English
Linguistics studies the structure, use, and acquisition of human language. Computer science studies formal languages and builds systems that process language. The connection supplies both mathematical tools and empirical questions: how can structure be represented, how can ambiguity be resolved, and what would a language-processing system tell us about human speakers?
Three distinctions help organize the subject. Syntax concerns how expressions are structured; semantics concerns meaning; pragmatics concerns use in context. A sentence can be grammatical yet false, and the same sentence can serve different purposes in different conversations. A parser or predictor therefore solves only part of what using language involves.
Background: Formal languages (§3.2), compilers (§4.7), machine learning (§5.2), and language models (§5.5). Cognitive science (§9.3) explains why matching linguistic behavior and explaining human acquisition are different goals.
The Relationship: From Shared Grammar to Contested Language
The Shared Foundation: Formal Language Theory
A formal language is a set of strings over an alphabet. A grammar describes how strings can be generated, while a recognizer decides or recognizes membership under its computational model. Chomsky’s work connected classes of grammars to questions about linguistic structure; automata theory and earlier work in logic supplied other parts of the shared mathematical development.
Regular languages can be recognized by finite automata. Context-free languages allow recursive structure and are recognized by pushdown automata with a stack. The hierarchy also includes context-sensitive languages, associated with linear-bounded automata, and unrestricted grammars, associated with Turing-recognizable languages. Recognition in the last case need not halt on a nonmember. The hierarchy concerns expressive power, not a ranking of languages by usefulness.
Balanced parentheses give a small example of recursive syntax. The rules S → (S)S and S → ε, where ε is the empty string, generate arbitrary finite balanced-parenthesis strings. A recognizer can track how many openings remain unmatched, reject if the count becomes negative, and require zero at the end. Arbitrarily deep nesting needs unbounded counting or stack capacity; a fixed finite-state recognizer cannot handle every depth.
Compiler front ends commonly use regular patterns for tokens and context-free grammars for much of syntax. Other checks, such as whether a name is declared or an expression has an appropriate type, require additional information. Practical regex libraries can also include features such as backreferences that go beyond classical regular expressions. The formal class and the tool with a similar name are not always identical.
For natural language, a grammar can distinguish possible structures without explaining how easily a person processes them. Center-embedded clauses may be grammatical yet difficult to understand because processing resources are limited. Claims about the formal power needed for human language also depend on which constructions and idealizations are included. A hierarchy alone does not establish a complete theory of language or cognition.
The Methodological Evolution: Rules, Statistics, Neural Networks
Rules, probabilities, and neural representations solve overlapping problems rather than forming three mutually exclusive eras. Handwritten rules can encode structural constraints, while statistical models estimate preferences among alternatives from data. Probabilistic grammars combine both. Neural models learn representations and scoring functions, sometimes within systems that also use explicit constraints.
Consider “I saw the person with a telescope.” The final phrase can describe the instrument used for seeing or the person being seen. A grammar may permit both structures. A probabilistic model ranks them using learned patterns; additional context can change the ranking. Producing a parse, selecting the likely interpretation, and establishing what actually happened are separate tasks.
Larger corpora and improved learning methods made data-driven approaches effective for many language tasks. This did not make linguistic analysis irrelevant: annotation conventions, tokenization choices, controlled tests, and distinctions among linguistic phenomena affect what is learned and how it is evaluated. A change in benchmark accuracy can be difficult to interpret without knowing which cases improved.
A neural language model’s architecture and training objective provide inductive biases, even when no explicit grammar is supplied. Predicting tokens can encourage internal structure useful for the task, but fluent continuation does not guarantee factual accuracy or successful reference to the world. Training data, context, and evaluation conditions remain part of the explanation.
Language Models as Evidence About Language and Learning
Language models provide evidence about what particular learning systems can acquire from particular data. They can challenge a claim that a specific hand-coded mechanism is necessary for a tested behavior. To draw that conclusion, state the claimed necessity and demonstrate the behavior under an appropriate test; a general impression of fluency is insufficient.
