Natural language processing (NLP) is a subfield of computer science and artificial intelligence concerned with enabling computers to process, generate, and analyze human language in text and speech. It integrates computational linguistics, statistics, and machine learning—especially deep learning—to build systems for tasks such as translation, question answering, summarization, and information extraction.
Last updated November 17, 2025
Natural language processing (NLP) is the area of computer science and artificial intelligence that builds algorithms and systems for analyzing, understanding, and generating human language in text and speech, combining methods from computational linguistics, statistics, and machine learning to approximate human-like language competence, according to Britannica and the open 3rd‑edition manuscript of Jurafsky and Martin’s textbook hosted by Stanford University (Speech and Language Processing).
Definition and scope
NLP encompasses the construction of models and pipelines that segment tokens, analyze morphology, assign parts of speech, parse syntax, resolve entities and coreference, map to semantic and discourse representations, and generate fluent outputs in applications such as translation and dialogue systems, as described in the Stanford textbook (Speech and Language Processing). According to Britannica, contemporary NLP systems rely on statistical learning and deep neural networks rather than exclusively on hand‑engineered rules. The field is closely related to Machine learning and Computational linguistics, with the premier professional society being the Association for Computational Linguistics (ACL), which describes itself as the international scientific society for work on language and computation and traces its name change to 1968 after its 1962 founding (ACL Member Portal; ACL 2023).
Historical development
Early modern discussion of machine intelligence and language was framed by the “imitation game” in Alan Turing’s 1950 article, which introduced what became known as the Turing test ([Computing Machinery and Intelligence](journal://Mind|Computing Machinery and Intelligence|1950); see also the Oxford journal PDF: Mind, 1950). A widely publicized early demonstration was the Georgetown–IBM machine translation experiment on January 7, 1954, which automatically translated Russian sentences into English and spurred funding, though it used a small vocabulary and simple rules (Wikipedia overview with primary references). In 1966 the U.S. Automatic Language Processing Advisory Committee issued the ALPAC report, which judged progress insufficient and recommended curtailing machine translation funding, reshaping research priorities in computational linguistics (National Academies Press).
From the late 1980s into the 1990s, statistical and corpus‑based methods displaced purely rule‑based approaches; hidden Markov models and probabilistic grammars became standard for tagging and parsing, a shift summarized in classic tutorials and texts (Rabiner 1989, Proceedings of the IEEE; [Foundations of Statistical NLP](book://Christopher D. Manning; Hinrich Schütze|Foundations of Statistical Natural Language Processing|MIT Press|1999)). Large shared corpora such as the Penn Treebank standardized evaluation and drove rapid progress in supervised learning for syntax and other tasks (Computational Linguistics 1993). The 2010s saw widespread adoption of neural methods, beginning with distributed word representations (word embeddings) and progressing to contextual encoders and large pretrained language models (word2vec; GloVe; ELMo). The 2017 introduction of the [Transformer (machine learning model)] architecture enabled more parallelizable training and state‑of‑the‑art results in machine translation and beyond (Attention Is All You Need). Transformer‑based pretraining produced models such as BERT, which set new benchmarks on a range of understanding tasks, and autoregressive large language models like GPT‑3 that demonstrated strong few‑shot performance via prompting (BERT paper; OpenAI GPT‑3 paper; corroborating overview: OpenAI blog).
Methods and models
Core methodological families in NLP include symbolic and rule‑based systems (grammars, lexicons), probabilistic models (e.g., n‑grams, HMMs, CRFs), and neural architectures (RNNs, CNNs, and Transformers), with the field’s mainstream practice now centered on pretrained neural models fine‑tuned or prompted for tasks, as surveyed by Jurafsky and Martin (Speech and Language Processing). Word embeddings such as word2vec and GloVe provide dense vectors that capture distributional semantics from large corpora, serving as inputs or initialization for downstream models (word2vec; GloVe). Contextual encoders like ELMo and BERT learn token representations conditioned on sentential context, improving transfer across tasks such as question answering and natural language inference (ELMo; BERT). The Transformer eliminates recurrence in favor of self‑attention and has become the dominant backbone for translation, summarization, and language modeling (Attention Is All You Need). Autoregressive scaling yielded few‑shot learners (e.g., GPT‑3) that can solve diverse tasks without gradient updates by conditioning on textual prompts (GPT‑3).
