第一章：人工智能与自然语言处理的历史演变 / Chapter 1: The Evolution of AI and Natural Language Processing

第一章：人工智能与自然语言处理的历史演变

Chapter 1: The Evolution of AI and Natural Language Processing

1.1 引言：从图灵测试到ChatGPT

从1950年阿兰·图灵（Alan Turing）提出著名的“图灵测试”开始，人工智能（Artificial Intelligence, AI）作为一个独立的学科正式登上了历史舞台。图灵在其论文《Computing Machinery and Intelligence》中提出了一个简单却深刻的问题：机器能够思考吗？这个问题不仅奠定了人工智能的哲学基础，更指明了人工智能发展的终极目标——让机器展现出与人类智能难以区分的能力。七十年后的今天，当我们回顾人工智能的发展历程时，会发现这条道路远比预想的更加曲折和丰富。从早期的符号推理系统到今天的深度学习大模型，人工智能经历了多次范式转换，每一次转换都深刻改变了我们对智能本质的理解。

In 1950, Alan Turing introduced the renowned "Turing Test" in his seminal paper "Computing Machinery and Intelligence," marking the birth of artificial intelligence as a formal discipline. Turing posed a profound question: Can machines think? This question not only established the philosophical foundation of AI but also pointed toward its ultimate goal—creating machines that exhibit capabilities indistinguishable from human intelligence. Looking back at seven decades of AI development, we find this journey far more winding and enriching than anticipated. From early symbolic reasoning systems to today's large deep learning models, AI has undergone multiple paradigm shifts, each fundamentally reshaping our understanding of intelligence.

自然语言处理（Natural Language Processing, NLP）作为人工智能的核心子领域之一，其发展轨迹更是充满了戏剧性的起伏。自然语言是人类智慧最集中的体现，它承载着知识、情感、意图和文化。从早期的“机翻笑话”到今天的ChatGPT，NLP领域经历了从规则到统计、从统计到深度学习、再从深度学习到大规模预训练的多次革命。每一次技术范式的转变，都带来了能力的质的飞跃，也带来了对“什么是语言理解”这一根本问题的重新思考。本章将带领读者回顾这段波澜壮阔的技术演进史，探讨人工智能和自然语言处理如何走到今天，并为理解后续章节将要深入讨论的大语言模型奠定基础。

Natural Language Processing (NLP), as one of the core subfields of AI, has experienced particularly dramatic ups and downs. Natural language represents the most concentrated expression of human intelligence, carrying knowledge, emotion, intention, and culture. From early "machine translation jokes" to today's ChatGPT, NLP has undergone multiple revolutions—from rules to statistics, from statistics to deep learning, and from deep learning to large-scale pretraining. Each paradigm shift has brought qualitative leaps in capability, along with reconsiderations of the fundamental question: What is language understanding? This chapter takes readers through this magnificent technological evolution, exploring how AI and NLP arrived at where they are today, and laying the foundation for understanding the large language models to be discussed in depth in subsequent chapters.

1.2 人工智能的三大流派

人工智能在其七十余年的发展历程中，逐渐形成了三大主要流派：符号主义（Symbolicism）、连接主义（Connectionism）和行为主义（Behaviorism）。这三大流派代表了不同的智能观和方法论，它们之间的竞争与合作、融合与分立，构成了人工智能学术史的主线。

During its more than seventy years of development, AI has gradually formed three main schools of thought: Symbolicism, Connectionism, and Behaviorism. These three schools represent different views of intelligence and methodological approaches. The competition and cooperation, fusion and separation among them constitute the main thread of AI's academic history.

1.2.1 符号主义：逻辑与知识的殿堂

符号主义，又称GOFAI（Good Old-Fashioned Artificial Intelligence，优良旧式人工智能），是人工智能领域最早且最具影响力的流派。其核心思想是：智能可以被形式化描述为对符号的操作。人类的思维过程可以分解为一系列逻辑推理步骤，而逻辑推理本质上是一种符号转换游戏。因此，只要我们能够将知识形式化为符号系统，并设计出正确的推理规则，就可以实现人工智能。

Symbolicism, also known as GOFAI (Good Old-Fashioned Artificial Intelligence), is the earliest and most influential school in AI. Its core idea is that intelligence can be formally described as symbol manipulation. Human thinking can be decomposed into a series of logical reasoning steps, and logical reasoning is essentially a game of symbol transformation. Therefore, as long as we can formalize knowledge into a symbolic system and design correct inference rules, we can achieve artificial intelligence.

符号主义的兴起与几位计算机科学先驱的工作密不可分。1956年夏天，在达特茅斯学院（Dartmouth College）举行的那次著名会议上，约翰·麦卡锡（John McCarthy）首次提出了“人工智能”这一术语。值得注意的是，这次会议的其他参与者还包括马文·明斯基（Marvin Minsky）、克劳德·香农（Claude Shannon）和纳撒尼尔·罗切斯特（Nathaniel Rochester）等人。他们中的大多数都是符号主义的信徒，相信可以通过逻辑和启发式搜索来解决智能问题。

The rise of Symbolicism is inseparable from the work of several computer science pioneers. In the summer of 1956, at the famous conference at Dartmouth College, John McCarthy first proposed the term "Artificial Intelligence." Notably, other participants included Marvin Minsky, Claude Shannon, and Nathaniel Rochester, most of whom were believers in Symbolicism, believing that intelligent problems could be solved through logic and heuristic search.

