data — and understands that the engineering success of the data-driven approaches came largely without the linguistic theory the rule-based approach was built on. This trajectory bears on the relationship between engineering success and scientific understanding: doing a thing well (processing language) turned out not to require, and not to validate, the theory of how the thing works (linguistics).

The third change is the ability to think clearly about what LLMs reveal about language. The practitioner who understands the confrontation between LLMs and linguistic theory can think about the deepest current question — whether these systems illuminate human language or merely imitate it — with the awareness of what each side has at stake and what the engineering success does and does not show. This protects against both the overclaim (LLMs have solved language, refuted Chomsky, shown language is just statistics) and the dismissal (LLMs are irrelevant to understanding language), and it engages the genuine, unresolved question of what the relationship between these systems and human language actually is.

The fourth change is appreciation of the question’s depth. The practitioner who sees that the question of how children acquire language, the question of whether language is innate or learned, and the question of what LLMs reveal are connected — that the engineering at the boundary bears on deep scientific questions about the human mind — appreciates that the relationship between computer science and linguistics reaches into the foundations of what language and mind are. The boundary is not merely technical; it is where some of the deepest questions about human language are now being contested, with the engineering artifacts of computer science as evidence in a scientific debate.

Resources

Foundational and Bridging Texts

Chomsky’s Syntactic Structures (1957) is the foundational text that launched both modern linguistics and, through formal language theory, a foundation of computer science. It is short and historically essential; reading it reveals the origin of the formal grammar concept that both fields use. The formal language theory it initiated is developed rigorously in the theory-of-computation texts of §3.2 (Sipser, Hopcroft-Ullman), which present the Chomsky hierarchy as computer science.

Jurafsky and Martin’s Speech and Language Processing (3rd ed., free at web.stanford.edu/~jurafsky/slp3) is the comprehensive bridge text for computational linguistics, covering the full trajectory from formal grammars and parsing through statistical methods to neural language models. It is the single best source for the relationship as it is practiced, and its coverage of the methodological evolution is exactly the trajectory this section describes. Updated to include large language models, it connects the formal foundations to the contemporary neural approach.

For the linguistics side and the Chomskyan tradition, Chomsky’s later works (Aspects of the Theory of Syntax, 1965, for the competence/performance distinction and the universal grammar program) and accessible treatments of the poverty-of-the-stimulus argument provide the theoretical claims that LLMs now confront. Steven Pinker’s The Language Instinct (1994) is the accessible presentation of the innate-language-faculty view, useful for understanding the position that the neural approach challenges.

Resource Role Type
Jurafsky & Martin, Speech and Language Processing (3rd ed., free) The comprehensive bridge text Depth
Chomsky, Syntactic Structures (1957) The shared foundation of formal grammar Depth
Theory of computation texts (§3.2) Chomsky hierarchy as computer science Depth
Pinker, The Language Instinct Accessible innate-faculty view Entry

The Confrontation: LLMs and Linguistic Theory

The contemporary confrontation is documented in the recent literature and the public exchanges. Chomsky’s dismissals of large language models (his interviews and the 2023 New York Times essay “The False Promise of ChatGPT,” co-authored, free) state the linguistic tradition’s skeptical position. The responses from the natural-language-processing and machine-learning communities — arguing that LLMs do illuminate language, or at least challenge the poverty-of-the-stimulus argument — are in the research literature and in the ongoing public debate. The work of researchers like Christopher Manning, who bridges computational linguistics and deep learning, represents the more synthetic positions.

For the empirical question of what LLMs know about linguistic structure, the interpretability and probing literature (work investigating whether LLMs represent syntactic structure, parts of speech, and other linguistic categories internally) is the relevant research, connecting to the interpretability material of §5.6. This research is attempting to answer empirically whether the systems that process language so well contain anything like linguistic structure — a question directly relevant to what they reveal about language.

For the broader question of language acquisition and what the poverty-of-the-stimulus argument requires, the debate in cognitive science — including the work questioning whether the stimulus is as impoverished as Chomsky claimed, and what statistical learning can achieve — provides the scientific context for the confrontation.

