Bilingual glossary & index Glossaire bilingue et index

Context evaluationNLP

The share of all decisions of a system that are correct, true positives and true negatives together.

Notes It can only be calculated where the true negatives can be counted, as in classification, and it is misleading when the class looked for is rare.

other attested French terms: justesse

See also precisionrecalltrue negativebaseline

Context agentsAI

AI systems that pursue a goal through one or more AI agents, deciding their next steps with limited human guidance.

Notes Such systems usually remember relevant information from earlier steps, use external tools or software and act on the results, often with several specialized agents working together.

other attested French terms: IA générative multiagent, IA agentique

See also AI agenttool use

Context agentsAI

An AI system that selects and carries out actions toward a goal, often by using tools or calling other programs and adapting later actions to the results it receives.

other attested English terms: agent

other attested French terms: agent intelligent autonome

See also agentic AItool use

Context ethics & societyAI

AI-generated content that looks polished or competent at first glance but has little substance, takes very little effort to produce compared with making it without AI, and can be produced in very large quantities.

other attested French terms: IA dégénérative, bouillie d'IA

See also information pollutionsynthetic content

Context AI safetylanguage models

The loss of ability on some tasks that can come with training a language model to behave as intended, seen as lower scores after alignment training on tests the model handled before it.

Notes The word comes from an older worry in alignment research, discussed by Askell et al. (2021), that a model made safer might also be made weaker. Ouyang et al. (2022) saw it happen: after preference tuning their models did worse on some standard tests, translation among them, and they recovered most of that ground by mixing pretraining data back into the alignment training. Side effects of alignment training, such as refusals that reach too far, less varied writing and longer answers, are often discussed alongside it.

See also alignment¹fine-tuningreinforcement learning from human feedbacksycophancy

Context AI

The process and goal of ensuring AI systems behave according to human values and intentions.

Notes Alignment training can cost a model some ability on other tasks, a trade-off called the alignment tax. Askell et al. (2021) raised the concern, Ouyang et al. (2022) measured the loss, and Bai et al. (2022b) found that it fell on their smaller models, while their larger ones did as well or better after that training.

See also preference tuningalignment taxalignment²

Context CAT toolstranslation

The process of pairing a source text with an existing translation of it, segment by segment, so that the pairs can be stored as translation units in a translation memory.

Notes Alignment is how a memory is built out of documents translated before any tool was involved, or translated by someone else. Software aligns automatically, by sentence or by word, and the output is checked by hand: one sentence split into two in the translation, or a paragraph moved, puts the two texts out of step from that point onwards.

other attested English terms: text alignment, translation alignment, bitext alignment

other attested French terms: alignement de texte

See also alignment¹translation memorytranslation unitparallel corpusTranslation Memory eXchange

Context regular expressions

The either/or of a regular expression, written with a vertical bar: /cat|dog/ matches one or the other, and the first alternative that matches at that point is the one taken.

Notes The bar has the lowest precedence of any operator, so on its own it separates everything to its left from everything to its right; a group is what limits its reach, which is why /^(cat|dog)$/ and /^cat|dog$/ do not mean the same thing.

other attested English terms: disjunction

See also regular expressioncapturing groupquantifieroperator precedence

Context linguisticsNLP

The property of a stretch of language that allows more than one interpretation.

Notes It occurs at every level of language, from a word with several grammatical classes or senses to a sentence whose structure can be analyzed in more than one way. Readers resolve most of it from context without noticing; a system has to compute each resolution.

See also polysemypart-of-speech taggingcontextual embedding

Context ethics & societyAI

The tendency of a trained model to reproduce a pattern in its training data more strongly than the pattern occurs in the data, so that a tendency in the data becomes a default in the output.

Notes Zhao et al. (2017) named it for social bias; the same sharpening applies to any pattern in the data, for instance the ideas or the languages a model produces most.

other attested English terms: bias amplification

other attested French terms: amplification des biais

See also training biasbias in AIhomogeneity bias

Context regular expressions

A regular-expression construct that matches a position in the text rather than a character, so it makes a match succeed or fail without consuming anything.

Notes The line anchors are ^ for the start of a line and $ for the end; under the multiline flag they apply line by line rather than to the whole text. Because they consume nothing, /^$/ finds an empty line, where the two positions fall together, and /^\d/ finds a digit at the start of a line while matching only that digit.

See also regular expressionword boundaryflag

Context computing

A standardized interface that allows software applications to communicate by sending requests and receiving responses, without needing to know how the other application works internally.

