Natural Language Processing

Natural language processing (NLP) is the discipline of making machines work with human language: understanding it (parsing, classification, entity and relation extraction, question answering), generating it (summarization, translation, drafting), and representing it (the embeddings that make text computable). Its modern history compresses into one arc: from hand-built linguistic rules, through statistical methods and task-specific neural models, to the transformer architecture and the large language models that now perform essentially every NLP task through one pretrained, instruction-following interface — the consolidation that reshaped this entire glossary's subject.

Document AI is applied NLP fused with vision: once recognition produces text, everything intelligent done with it is NLP's toolkit at work. Classification sorts the documents; NER and relation extraction structure their contents; summarization condenses them; question answering interrogates them; and the language-model reasoning behind agents, copilots, and multi-step analysis is NLP's current frontier operating on document content. The document setting adds its characteristic pressures — text arriving via OCR with noise, meaning distributed across layout as well as language, domain registers (legal, clinical, financial) dense with terms general corpora underrepresent — which is why document NLP leans on the adaptations this glossary catalogs: layout-aware models, domain fine-tuning, grounding disciplines, and validation layers that catch what fluency hides.

For practitioners orienting in the vocabulary: NLP names the field; LLMs are its current dominant instruments; NLU and NLG (understanding and generation) are its traditional halves; and the classical task taxonomy (tokenization, tagging, parsing, classification, extraction, QA, summarization) survives as the vocabulary of evaluation and system design even where one model performs all of it. The durable insight the field's history teaches — and document AI re-learns continuously — is that language capability is bounded by data and evaluation more than architecture: the models are general; the labeled examples, domain corpora, and honest benchmarks are where specific performance actually comes from.