Resource Role Type
Chomsky et al., “The False Promise of ChatGPT” (2023, free) The linguistic tradition’s skeptical position Reference
LLM linguistic-probing literature (§5.6) What LLMs know about linguistic structure Reference
Manning and synthetic-position work Bridging computational linguistics and deep learning Reference
ACL Anthology (free) Computational linguistics and NLP primary literature Reference
Opitz et al., “Natural Language Processing RELIES on Linguistics” (free) Current argument for linguistics in NLP Reference
Language Log (free) Linguistic public scholarship and debate Auxiliary
Language acquisition debate (cognitive science) Scientific context for the confrontation Depth

Traps

Trap Why it misleads Better response
Forgetting that formal language theory came from linguistics Computer scientists use the Chomsky hierarchy, context-free grammars, and parsing theory as native computer science, often without knowing that this entire foundation was developed by a linguist to characterize human language. This obscures the depth of the connection between the fields and the dual nature of the theory. Recognize formal language theory as a foundation shared between the fields, developed by Chomsky to understand human language and built upon by computer science for parsing and language design. The Chomsky hierarchy classifies both human languages and programming languages because it is a theory of grammar that belongs to both.
Claiming LLMs have “solved” or “refuted” linguistics The engineering success of large language models leads to the overclaim that they have solved language, shown it to be “just statistics,” and refuted the Chomskyan tradition. This conflates producing language with understanding the human language capacity: a system that generates fluent text is not thereby a model of how humans acquire or process language, and its success does not directly answer the scientific questions linguistics asks about the human capacity. Distinguish engineering success from scientific understanding. LLMs process language remarkably well; whether they illuminate human language is a separate question that their fluency does not settle. They do challenge specific claims (such as the strongest versions of poverty-of-the-stimulus), but challenging a claim is not refuting a field. Engage the genuine question of what they reveal rather than declaring it answered.
Dismissing LLMs as irrelevant to understanding language The opposite overclaim, exemplified by Chomsky’s dismissal: that LLMs are mere engineering, producing fluent output without illuminating anything about language, and are therefore irrelevant to the science. This is too quick given how much they achieve — a system that learns language from data this successfully is at least a challenge to claims about what language learning requires, and the probing research finds linguistic structure in these systems that bears on the questions. Take the challenge LLMs pose seriously without overclaiming on their behalf. Their success is at least evidence about what learning from data can achieve, which bears on the poverty-of-the-stimulus argument and on claims about innate grammar. The empirical question of what linguistic structure they contain is being investigated and is relevant. Dismissing them as irrelevant forecloses a genuine and consequential question.

中文

计算机科学与语言学的关系，始于一个共同对象——形式语法（formal grammar）——以及一位共同奠基者：诺姆·乔姆斯基（Noam Chomsky）。他在 1950 年代的工作同时属于这两个领域。乔姆斯基试图理解人类语言，在这一过程中，他发展出一套关于语法及其生成语言的数学理论，而这套理论后来成为计算机科学的基础：乔姆斯基层级（Chomsky hierarchy）按照识别形式语言所需的计算能力对其分类，它支撑了解析理论、编程语言设计和编译器构造。乔姆斯基用来刻画人类语言的同一套数学，也刻画了计算机所处理的语言；这种共同基础，是两个领域关系最深的根源。

此后，这种关系沿着一条非常引人注目的轨迹发展，而推动它变化的，是计算语言学（computational linguistics）方法的变化——也就是让计算机处理人类语言的尝试。它最初采用乔姆斯基式的、基于规则的方法：如果语言受形式语法支配，那么处理语言就是实现这套语法。到了 1990 年代，它转向统计方法，很大程度上放弃显式语法，转而从数据中学习模式。最终，它在大语言模型中达到当前阶段：这些系统能以极高熟练度处理语言，却没有任何类似语言学所描述的语法结构。这造成了一场真正的危机，或者至少是一场真正的对峙：一边是神经方法在工程上的成功，另一边是语言学关于语言如何运作的理论主张。两个领域之间的边界，如今已经成为一个活跃且后果重大的争论现场：语言到底是什么，这些系统又揭示了什么？