See also function callingtool usetoken-based billing

Context AI

A field of research and practice concerned with developing computer systems that perform functions associated with intelligence, such as perception, learning, reasoning, language use, and decision-making. By extension, the term also refers to the systems themselves, as in the increasingly common expression "an AI". For this second sense, the OECD defines an AI system as "a machine-based system that, for explicit or implicit objectives, infers, from the input it receives, how to generate outputs such as predictions, content, recommendations, or decisions that can influence physical or virtual environments. Different AI systems vary in their levels of autonomy and adaptiveness after deployment" (Grobelnik et al., 2024).

Context text encoding

A 7-bit coded character set of 128 characters for information interchange, standardized as ANSI X3.4 and adopted as the US national variant of ISO/IEC 646. It contains the unaccented letters of the English alphabet, the digits, common punctuation and a set of control codes, and no accented letter. The first 128 characters of Unicode are the ASCII ones, in the same order, which is why UTF-8 stores them in a single byte.

other attested English terms: American Standard Code for Information Interchange

other attested French terms: ASCII, code américain normalisé pour l'échange d'information

See also character encodingUTF-8word boundary

Context deep learninglanguage models

A rule or numerical structure that prevents an attention layer from assigning attention to specified token positions.

Notes For example, a causal attention mask blocks access to later positions during next-token prediction, while a padding mask prevents the model from attending to placeholder tokens added to make sequences the same length. An attention mask controls the flow of information within an attention calculation; it is different from the masking used in masked language modeling, where selected input tokens are hidden or altered to create a prediction task.

See also attention mechanismcausal language modelmasked language model

Context deep learninglanguage models

A "method in neural networks that helps a model focus on the most relevant parts of the input when producing an output. For example, when translating a sentence, it 'pays attention' to the words that matter most for the next translated word, which improves accuracy and handling of long sequences." (Stanford HAI).

See also self-attentioncross-attentionattention mask

Context XML

In XML, a name-value pair written in an element's start-tag or empty-element tag to provide additional information about that element, such as its language, identifier, or creation date.

See also tageXtensible Markup Language

Context prompting

The use of an algorithm to generate, evaluate, and iteratively revise candidate prompts in order to improve a model's performance on a specified task or evaluation measure.

See also prompt engineering

Context speech

Technology that analyses the human voice captured by a microphone and transcribes it into machine-readable text; also called speech-to-text (STT).

Context NLPterminology

The identification, by a program, of candidate terms in a text or a corpus, for a person to validate.

Notes The output is a ranked list of candidates, or terms marked where they occur in the text.

other attested English terms: automatic term recognition, terminology extraction

other attested French terms: dépouillement automatique, extraction terminologique

See also termcandidate termtermhoodunithoodnamed entity recognition

Context language models

A method of generating a sequence, such as text, one token at a time.

Notes At each step, the model uses the tokens generated so far to assign probabilities to possible next tokens, selects one using a decoding method, and repeats the process until it reaches a stopping point. Autoregressive indicates that each generated token depends on the preceding sequence; it does not mean that the model considers only the immediately preceding token.

other attested English terms: causal generation

See also next-token predictiondecodingcausal language model

Context language models

A language model that has completed pretraining but has not yet undergone post-training.

Notes It can generate text by predicting successive tokens, but it has not received additional training intended to improve instruction-following, alignment with human preferences, or complex reasoning.

See also pre-trainingpost-trainingfoundation model

Context evaluationmachine learning

The score that a very simple method reaches on a test, with which the score of a system is compared.

Notes The word is also used for the simple method itself. For classification, common baselines are a random guess and always answering the most frequent class.

other attested French terms: modèle étalon

See also benchmarkgold standardtest set

Context language models

A decoding method that keeps a fixed number of promising partial sequences at each generation step, extends them with the possible next tokens and keeps only the highest-scoring ones for the next step.

Notes The number of sequences kept is the beam width. Beam search is common in machine translation and speech recognition, while chatbots usually sample their next token instead.

other attested French terms: algorithme de recherche en faisceau

See also decodinggreedy decodingmachine translation

Context evaluationAI

A standardized evaluation used to measure and compare how AI systems perform on particular tasks or capabilities.

Notes It generally combines a test set (such as questions, examples, or another task) with a scoring method such as accuracy. A benchmark may overstate a model's abilities if its test material appeared in the model's training data (see data contamination). Moreover, strong performance on a particular benchmark does not necessarily translate into strong performance in real-world situations.

other attested French terms: test d'évaluation des performances

See also data contaminationhuman preference evaluationtest set

Context ethics & societyAI

A systematic tendency in an AI system, or in the processes surrounding it, that makes certain outputs, errors, or effects more likely than others.

Notes In responsible AI, the term usually means unwanted bias, which makes results inaccurate, unrepresentative or unfair for some people or groups. It need not be intentional, and statistical bias does not necessarily amount to discrimination.

See also training biashomogeneity bias

Context machine learning

A real number added to a unit's weighted inputs, learned during training, which shifts its output independently of any single input.