Human language acquisition asks additional questions. Children learn under particular perceptual, social, and developmental conditions with a different amount and organization of input. A model trained on a large corpus is not automatically a model of that process. Conversely, a difference from children does not make the model useless for testing every linguistic hypothesis. It may isolate one question about learnability or representation.
Theories of innate linguistic structure also differ; “universal grammar” is not one unchanged mechanism shared by all linguistic research. Every learner has some prior structure, including its architecture and learning procedure. The relevant question is which assumptions are needed to explain specified observations, not whether a system learns with literally no assumptions.
Controlled comparisons are more informative than a general contest between disciplines. For subject-verb agreement, vary the number of intervening nouns while keeping the intended relation clear. Compare success, errors, sensitivity to training data, and performance on unfamiliar constructions. Check that test items do not merely repeat training material. Such experiments can reveal both useful generalization and limits without treating a model as a complete account of language.
What This Perspective Changes
When building a language-processing system, separate the structure it recognizes, the interpretations it ranks, and the claims about meaning or truth it can support. Choose evaluation cases that test those distinctions.
When reading a claim about human language, ask what population, data, task, and learning conditions were studied. A shared formalism can transfer methods between linguistics and CS while leaving different empirical questions to be answered in each field.
Resources
Foundational and Bridging Texts
Jurafsky and Martin’s Speech and Language Processing materials connect linguistic tasks with computational methods. Chomsky’s Syntactic Structures supplies historical context for formal grammar; automata texts explain the machine–language correspondences precisely. Pinker presents an account emphasizing an innate language faculty. Treat it as a position to understand alongside other acquisition research, rather than as the definition of linguistics.
| Resource | Role | Type |
|---|---|---|
| Jurafsky & Martin, Speech and Language Processing (online draft) | Linguistic tasks and computational methods | Depth |
| Chomsky, Syntactic Structures (1957) | Historical development of formal grammar | Depth |
| Theory of computation texts (§3.2) | Formal languages and their recognizing machines | Depth |