本节将通过三个视角考察这种关系：形式语言理论这一共同基础、计算语言学的方法演变，以及大语言模型与语言学理论之间的对峙。它连接到计算理论（§3.2），在那里，乔姆斯基层级按照语言识别所需的计算能力进行分类；也连接到大语言模型（§5.5），这是这种关系当前最具挑衅性的章节。

背景知识：计算理论（§3.2）提供形式语言理论；编译器（§4.7）提供解析和语言实现；大语言模型（§5.5）提供当代与语言学接触的关键点：句法、语义和语言习得。

这种关系：从共同语法到争议中的语言

共同基础：形式语言理论

两者关系最深的根源，是形式语言理论。乔姆斯基发展它，是为了刻画人类语言；而它后来成为计算机科学的基础。这个理论是一座两个领域共同使用的数学建筑，它的“双重身份”是二者共享基础中最清楚的例子。

乔姆斯基在 Syntactic Structures（1957）及其相关工作中的洞见是：语法可以被看作一个形式系统——一组生成合格句子的规则——而语法可以按照其生成能力进行分类。乔姆斯基层级（§3.2）把语法分成四种能力逐渐增强的类型：正则语法（由有限自动机识别）、上下文无关语法（由下推自动机识别）、上下文有关语法（由线性有界自动机识别），以及无限制语法（由图灵机识别）。每一级都对应一类形式语言，以及识别这类语言的一类机器，从而把语法这一语言学概念与计算这一计算机科学概念连接到同一个层级之中。

对于计算机科学来说，这个层级是基础性的。编程语言的语法由上下文无关语法指定；编程语言的解析（§4.7）就是对上下文无关语言的算法识别；解析之前的词法分析，则用有限自动机识别正则语言。编译器的整个前端都是形式语言理论的应用，而这个理论准确告诉编译器作者：识别每一种语言结构需要什么类型的机器。计算中随处使用的正则表达式，正是层级底部正则语言的实践形态。形式语言理论不是计算机科学偶然借用的语言学成果；它是两个领域共享的基础，由一位语言学家发展，并由两个领域共同建立其后续工作。

对于语言学来说，这个层级提出了一个问题：人类语言处在什么位置？乔姆斯基认为，人类语言比上下文无关语言更强大；自然语言句法至少需要上下文有关能力，或者更准确地说，需要后来发展出的“弱上下文有关”（mildly context-sensitive）能力。于是，人类语言的形式复杂性问题——到底需要多少计算能力才能刻画它——成为数理语言学中的一个研究项目。这个问题之所以重要，是因为它关系到语言需要何种认知机制，也连接到下文关于语言与心智的更深层乔姆斯基式主张。

方法演变：规则、统计、神经网络

计算语言学——让计算机处理人类语言——经历了两次方法革命。从规则到统计，再到神经网络的轨迹，既是这种关系的近代史，也标志着“以计算方式理解语言”到底意味着什么发生了深刻变化。

早期方法从 1950 年代到 1980 年代占据主导，它是基于规则的，并且以语言学理论为基础。如果语言受语法支配，那么处理语言就意味着实现语法：手写句法规则，构建应用这些规则的解析器，并设计试图以符号方式表示意义的系统。这种方法体现了乔姆斯基式观点：语言是一个受规则支配的系统；它试图通过编码语言学已经发现的规则来处理语言。它取得了有限成功：基于规则的系统很脆弱，难以处理真实语言中的歧义和变异，而且需要巨大的人工规则构造工作，却无法很好泛化。这些失败与符号 AI 的一般失败（§5.1，§9.3）相互呼应——规则无法捕捉真实语言开放而混乱的特性。