Core tasks
Common tasks include tokenization and morphological analysis, part‑of‑speech tagging, syntactic parsing, named entity recognition, coreference resolution, semantic role labeling, sentiment analysis, natural language inference, question answering, and machine translation, as summarized by the field’s standard textbook (Speech and Language Processing). Benchmarks such as SQuAD advanced reading comprehension by providing large‑scale QA datasets for span extraction and comparison of models under common protocols (SQuAD 2016). In translation and summarization, automatic metrics enabled rapid iteration; for example, BLEU uses n‑gram precision with brevity penalties to correlate with human MT judgments, and ROUGE families compute n‑gram and sequence overlaps for summaries (BLEU, ACL 2002; ROUGE, ACL 2004). GLUE and SuperGLUE consolidated diverse English understanding tasks into leaderboards for general language understanding assessment, accelerating progress on transfer and prompting methods (GLUE, 2018; SuperGLUE, 2019).
Data resources and infrastructure
Annotated corpora and lexical resources underpin supervised and semi‑supervised NLP research; the Penn Treebank standardized syntactic annotation and evaluation, while WordNet provides a large, relational lexical database of English synsets and semantic relations used in word sense disambiguation and semantic similarity tasks (Penn Treebank, CL 1993; WordNet homepage, Princeton). Web‑scale unlabeled data is commonly sourced from open repositories such as Common Crawl, a nonprofit collection of hundreds of billions of web pages that supports large‑scale pretraining and analysis (Common Crawl overview; AWS registry entry: Open Data on AWS). ACL organizes open bibliographic and proceedings infrastructure via the ACL Anthology, facilitating dissemination and citation of research outputs across subfields (ACL Member Portal).
Evaluation practices
Task‑specific accuracy and F‑scores are complemented by shared metrics and leaderboards; BLEU remains a standard for machine translation and ROUGE for summarization, despite known limitations around sensitivity to paraphrase and discourse structure (BLEU, ACL 2002; ROUGE, ACL 2004). Multi‑task suites such as GLUE and SuperGLUE provide aggregated scores to track generalization across natural language inference, similarity, and commonsense tasks, which have been significantly advanced by pretrained Transformers (GLUE; SuperGLUE; BERT). For language modeling, perplexity remains a common intrinsic metric, historically rooted in probabilistic modeling of text ([Foundations of Statistical NLP](book://Christopher D. Manning; Hinrich Schütze|Foundations of Statistical Natural Language Processing|MIT Press|1999)).
Applications
NLP systems appear in search and Information retrieval (including query understanding and ranking), in virtual assistants and chatbots, in machine translation, in summarization and content moderation, and in assistive technologies such as screen readers, leveraging advances in pretrained and promptable models ([Introduction to Information Retrieval](book://Christopher D. Manning; Prabhakar Raghavan; Hinrich Schütze|Introduction to Information Retrieval|Cambridge University Press|2008); Britannica). Transformer architectures and large pretrained models power production‑scale translation and generative text systems, reflecting research findings such as the advantages of attention‑only models and the emergent few‑shot behavior of scaled autoregressive LMs (Attention Is All You Need; GPT‑3).
Community and publications
The field’s flagship venues include the Annual Meeting of the ACL (with regional NAACL, EACL, and AACL chapters) and journals such as Computational Linguistics and Transactions of the ACL, as outlined in the organization’s description and conference materials (ACL Member Portal; ACL 2023). The ACL Anthology serves as the primary open repository for conference and journal papers, including foundational datasets and metrics papers such as Penn Treebank, BLEU, and ROUGE (Penn Treebank, CL 1993; BLEU; ROUGE).
Ethical and societal considerations
Scholars have documented biases and harms that can arise from training data and model behavior, including gender and racial stereotypes in embeddings and downstream systems, calling for clearer definitions, evaluation, and mitigation strategies in NLP research (Gonen & Goldberg 2019; Blodgett et al. 2020). Critical discussions of large language models emphasize environmental costs, data governance, and the risk of producing fluent but misleading outputs, urging dataset documentation and value‑sensitive design (“Stochastic Parrots,” FAccT 2021; see also the University of Washington summary: UW News). Ongoing debates about the use of large web corpora—such as Common Crawl—for training models intersect with copyright, consent, and access questions central to responsible NLP practice (Common Crawl overview; reporting on publisher challenges: Wired, 2024).