| Pinker, The Language Instinct | An account emphasizing an innate language faculty | Entry |
Language Models and Linguistic Evidence
Use debate essays to identify competing claims, then look for experiments that distinguish them. Linzen, Dupoux, and Goldberg’s 2016 subject–verb agreement study illustrates controlled testing of syntactic dependencies; its particular results concern its models and training conditions. ACL Anthology supplies primary papers, while Language Log and the other listed discussions provide interpretation and argument. Performance evidence, explanations of learning, and claims about human language acquisition should be evaluated separately.
| Resource | Role | Type |
|---|---|---|
| Chomsky et al., “The False Promise of ChatGPT” (2023) | A skeptical argument about language-model explanation | Reference |
| LLM linguistic-probing literature (§5.6) | Experiments on linguistic information and behavior | Reference |
| Manning and synthetic-position work | Accounts connecting linguistic structure and learned models | Reference |
| ACL Anthology (free) | Primary papers and evaluation methods | Reference |
| Opitz et al., “Natural Language Processing RELIES on Linguistics” (free) | An argument for linguistic knowledge in NLP | Reference |
| Language Log (free) | Public discussion of linguistic evidence and claims | Auxiliary |
| Language acquisition debate (cognitive science) | Evidence and theories about learning language | Depth |
Traps
| Trap | Why it misleads | Better response |
|---|---|---|
| Treating grammar, meaning, and truth as one task | A valid structure can be ambiguous or false. | Evaluate syntax, interpretation, and factual support separately. |
| Assuming learned models have no prior structure | Architecture, objectives, and data organization constrain learning. | Identify the relevant inductive biases and compare alternatives. |
| Using engineering success to settle human acquisition | Training conditions and target explanations may differ. | Test a specified hypothesis with controlled tasks and relevant evidence. |
中文
语言学研究人类语言的结构、使用与习得。计算机科学研究形式语言，也构建处理语言的系统。这种联系同时带来数学工具与经验问题：怎样表示结构，怎样消除歧义，以及语言处理系统能告诉我们哪些关于人类说话者的事情？
三个区别有助于组织主题。句法涉及表达式怎样构成，语义涉及意义，语用涉及情境中的使用。句子可以符合语法却不真实；同一句话在不同对话中，也可能具有不同目的。因此，解析器或预测器只解决了语言使用的一部分。
背景知识：形式语言（§3.2）、编译器（§4.7）、机器学习（§5.2）与语言模型（§5.5）。认知科学（§9.3）解释了为什么匹配语言行为与解释人类习得是不同目标。
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
形式语言是字母表上的字符串集合。文法描述怎样生成字符串，识别器则在某种计算模型下判定或识别成员关系。Chomsky 的工作把文法类别与语言结构问题联系起来；自动机理论和更早的逻辑工作，也参与形成了共同的数学基础。
正则语言可以由有限自动机识别。上下文无关语言允许递归结构，可由带栈的下推自动机识别。层级还包括与线性有界自动机对应的上下文有关语言，以及与 Turing 可识别语言对应的无限制文法。最后一种情形的识别器，面对非成员输入时未必终止。该层级讨论表达能力，不是根据实用价值给语言排名。
