Natural language processing (NLP) is the branch of AI concerned with getting computers to read, understand, and generate human language: translation, search, summarization, question answering, and conversation. Its history is one of the clearest examples of a field changing its core method several times.
The first era was rule-based. Researchers tried to encode the grammar and vocabulary of language by hand, expecting that enough carefully written rules would add up to understanding. Early machine translation pursued this path, and the 1966 ALPAC report, which judged the results disappointing, helped trigger a long funding winter for the field.
The second era was statistical. In 1990 an IBM team published “A Statistical Approach to Machine Translation,” reframing the problem as one of probability learned from large collections of translated text rather than hand-coded rules. This data-driven philosophy spread across NLP and dominated for two decades. The third era was neural: methods like word2vec (2013) and sequence-to-sequence models learned dense numerical representations of words and sentences. The fourth era arrived with the Transformer and models like BERT (2018), whose paper reported state-of-the-art results on eleven NLP benchmarks by pre-training a deep bidirectional model on huge amounts of plain text, and then with the large language models that followed.
For business readers, NLP is the technology behind chatbots, document search, sentiment analysis, and the language abilities of modern AI assistants. Its long arc, from rules to statistics to neural networks to large language models, is the single best illustration of how the data-and-learning approach gradually overtook hand-built expertise across all of AI.