Natural language processing (NLP) is a subfield of computer science and artificial intelligence concerned with enabling computers to process, generate, and analyze human language in text and speech. It integrates computational linguistics, statistics, and machine learning—especially deep learning—to build systems for tasks such as translation, question answering, summarization, and information extraction.
Last updated November 17, 2025
Natural language processing (NLP) is the area of computer science and artificial intelligence that builds algorithms and systems for analyzing, understanding, and generating human language in text and speech, combining methods from computational linguistics, statistics, and machine learning to approximate human-like language competence, according to Britannica and the open 3rd‑edition manuscript of Jurafsky and Martin’s textbook hosted by Stanford University (Speech and Language Processing).
Definition and scope
NLP encompasses the construction of models and pipelines that segment tokens, analyze morphology, assign parts of speech, parse syntax, resolve entities and coreference, map to semantic and discourse representations, and generate fluent outputs in applications such as translation and dialogue systems, as described in the Stanford textbook (Speech and Language Processing). According to Britannica, contemporary NLP systems rely on statistical learning and deep neural networks rather than exclusively on hand‑engineered rules. The field is closely related to Machine learning and Computational linguistics, with the premier professional society being the Association for Computational Linguistics (ACL), which describes itself as the international scientific society for work on language and computation and traces its name change to 1968 after its 1962 founding (ACL Member Portal; ACL 2023).
Historical development
Early modern discussion of machine intelligence and language was framed by the “imitation game” in Alan Turing’s 1950 article, which introduced what became known as the Turing test ([Computing Machinery and Intelligence](journal://Mind|Computing Machinery and Intelligence|1950); see also the Oxford journal PDF: Mind, 1950). A widely publicized early demonstration was the Georgetown–IBM machine translation experiment on January 7, 1954, which automatically translated Russian sentences into English and spurred funding, though it used a small vocabulary and simple rules (Wikipedia overview with primary references). In 1966 the U.S. Automatic Language Processing Advisory Committee issued the ALPAC report, which judged progress insufficient and recommended curtailing machine translation funding, reshaping research priorities in computational linguistics (National Academies Press).
From the late 1980s into the 1990s, statistical and corpus‑based methods displaced purely rule‑based approaches; hidden Markov models and probabilistic grammars became standard for tagging and parsing, a shift summarized in classic tutorials and texts (Rabiner 1989, Proceedings of the IEEE; [Foundations of Statistical NLP](book://Christopher D. Manning; Hinrich Schütze|Foundations of Statistical Natural Language Processing|MIT Press|1999)). Large shared corpora such as the Penn Treebank standardized evaluation and drove rapid progress in supervised learning for syntax and other tasks (Computational Linguistics 1993). The 2010s saw widespread adoption of neural methods, beginning with distributed word representations (word embeddings) and progressing to contextual encoders and large pretrained language models (word2vec; GloVe; ELMo). The 2017 introduction of the [Transformer (machine learning model)] architecture enabled more parallelizable training and state‑of‑the‑art results in machine translation and beyond (Attention Is All You Need). Transformer‑based pretraining produced models such as BERT, which set new benchmarks on a range of understanding tasks, and autoregressive large language models like GPT‑3 that demonstrated strong few‑shot performance via prompting (BERT paper; OpenAI GPT‑3 paper; corroborating overview: OpenAI blog).
Methods and models
Core methodological families in NLP include symbolic and rule‑based systems (grammars, lexicons), probabilistic models (e.g., n‑grams, HMMs, CRFs), and neural architectures (RNNs, CNNs, and Transformers), with the field’s mainstream practice now centered on pretrained neural models fine‑tuned or prompted for tasks, as surveyed by Jurafsky and Martin (Speech and Language Processing). Word embeddings such as word2vec and GloVe provide dense vectors that capture distributional semantics from large corpora, serving as inputs or initialization for downstream models (word2vec; GloVe). Contextual encoders like ELMo and BERT learn token representations conditioned on sentential context, improving transfer across tasks such as question answering and natural language inference (ELMo; BERT). The Transformer eliminates recurrence in favor of self‑attention and has become the dominant backbone for translation, summarization, and language modeling (Attention Is All You Need). Autoregressive scaling yielded few‑shot learners (e.g., GPT‑3) that can solve diverse tasks without gradient updates by conditioning on textual prompts (GPT‑3).
