Timeline of Natural Language Processing
Natural language processing from machine-translation proposals in 1949 to reasoning models in 2025: linguistic rules, text corpora, statistical methods, neural representations, and conversational systems.
- Category
- Engineering & Technology
- Tag
- computing
- Events
- 24 events 1949 – 2025
- Generated
- by gpt-6
1949: Weaver proposes translation by computer
Warren Weaver circulated his memorandum "Translation," connecting automated translation with information theory, cryptanalysis, and the use of context to resolve ambiguity. His proposals helped organize research into processing human languages with digital computers.
https://en.wikipedia.org/wiki/Warren_Weaver1950: Turing proposes a conversational test
Alan Turing published "Computing Machinery and Intelligence," replacing the question of whether machines think with an imitation game conducted through written exchanges. The proposal made language behavior a lasting benchmark in debates about machine intelligence.
https://en.wikipedia.org/wiki/Computing_Machinery_and_Intelligence1954: Georgetown and IBM demonstrate machine translation
Georgetown University and International Business Machines (IBM) demonstrated Russian-to-English translation on an IBM 701. Carefully selected sentences and a small vocabulary showed that translation could be automated in a restricted setting, attracting funding despite the system's narrow scope.
https://en.wikipedia.org/wiki/Georgetown%E2%80%93IBM_experiment1957: Chomsky publishes a formal account of syntax
Noam Chomsky published "Syntactic Structures," presenting a generative approach to grammar. Formal rules for building sentences influenced computational parsing and sharpened the distinction between grammatical structure and the meanings of sentences.
https://en.wikipedia.org/wiki/Syntactic_Structures1964: The Brown Corpus makes text machine-readable
W. Nelson Francis and Henry Kučera released the Brown Corpus, a machine-readable collection of roughly one million words from American texts published in 1961. Its systematic sampling gave researchers a shared basis for measuring language use and evaluating computational methods.
https://en.wikipedia.org/wiki/Brown_Corpus1966: Weizenbaum publishes the ELIZA program
Joseph Weizenbaum described ELIZA, which matched keywords and transformed input sentences into conversational responses. Its therapist-like DOCTOR script exposed how readily people could attribute understanding to a program that followed relatively simple text-processing rules.
https://en.wikipedia.org/wiki/ELIZA1966: A federal report challenges translation research
The Automatic Language Processing Advisory Committee (ALPAC) concluded that automatic translation had not justified expectations of quick practical success. Its report redirected United States funding toward basic language research and tools to assist human translators.
https://en.wikipedia.org/wiki/ALPAC1970: Winograd connects language to a simulated world
Terry Winograd completed SHRDLU at the Massachusetts Institute of Technology. The program answered questions and followed instructions about simulated blocks, combining grammar, reasoning, and a model of its environment while remaining limited to that small world.
https://en.wikipedia.org/wiki/SHRDLU1985: Princeton begins building WordNet
George A. Miller and colleagues began WordNet at Princeton University. The lexical database organized words into sets of related meanings and linked concepts such as broader and narrower categories, giving language software a reusable resource beyond alphabetical dictionaries.
https://en.wikipedia.org/wiki/WordNet1990: IBM publishes statistical translation results
Researchers at International Business Machines (IBM) described a statistical approach to translation trained on parallel English and French parliamentary text. The work helped shift machine translation toward learning probable word correspondences from examples rather than relying entirely on hand-written linguistic rules.
https://en.wikipedia.org/wiki/Statistical_machine_translation1993: The Penn Treebank supplies annotated sentences
Mitchell Marcus, Beatrice Santorini, and Mary Ann Marcinkiewicz published the Penn Treebank project's design and results. Its manually checked grammatical annotations supported the training and comparison of statistical parsers, making shared corpora central to language research.