1990 年代的统计革命基本放弃了显式语法，转而从数据中学习模式。统计方法不再手写规则，而是在大型文本语料上训练概率模型，让系统从样本中学习语言规律，而不是由人手动编码这些规律。IBM 语音识别团队的 Fred Jelinek 有一句著名挑衅性说法，概括了这种转向：“每当我解雇一个语言学家，语音识别器的性能就会上升。”统计方法在许多任务上比基于规则的方法效果更好，而且它们大体上并不依赖基于规则方法所依托的语言学理论——这是第一个迹象：语言处理的工程成功，可能并不需要语言学提供的语言理论，也未必能验证这种理论。

以大语言模型（§5.5）为顶点的神经网络革命，完成了这条轨迹。神经语言模型从数据中学习语言，比统计方法更少内置语言结构——没有语法，没有词类，没有句法范畴，只有从海量语料中学习预测下一个 token 的任务。它们处理语言的能力远超之前一切方法：它们生成流畅文本，回答问题，翻译，并表现出某种表面上的理解，而这一切并不依赖任何类似语言学所描述的语法和规则的东西。从规则到统计再到神经网络，这条轨迹就是一条远离显式语言学理论、转向从数据中学习的轨迹；而它的终点——那些出色处理语言却不包含语言学理论的系统——正是当前这段关系发生对峙的根源。

对峙：大语言模型与语言学理论

大语言模型的成功与语言学理论产生了真正的对峙，尤其是与长期主导理论语言学的乔姆斯基传统之间的对峙。这是本章所有学科边界中最具后果的冲突之一。

乔姆斯基传统对语言提出了强主张：人类语言受一种先天普遍语法（universal grammar）支配，这是一种语言专属的认知禀赋，用来解释儿童为何能从有限且不完美的输入中如此迅速地习得语言。这就是“刺激贫乏论证”（poverty of the stimulus）：儿童获得的输入太贫乏，无法在没有先天语法禀赋的情况下解释他们实际学到的东西。在这种观点中，语言是一套专门化的认知系统，由先天原则结构化；学习一门语言，不是从数据中学习模式，而是在一套先天给定的语法中设定参数。

从表面上看，大语言模型挑战了这幅图景。它们完全从数据中学习语言，达到了高度流畅和表面上相当强的能力，却没有先天语法，没有语言专属结构，只有一种通用学习架构和海量文本。如果一个没有普遍语法的系统能够仅凭数据把语言学到这种程度，这是否削弱了刺激贫乏论证，以及关于先天普遍语法的论证？神经方法的支持者认为，至少在某种程度上是的；或者至少它说明，从数据中学习的能力远比乔姆斯基传统曾设想的更强。乔姆斯基观点的维护者则回应说：大语言模型接触的数据量远远超过任何儿童，一个儿童并不会听到数万亿个词；大语言模型学到的东西可能与儿童学到的东西非常不同；流畅并不等于语言学所研究的特定能力；大语言模型在生成语言上的成功，并不说明它们拥有与人类相同的底层系统，也不说明人类语言习得的方式与大语言模型相同。乔姆斯基本人曾把大语言模型斥为与语言科学无关——它们只是产生流畅输出的工程系统，却不能照亮人类语言能力。考虑到大语言模型已经取得的成就，这个立场在许多人看来过于迅速，但它确实指向一个真实问题：工程成功是否能照亮底层科学？

这场对峙是真实的，也尚未解决；它之所以尖锐，是因为双方都有严肃理由。神经方法拥有工程成功：这些系统确实有效，而且它们不依赖语言学曾认为必要的装置，这至少挑战了关于这种装置“必要性”的主张。语言学传统则拥有科学问题：人类语言能力是什么，它如何被习得，它的结构是什么——这些关于人的问题，并不会被大语言模型的工程成功自动回答，因为一个能够生成语言的系统，并不因此就是人类如何生成语言的模型。边界处最深的问题是：大语言模型究竟照亮了人类语言，还是只是模仿了语言输出？它们是在告诉我们语言是什么，还是只是在执行一种惊人的技巧，而对人类语言能力保持沉默？这个问题仍然开放，后果重大，并且是当代思想生活中最有意思的学科边界争论之一。

