Natural language processing, or NLP, is the subfield of Artificial intelligence focused on enabling computers to process, interpret, and generate human language, spoken or written. It covers tasks ranging from narrow ones such as spell-checking and part-of-speech tagging to open-ended ones such as translation, summarization, and conversation, and it is the parent field of the Large language model technology that dominates AI discourse in the mid-2020s.
Early era: rules and symbols
The earliest NLP systems, from the 1950s through the 1980s, were built on hand-crafted linguistic rules and symbolic logic, an approach aligned with Symbolic AI. ELIZA, a 1966 chatbot created by Joseph Weizenbaum, used simple pattern matching to simulate a Rogerian psychotherapist and became an early, if limited, demonstration of the field's ambitions. Rule-based Machine translation systems were also developed during this period, with mixed success.
Statistical era
From the late 1980s through the 2000s, NLP shifted toward statistical methods, using probabilities learned from text corpora rather than hand-written rules to handle tasks like translation, tagging, and parsing. This period produced techniques such as n-gram language models and statistical machine translation, and established large annotated corpora as central research infrastructure, a precursor to the modern emphasis on training data scale.
Neural and transformer era
Neural approaches began replacing statistical pipelines in the early 2010s, aided by word-level embeddings such as Word2vec (2013), which represented words as dense vectors capturing semantic relationships, and by Recurrent neural network and LSTM architectures for Seq2seq tasks such as translation. The 2017 introduction of the Transformer (architecture) architecture proved transformative for the field, enabling far more efficient training on large text corpora. BERT (2018) showed the power of large-scale pretraining for language understanding, while the GPT series demonstrated that a single Autoregressive model trained on next-token prediction at sufficient scale could perform an enormous range of NLP tasks without task-specific architectures, culminating in general-purpose large language models.
Modern scope
Contemporary NLP research and applications span machine translation, Speech recognition, semantic search, sentiment analysis, summarization, question answering, and dialogue systems underlying chatbots and voice assistants. Since the rise of large language models, many previously distinct NLP subtasks are now commonly addressed by prompting a single general-purpose model rather than training specialized architectures, a shift that reduced the need for narrow supervised systems but raised new evaluation and reliability questions, including Hallucination (AI).
Challenges
NLP systems continue to struggle with low-resource languages that lack large training corpora, culturally specific idioms and context, and consistent factual reliability. Bias present in training text can be reproduced or amplified in model outputs, a concern raised prominently in the Stochastic parrot critique, and remains an active area of research alongside broader efforts in Algorithmic bias mitigation.