https://en.wikipedia.org/wiki/Treebank2001: Conditional random fields model labeled sequences
John Lafferty, Andrew McCallum, and Fernando Pereira introduced conditional random fields, a probabilistic framework for labeling structured data. The method supported tasks such as identifying names and assigning parts of speech while considering dependencies between neighboring labels.
https://en.wikipedia.org/wiki/Conditional_random_field2002: BLEU standardizes automatic translation scoring
Kishore Papineni and colleagues introduced bilingual evaluation understudy (BLEU), which compared machine translations with human reference translations. The inexpensive automatic score made repeated experiments easier, although a numerical match could not fully measure translation quality.
https://en.wikipedia.org/wiki/BLEU2003: Neural language models learn word representations
Yoshua Bengio and colleagues published a neural probabilistic language model that jointly learned word vectors and next-word predictions. Similar words could share statistical strength, helping address the scarcity of examples that affected models based on exact word sequences.
https://en.wikipedia.org/wiki/Language_model#Neural_models2006: Google brings translation to the web
Google launched Google Translate, making automatic translation available through a widely used web service. The product later moved from its initial translation technology to Google's statistical systems and then neural models, bringing research methods into everyday multilingual communication.
https://en.wikipedia.org/wiki/Google_Translate2011: Watson defeats human Jeopardy champions
The Watson system from International Business Machines (IBM) defeated Ken Jennings and Brad Rutter in televised "Jeopardy!" matches. It combined language analysis, information retrieval, and evidence scoring to answer varied clues without an internet connection during play.
https://en.wikipedia.org/wiki/IBM_Watson2013: Word2vec makes word embeddings practical
Tomáš Mikolov and colleagues introduced efficient methods for learning word vectors from large text collections. Their word2vec software helped make reusable numerical representations of words a standard component of language applications and research.
https://en.wikipedia.org/wiki/Word2Vec2014: Neural translation learns where to attend
Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio proposed an attention mechanism that let a translation model focus on relevant parts of an input sentence as it generated words. The approach reduced the bottleneck of compressing an entire sentence into one fixed-length representation.
https://en.wikipedia.org/wiki/Attention_(machine_learning)2017: Researchers introduce attention-based transformers
Ashish Vaswani and colleagues published "Attention Is All You Need," introducing the transformer architecture for translation. Attention mechanisms connected words directly and allowed more parallel training, providing the foundation for many later language models.
https://en.wikipedia.org/wiki/Attention_Is_All_You_Need2018: BERT learns from both sides of words
Google researchers introduced Bidirectional Encoder Representations from Transformers (BERT), pretraining a model to recover masked words using surrounding context. Adapting the resulting representations improved a range of language-understanding benchmarks without designing a separate architecture for each task.
https://en.wikipedia.org/wiki/BERT_(language_model)2020: Retrieval augments neural text generation
Patrick Lewis and colleagues introduced retrieval-augmented generation, combining a language model with passages retrieved from an external document collection. The method separated some factual knowledge from model parameters and provided a framework for generating answers grounded in retrieved material.
https://en.wikipedia.org/wiki/Retrieval-augmented_generation2022: ChatGPT brings conversational generation to the public
OpenAI released ChatGPT as a public research preview on November 30. Training with human feedback helped the system respond to instructions and follow-up questions, while its fluent answers could still contain factual errors or unsupported claims.
https://en.wikipedia.org/wiki/ChatGPT2023: Meta releases Llama 2 model weights
Meta released Llama 2 with downloadable weights and a license permitting many research and commercial uses. Developers gained a widely available starting point for adapting and running large language models, although the license included restrictions beyond conventional open-source software licenses.
https://en.wikipedia.org/wiki/Llama_(language_model)2025: DeepSeek releases a reasoning-focused language model
The Chinese company DeepSeek released DeepSeek-R1 and described training methods that used reinforcement learning to improve reasoning on mathematical and programming tasks. Downloadable model weights and smaller distilled versions widened access to this approach to language-model training.
https://en.wikipedia.org/wiki/DeepSeek_(AI)#DeepSeek-R1