配对括号提供了递归句法的小例子。规则 S → (S)S 与 S → ε 中，ε 表示空串；它们可以生成任意有限的平衡括号串。识别器可以记录尚未配对的左括号数量；计数变负时拒绝，结束时要求归零。任意深度的嵌套需要无界计数或栈容量，固定的有限状态识别器不能处理所有深度。
编译器前端通常用正则模式识别词元，用上下文无关文法描述大部分句法。但名称是否声明、表达式类型是否合适等检查，需要额外信息。实用正则表达式库还可能包含反向引用等超出经典正则表达式能力的特性。形式类别与具有相似名称的工具，不一定完全相同。
对于自然语言，文法可以区分可能的结构，却不能单独解释人处理它们有多容易。中心嵌套从句可能符合语法，却因处理资源有限而难以理解。关于人类语言需要多强形式能力的结论，也取决于纳入哪些构式与理想化假设。仅有层级，不能建立完整的语言或认知理论。
方法演变：规则、统计、神经网络
规则、概率与神经表示解决相互重叠的问题，并非三个彼此排斥的时代。手写规则可以编码结构约束，统计模型则从数据估计不同候选之间的偏好。概率文法结合了两者。神经模型学习表示与评分函数，也可能置于同时使用显式约束的系统中。
考虑“I saw the person with a telescope”。末尾短语可以说明观察使用的工具，也可以说明被看到的人带着望远镜。文法可能允许两种结构；概率模型利用学到的规律排序，额外情境则可能改变排序。生成解析、选择可能的解释，以及确定实际发生了什么，是不同任务。
更大语料与学习方法的改进，使数据驱动方法在许多语言任务中有效。但这没有使语言分析失去作用：标注约定、分词选择、受控测试，以及语言现象之间的区别，都会影响学习内容与评估方式。如果不知道哪些案例改善了，基准准确率的变化也可能难以解释。
即使没有显式文法，神经语言模型的架构与训练目标也提供归纳偏置。预测词元可能促使模型形成对任务有用的内部结构，但流畅续写不能保证事实准确，也不能保证成功指涉世界。训练数据、上下文和评估条件仍然属于解释的一部分。
把语言模型作为研究语言与学习的证据
语言模型提供了关于特定学习系统能从特定数据获得什么的证据。如果有人声称某种手写机制是某项受测行为的必要条件，模型可能挑战这一说法。要得出该结论，需要明确所谓必要条件，并用适当测试展示行为；笼统的流畅印象还不够。
人类语言习得还提出其他问题。儿童在特定感知、社会与发展条件下学习，输入的数量和组织方式也不同。在大规模语料上训练的模型，不会自动成为该过程的模型。反过来，与儿童不同，也不代表模型对检验所有语言学假设都无用；它可能单独检验某个可学习性或表示问题。
关于先天语言结构的理论也各不相同，“普遍语法”并不是所有语言研究共同采用、从未变化的单一机制。每种学习者都有某些先验结构，包括架构与学习过程。相关问题是解释指定观察需要哪些假设，而不是系统是否字面上完全没有假设。
受控比较比学科之间的笼统对抗更有信息量。例如，研究主谓一致时，可以改变中间名词的数量，同时保持目标关系明确。比较成功、错误、对训练数据的敏感性，以及面对陌生构式的表现，并检查测试项目是否只是重复训练材料。这样的实验既能揭示有用的泛化，也能发现限制，无需把模型当成语言的完整解释。
这种视角会改变什么
构建语言处理系统时，区分它识别的结构、排序的解释，以及它能够支持的意义或真实性判断。选择能检验这些区别的评估案例。
阅读关于人类语言的结论时，应问研究了哪些人群、数据、任务和学习条件。共同形式体系可以让方法在语言学与 CS 之间迁移，同时保留各自仍需回答的经验问题。
资源
基础与桥接文本
Jurafsky 与 Martin 的 Speech and Language Processing 材料联系语言任务与计算方法。Chomsky 的《Syntactic Structures》提供形式语法的历史背景；自动机教材精确解释机器与语言之间的对应。Pinker 提出强调先天语言能力的解释，应结合其他习得研究理解这一立场，而不是把它当作语言学的定义。
| 资源 | 作用 | 类型 |
|---|---|---|
| Jurafsky & Martin，Speech and Language Processing（在线草稿） | 语言任务与计算方法 | 深入 |
| Chomsky, Syntactic Structures（1957） | 形式语法的历史发展 | 深入 |
| 计算理论教材（§3.2） | 形式语言及其识别机器 | 深入 |
| Pinker, The Language Instinct | 强调先天语言能力的解释 | 入门 |
语言模型与语言学证据
可以从争论文章中找出相互竞争的主张，再寻找能够区分它们的实验。Linzen、Dupoux 和 Goldberg 的 2016 年主谓一致研究展示了怎样控制条件来检验句法依赖；具体结果针对其中的模型与训练条件。ACL Anthology 提供原始论文，Language Log 与其他所列讨论提供解释和论证。表现证据、学习机制解释，以及人类语言习得主张，应分别评价。
| 资源 | 作用 | 类型 |
|---|---|---|
| Chomsky 等，The False Promise of ChatGPT（2023） | 质疑语言模型解释力的论证 | 参考 |
| 大语言模型语言学 probing 文献（§5.6） | 有关语言信息与行为的实验 | 参考 |
| Manning 及综合立场相关工作 | 联系语言结构与学习模型的解释 | 参考 |
| ACL Anthology（免费） | 原始论文与评价方法 | 参考 |
| Opitz 等，”Natural Language Processing RELIES on Linguistics”（免费） | 论证语言学知识在 NLP 中作用的文章 | 参考 |
| Language Log（免费） | 对语言学证据与主张的公开讨论 | 辅助 |
| 语言习得争论（认知科学） | 语言学习的证据与理论 | 深入 |
陷阱
| 误区 | 为什么会误导 | 更好的做法 |
|---|---|---|
| 把文法、意义和真实性当作同一任务 | 合法结构也可能有歧义或不真实。 | 分别评估句法、解释与事实依据。 |
| 认为学习模型没有先验结构 | 架构、目标与数据组织都会约束学习。 | 明确相关归纳偏置，并比较替代方案。 |
| 用工程成功直接解决人类习得问题 | 训练条件与解释目标可能不同。 | 用受控任务与相关证据检验明确假设。 |