这种视角会改变什么

看清计算机科学与语言学的关系，会改变实践者对形式语言以及现代系统所处理的自然语言的理解。

第一种改变，是认识到二者的共同基础。实践者如果看见形式语言理论——乔姆斯基层级、语法分类以及识别这些语法的机器分类——原本是为了刻画人类语言而发展出来的，并且后来成为计算机科学的基础，就会理解这个理论的起源和双重性质。编程语言的解析、正则表达式的设计、每一个编译器的前端，都是一位语言学家为了理解人类语言而发展出的理论的应用。这个共同根源是两个领域之间最直接的连接之一，而看见它会加深对形式语言理论的理解。

第二种改变，是理解方法演变及其意义。实践者如果理解从规则到统计再到神经网络的转变，就能看见“计算式处理语言”意味着什么发生了深刻变化：从编码语言学理论，变成从数据中学习。并且，他会理解，数据驱动方法在工程上的成功，基本上并不依赖基于规则方法所依托的语言学理论。这条轨迹关系到工程成功与科学理解之间的关系：把一件事做得很好，比如处理语言，事实证明并不需要、也不验证关于这件事如何运作的理论，也就是语言学理论。

第三种改变，是获得清楚思考大语言模型揭示了什么语言问题的能力。实践者如果理解大语言模型与语言学理论之间的对峙，就能带着对双方关切和工程成功边界的认识，思考当前最深的问题：这些系统究竟照亮了人类语言，还是只是模仿了它？这种认识能防止两种过度主张：一种是说大语言模型已经解决了语言、推翻了乔姆斯基、证明语言不过是统计；另一种是说大语言模型与理解语言毫无关系。真正需要面对的，是这些系统与人类语言之间到底是什么关系这一尚未解决的问题。

第四种改变，是认识到这个问题的深度。实践者如果看见儿童如何习得语言、语言究竟是先天的还是后天习得的、以及大语言模型究竟揭示了什么，这些问题彼此相连——也就是说，学科边界处的工程实践正在影响关于人类心智的深层科学问题——就会意识到，计算机科学与语言学的关系触及语言和心智是什么的基础。这个边界不只是技术性的；它是关于人类语言的某些最深问题正在被争论的地方，而计算机科学制造出的工程产物，正在成为这场科学争论中的证据。

资源

基础与桥接文本

乔姆斯基的 Syntactic Structures（1957）是奠基性文本，它开创了现代语言学，也通过形式语言理论奠定了计算机科学的一个基础。它篇幅很短，但在历史上极其重要；阅读它可以看见两个领域共同使用的形式语法概念从何而来。它开启的形式语言理论，在 §3.2 的计算理论教材中得到严格发展，例如 Sipser、Hopcroft-Ullman 等教材，它们把乔姆斯基层级作为计算机科学来讲授。

Jurafsky 和 Martin 的 Speech and Language Processing（第 3 版，可在 web.stanford.edu/~jurafsky/slp3 免费获取）是计算语言学最全面的桥接教材，覆盖从形式语法和解析，到统计方法，再到神经语言模型的完整轨迹。它是理解这种关系在实践中如何展开的最佳单一来源，其对方法演变的覆盖，正好就是本节所描述的轨迹。该书已更新并纳入大语言模型，将形式基础连接到当代神经方法。

从语言学一侧和乔姆斯基传统来看，乔姆斯基后期著作（如 Aspects of the Theory of Syntax，1965，讨论能力/表现区分与普遍语法计划）以及关于刺激贫乏论证的易读处理，提供了大语言模型如今正在面对的理论主张。Steven Pinker 的 The Language Instinct（1994）是关于先天语言能力观点的易读介绍，有助于理解神经方法所挑战的立场。

资源 作用 类型
Jurafsky & Martin, Speech and Language Processing（第 3 版，免费） 全面的桥接教材 深入
Chomsky, Syntactic Structures（1957） 形式语法的共同基础 深入
计算理论教材（§3.2） 作为计算机科学的乔姆斯基层级 深入
Pinker, The Language Instinct 易读的先天语言能力观点 入门