符号主义的代表性成果包括通用问题求解器（General Problem Solver, GPS）、专家系统（Expert Systems）和逻辑编程语言Prolog。GPS由艾伦·纽维尔（Allen Newell）和赫伯特·西蒙（Herbert Simon）于1957年开发，是第一个将问题求解过程形式化的程序。它采用手段-目的分析（Means-Ends Analysis）的方法，通过比较当前状态和目标状态之间的差异，选择合适的操作符来减少差异。这种方法虽然在实际应用中效率有限，但它开创了将智能行为形式化的先河。

Representative achievements of Symbolicism include the General Problem Solver (GPS), Expert Systems, and the logic programming language Prolog. GPS, developed by Allen Newell and Herbert Simon in 1957, was the first program to formalize the problem-solving process. It used the Means-Ends Analysis method, comparing differences between the current state and goal state to select appropriate operators to reduce those differences. Although this method had limited efficiency in practical applications, it pioneered the formalization of intelligent behavior.

专家系统是符号主义最成功的应用形式之一。1970年代至1980年代间，以MYCIN（用于诊断细菌感染的专家系统）和DENDRAL（用于推断分子结构的专家系统）为代表的一批专家系统展现了符号方法的强大能力。这些系统将领域专家的知识编码为“如果-那么”（If-Then）规则，通过前向链或后向链推理来解决问题。MYCIN的知识库包含约450条规则，其在诊断细菌感染方面的准确率据报道可达65%左右，接近甚至超过了一些人类传染病专家的水平。

Expert systems represent one of the most successful applications of Symbolicism. During the 1970s and 1980s, a batch of expert systems including MYCIN (for diagnosing bacterial infections) and DENDRAL (for inferring molecular structures) demonstrated the powerful capabilities of symbolic methods. These systems encoded domain expert knowledge into If-Then rules, solving problems through forward or backward chaining inference. MYCIN's knowledge base contained approximately 450 rules, and its accuracy in diagnosing bacterial infections was reported to reach about 65%, close to or even exceeding the level of some human infectious disease experts.

然而，符号主义也面临着难以克服的局限性。最根本的问题在于知识的获取和表示。人类专家的很多知识是隐性的、直觉性的，难以用明确的规则来表达。随着问题领域的扩大，规则的数量呈指数级增长，陷入了所谓的“知识获取瓶颈”。此外，符号系统缺乏处理不确定性、模糊性和噪声的能力，而这些都是现实世界问题的常态特征。

However, Symbolicism also faces insurmountable limitations. The most fundamental problem lies in knowledge acquisition and representation. Much of human expert knowledge is tacit and intuitive, making it difficult to express in explicit rules. As the problem domain expands, the number of rules grows exponentially, leading to the so-called "knowledge acquisition bottleneck." Additionally, symbolic systems lack the ability to handle uncertainty, ambiguity, and noise—characteristics that are normal in real-world problems.

1.2.2 连接主义：神经网络与学习的革命

连接主义代表了与符号主义截然不同的智能观。其核心主张是：智能不是来自预先编程的规则，而是来自大量简单单元之间的相互作用。1943年，神经生理学家沃伦·麦卡洛克（Warren McCulloch）和数学家沃尔特·皮茨（Walter Pitts）提出了第一个人工神经元模型——M-P神经元模型。这个模型受到了生物神经元结构的启发，将神经元抽象为一个多输入、单输出的处理单元，通过加权求和和阈值函数来产生输出。

Connectionism represents a fundamentally different view of intelligence from Symbolicism. Its core claim is that intelligence does not come from pre-programmed rules but from interactions among numerous simple units. In 1943, neurophysiologist Warren McCulloch and mathematician Walter Pitts proposed the first artificial neuron model—the M-P neuron model. Inspired by biological neuron structure, this model abstracted neurons as processing units with multiple inputs and single output, producing outputs through weighted sum and threshold functions.

$$y = f\left(\sum_{i=1}^{n} w_i x_i - \theta\right)$$

其中，$x_i$ 表示输入，$w_i$ 表示连接权重，$\theta$ 表示阈值，$f$ 表示激活函数。这个简单的数学模型虽然粗糙，却抓住了神经信息处理的核心思想——信息通过加权连接传递，并通过非线性变换产生输出。

Where $x_i$ represents inputs, $w_i$ represents connection weights, $\theta$ represents the threshold, and $f$ represents the activation function. Although this simple mathematical model was crude, it captured the core idea of neural information processing—information is transmitted through weighted connections and produces outputs through nonlinear transformation.

1958年，弗兰克·罗森布拉特（Frank Rosenb