Natural Language Processing (NLP) is the branch of artificial intelligence that gives machines the ability to read, understand, and generate human language. It sits at the intersection of computer science, linguistics, and machine learning. NLP is behind every spell checker, search engine, voice assistant, and chatbot you've ever used. The field tackles a deceptively hard problem: human language is ambiguous (the same word means different things in different contexts), contextual (meaning depends on what came before), and grounded in world knowledge (understanding "the trophy doesn't fit in the suitcase because it's too big" requires knowing that trophies have physical size). These properties make language the hardest data type in AI.

NLP began with Alan Turing's 1950 question "Can machines think?" The 1954 Georgetown-IBM experiment translated 60+ Russian sentences using just 250 words and 6 grammar rules — and researchers predicted machine translation would be solved within 3–5 years. They were off by about 60 years. ELIZA (1966) by Joseph Weizenbaum used simple pattern matching to mimic a psychotherapist, convincing some users they were talking to a real person. SHRDLU (1970s) could understand natural language commands about blocks on a table — but only blocks on a table. Noam Chomsky's formal grammars (1957) provided the theoretical foundation: language could be described by rules. The problem was that real language has too many rules, too many exceptions, and too much ambiguity for hand-crafted systems to handle.

The statistical revolution replaced hand-crafted rules with probabilities learned from data. Instead of writing grammar rules, researchers trained models on large text corpora and let the statistics emerge. Hidden Markov Models (HMMs) dominated POS tagging and speech recognition. Naive Bayes and logistic regression powered text classification. Statistical machine translation (IBM Models, phrase-based MT) replaced rule-based translation. The key insight was Frederick Jelinek's famous quote: "Every time I fire a linguist, the performance of the speech recognizer goes up." Data and statistics outperformed hand-crafted linguistic knowledge. This era also introduced rigorous evaluation — BLEU scores for translation, F1 for classification — making NLP a measurable science.

Word2Vec (2013) was the watershed moment: words could be represented as dense vectors where geometry captured meaning — "king − man + woman = queen" worked as actual vector arithmetic. This eliminated the need for hand-crafted features. Recurrent Neural Networks (RNNs) and their improved variants LSTMs and GRUs could process sequences of variable length, making them natural fits for language. Sequence-to-sequence models with attention (2014–2015) revolutionized machine translation, and ELMo (2018) introduced contextual embeddings — the same word gets different vectors depending on its context. The deep learning era automated feature engineering: instead of telling the model what to look for, you let it discover patterns in raw text.

The 2017 paper "Attention Is All You Need" introduced the Transformer architecture, which replaced recurrence with self-attention — every token can attend to every other token in parallel. This unlocked massive parallelism and enabled training on unprecedented data scales. BERT (2018) used the encoder side for understanding tasks (classification, NER, question answering), achieving state-of-the-art on 11 NLP benchmarks simultaneously. GPT (2018–2020) used the decoder side for generation. T5 (2019) unified all NLP tasks as text-to-text. The transformer didn't just improve NLP — it unified it. One architecture, pre-trained on massive text, could be fine-tuned for virtually any language task.

Despite the enormous variety of NLP applications, most problems reduce to a handful of fundamental task types. Text classification: assign a label to a document (sentiment, topic, spam). Sequence labeling: assign a label to each token (POS tagging, NER). Sequence-to-sequence: transform one sequence into another (translation, summarization). Text generation: produce text from a prompt or context. Information extraction: pull structured data from unstructured text (relation extraction, event detection). Semantic similarity: measure how similar two texts are (paraphrase detection, search). Understanding which task type your problem maps to is the first step in choosing the right approach.

A standard NLP pipeline transforms raw text into actionable predictions through a series of stages. Data acquisition: collect raw text plus metadata (source, timestamp, language). Preprocessing: clean, normalize, and tokenize the text. Representation: convert tokens into numerical vectors the model can process. Modeling: apply a machine learning or deep learning model to the vectors. Post-processing: format outputs, apply business rules, filter low-confidence predictions. Evaluation: measure performance against ground truth using task-appropriate metrics. Each stage has its own design decisions and failure modes. The pipeline is only as strong as its weakest stage — a perfect model on poorly preprocessed text will produce poor results.

This course follows the evolution of NLP from foundations to the modern landscape. Chapters 2–3 cover the pipeline's foundation: text preprocessing and representation (from bag-of-words to Word2Vec). Chapters 4–5 tackle the core tasks: text classification and sequence labeling, with both classical and neural approaches. Chapter 6 covers language models and generation — the foundation that led to GPT. Chapter 7 is the transformer revolution: BERT, GPT, and T5. Chapters 8–9 cover the modern workflow: transfer learning, fine-tuning, and evaluation metrics. Chapter 10 surveys the current landscape: instruction tuning, few-shot NLP, multilingual models, and where the field is heading next.