对峙：大语言模型与语言学理论

当代这场对峙记录在近期文献和公开论争之中。乔姆斯基对大语言模型的否定态度——包括他的访谈，以及 2023 年与他人合写、可免费阅读的《The False Promise of ChatGPT》一文——表达了语言学传统中的怀疑立场。自然语言处理和机器学习共同体的回应，则出现在研究文献和持续的公共争论中：他们主张大语言模型确实照亮了语言，或者至少挑战了刺激贫乏论证。像 Christopher Manning 这样横跨计算语言学和深度学习的研究者，代表了更综合性的立场。

关于大语言模型对语言结构到底知道什么这个经验问题，相关研究是可解释性和 probing 文献：它们考察大语言模型内部是否表征了句法结构、词类以及其他语言学范畴。这些研究连接到 §5.6 的可解释性材料，试图用经验方式回答一个与“它们揭示了什么语言问题”直接相关的问题：这些处理语言如此出色的系统，内部是否包含某种类似语言结构的东西？

至于语言习得和刺激贫乏论证到底需要什么这一更广泛的问题，认知科学中的争论提供了科学背景：其中包括质疑儿童获得的刺激是否真的像乔姆斯基所说的那样贫乏，以及统计学习到底能做到什么的研究。

资源 作用 类型
Chomsky 等，”The False Promise of ChatGPT”（2023，免费） 语言学传统的怀疑立场 参考
大语言模型语言学 probing 文献（§5.6） 大语言模型对语言结构知道什么 参考
Manning 及综合立场相关工作 连接计算语言学与深度学习 参考
ACL Anthology（免费） 计算语言学与 NLP 一手文献 参考
Opitz 等，”Natural Language Processing RELIES on Linguistics”（免费） 当前关于 NLP 仍需语言学的论证 参考
Language Log（免费） 语言学公共学术写作与争论 辅助
语言习得争论（认知科学） 这场对峙的科学背景 深入

陷阱

陷阱 为什么会误导 更好的回应
忘记形式语言理论来自语言学 计算机科学家使用乔姆斯基层级、上下文无关语法和解析理论时，常常把它们当成计算机科学内部原生的东西，却不知道这一整套基础最初是由一位语言学家为了刻画人类语言而发展出来的。这会遮蔽两个领域之间联系的深度，也会遮蔽该理论的双重性质。 把形式语言理论理解为两个领域共享的基础：它由乔姆斯基为了理解人类语言而发展，又被计算机科学用于解析和语言设计。乔姆斯基层级之所以既能分类人类语言，也能分类编程语言，是因为它是一套同时属于两个领域的语法理论。
声称大语言模型已经“解决”或“推翻”了语言学 大语言模型在工程上的成功容易导致一种过度主张：它们已经解决语言，证明语言“只是统计”，并推翻乔姆斯基传统。这把生成语言和理解人类语言能力混为一谈：一个能生成流畅文本的系统，并不因此就是人类如何习得或处理语言的模型；它的成功也不能直接回答语言学关于人类能力提出的科学问题。 区分工程成功和科学理解。大语言模型处理语言的能力非常惊人；但它们是否照亮了人类语言，是另一个问题，不会被流畅性本身解决。它们确实挑战了某些具体主张，例如最强版本的刺激贫乏论证，但挑战一个主张，并不等于推翻整个领域。应该面对它们到底揭示了什么这个真实问题，而不是宣布问题已经终结。
把大语言模型斥为与理解语言无关 这是相反方向的过度主张，乔姆斯基的否定态度就是典型例子：认为大语言模型只是工程产物，能生成流畅输出，却不能照亮任何关于语言的问题，因此与语言科学无关。考虑到它们已经取得的成果，这个判断过于迅速——一个系统如果能如此成功地从数据中学习语言，至少会挑战关于语言学习需要什么的主张；而 probing 研究也在这些系统中发现了与语言学问题相关的结构。 认真对待大语言模型提出的挑战，但不替它们过度主张。它们的成功至少说明了从数据中学习能达到什么，这与刺激贫乏论证和先天语法主张有关。它们内部包含何种语言结构，是一个正在被研究的经验问题，并且确实重要。把它们直接斥为无关，会提前关闭一个真实且后果重大的问题。