Core tasks
Common tasks include tokenization and morphological analysis, part‑of‑speech tagging, syntactic parsing, named entity recognition, coreference resolution, semantic role labeling, sentiment analysis, natural language inference, question answering, and machine translation, as summarized by the field’s standard textbook (Speech and Language Processing). Benchmarks such as SQuAD advanced reading comprehension by providing large‑scale QA datasets for span extraction and comparison of models under common protocols (SQuAD 2016). In translation and summarization, automatic metrics enabled rapid iteration; for example, BLEU uses n‑gram precision with brevity penalties to correlate with human MT judgments, and ROUGE families compute n‑gram and sequence overlaps for summaries (BLEU, ACL 2002; ROUGE, ACL 2004). GLUE and SuperGLUE consolidated diverse English understanding tasks into leaderboards for general language understanding assessment, accelerating progress on transfer and prompting methods (GLUE, 2018; SuperGLUE, 2019).
Data resources and infrastructure
Annotated corpora and lexical resources underpin supervised and semi‑supervised NLP research; the Penn Treebank standardized syntactic annotation and evaluation, while WordNet provides a large, relational lexical database of English synsets and semantic relations used in word sense disambiguation and semantic similarity tasks (Penn Treebank, CL 1993; WordNet homepage, Princeton). Web‑scale unlabeled data is commonly sourced from open repositories such as Common Crawl, a nonprofit collection of hundreds of billions of web pages that supports large‑scale pretraining and analysis (Common Crawl overview; AWS registry entry: Open Data on AWS). ACL organizes open bibliographic and proceedings infrastructure via the ACL Anthology, facilitating dissemination and citation of research outputs across subfields (ACL Member Portal).
Evaluation practices
Task‑specific accuracy and F‑scores are complemented by shared metrics and leaderboards; BLEU remains a standard for machine translation and ROUGE for summarization, despite known limitations around sensitivity to paraphrase and discourse structure (BLEU, ACL 2002; ROUGE, ACL 2004). Multi‑task suites such as GLUE and SuperGLUE provide aggregated scores to track generalization across natural language inference, similarity, and commonsense tasks, which have been significantly advanced by pretrained Transformers (GLUE; SuperGLUE; BERT). For language modeling, perplexity remains a common intrinsic metric, historically rooted in probabilistic modeling of text ([Foundations of Statistical NLP](book://Christopher D. Manning; Hinrich Schütze|Foundations of Statistical Natural Language Processing|MIT Press|1999)).
Applications
NLP systems appear in search and Information retrieval (including query understanding and ranking), in virtual assistants and chatbots, in machine translation, in summarization and content moderation, and in assistive technologies such as screen readers, leveraging advances in pretrained and promptable models ([Introduction to Information Retrieval](book://Christopher D. Manning; Prabhakar Raghavan; Hinrich Schütze|Introduction to Information Retrieval|Cambridge University Press|2008); Britannica). Transformer architectures and large pretrained models power production‑scale translation and generative text systems, reflecting research findings such as the advantages of attention‑only models and the emergent few‑shot behavior of scaled autoregressive LMs (Attention Is All You Need; GPT‑3).
Community and publications
The field’s flagship venues include the Annual Meeting of the ACL (with regional NAACL, EACL, and AACL chapters) and journals such as Computational Linguistics and Transactions of the ACL, as outlined in the organization’s description and conference materials (ACL Member Portal; ACL 2023). The ACL Anthology serves as the primary open repository for conference and journal papers, including foundational datasets and metrics papers such as Penn Treebank, BLEU, and ROUGE (Penn Treebank, CL 1993; BLEU; ROUGE).
Ethical and societal considerations
Scholars have documented biases and harms that can arise from training data and model behavior, including gender and racial stereotypes in embeddings and downstream systems, calling for clearer definitions, evaluation, and mitigation strategies in NLP research (Gonen & Goldberg 2019; Blodgett et al. 2020). Critical discussions of large language models emphasize environmental costs, data governance, and the risk of producing fluent but misleading outputs, urging dataset documentation and value‑sensitive design (“Stochastic Parrots,” FAccT 2021; see also the University of Washington summary: UW News). Ongoing debates about the use of large web corpora—such as Common Crawl—for training models intersect with copyright, consent, and access questions central to responsible NLP practice (Common Crawl overview; reporting on publisher challenges: Wired, 2024).