Notes Not the bias of bias in AI, which is about unfair outcomes: this one is a quantity inside the model, and the two senses share only the word. Weights and bias terms are the two kinds of number that make up a model's parameters.

other attested English terms: intercept

See also weightparameterbias in AItraining bias

Context language models

A language model that can use context from token positions both before and after a given position when computing its representation or prediction.

Notes In a bidirectional transformer encoder, self-attention can draw on the entire available input sequence. This is useful for representing and analyzing existing text, as in BERT and other masked language models. Bidirectional does not necessarily mean that the text is processed in two separate passes; in a transformer encoder it means that a token representation may incorporate information from positions on both sides. A bidirectional language model is not necessarily a masked language model, although masked language modeling is a common way to train one.

other attested English terms: bidirectional model

other attested French terms: modèle de langage bidirectionnel, modèle bidirectionnel

See also masked language modelencoder-only modelcausal language modelcontextual embedding

Context AImachine learning

A system whose internal decision-making cannot be inspected: what goes in and what comes out can both be observed, and the process connecting them cannot be read.

Notes Trained neural networks are the standard case, since their behaviour is stored across millions of connection strengths and no single one of them means anything on its own; how such a system handles a given kind of text is found out by testing it rather than by reading it. The opacity is sometimes a property of the method and sometimes a commercial choice, when a provider withholds the workings of a model it sells.

other attested French terms: boîte opaque

See also neural networkdeep learningparameter

Context CAT toolsterminology

A list of target-language terms that must not be used in a translation, which a CAT tool's quality assurance checks flag wherever they appear.

Notes It complements the termbase, which says what to use: the blocklist says what not to use, such as a term a client rejects, a proscribed calque or an old product name. Trados and memoQ call such entries forbidden terms. In Wordfast Pro a blocklist has one language only, and a forbidden term may carry a suggested replacement. It colours nothing in the editor; its terms show up only in the checks.

other attested English terms: forbidden terms list

other attested French terms: liste de rejet, liste de termes interdits

See also termbasequality assurance

Context language modelsNLP

A subword tokenization method that learns a vocabulary from training text by repeatedly joining frequently occurring adjacent characters or bytes.

Notes It then applies the learned combinations to new text, allowing common words or sequences to become single tokens while representing rarer or unfamiliar words as several smaller tokens.

See also subword tokenizationtokenization²vocabulary

Context NLPterminology

A unit that a term extraction program proposes as a possible term, before a person has validated it.

other attested English terms: term candidate

other attested French terms: candidat-terme, terme candidat

See also automatic term extractionterm

Context regular expressions

In a regular expression, a part of a pattern enclosed in parentheses, so that a quantifier or an alternation applies to the whole of it rather than to one character: /(ha)+/ repeats the pair, /(anti|pro)nuclear/ puts the choice inside the shared word.

Notes What each group matched is also kept, which is what lets a find-and-replace put those pieces back into the replacement.

other attested French terms: groupe

See also regular expressionquantifieralternation

Context environment

The total amount of greenhouse gases released into the atmosphere as a result of human activities, usually expressed in carbon dioxide equivalents (CO₂e).

Notes "A carbon footprint is the total amount of greenhouse gases [...] released into the atmosphere as a result of human activities. These emissions are primarily carbon dioxide (CO₂), but also include methane (CH₄), nitrous oxide (N₂O) and other gases that contribute to climate change. To make these different greenhouse gases easier to compare, a carbon footprint is usually expressed in carbon dioxide equivalents (CO₂e) [...], based on each gas's global warming potential [...]. The concept can be applied at different levels: individuals have a personal carbon footprint shaped by everyday choices such as transport, home energy use, diet and consumption habits. Products carry a carbon footprint across their life cycle, from raw material extraction and manufacturing to transport, use, disposal or recycling. Organizations generate carbon emissions through their operations, supply chains, buildings, services and products. Even specific events and activities, such as a conference, concert or flight, can be assessed for their carbon impact." (ISO).

See also training emissionsinference emissionsdata centre

Context language models

A language model that predicts each token using only the token positions that precede it in the sequence; it cannot use later tokens when making that prediction.

Notes This restriction makes the model suitable for autoregressive generation, in which it generates a token, adds it to the available context and then predicts the next one. In language modelling, causal refers to the restriction on access to later token positions; it does not mean that the model discovers cause-and-effect relationships. Causal language model and autoregressive language model are often used interchangeably: more precisely, causal describes the permitted information flow, while autoregressive describes the factorization or generation process.

other attested French terms: modèle de langage causal

See also autoregressive generationdecoder-only modelattention masknext-token prediction

Context promptinglanguage models

A sequence of intermediate, reasoning-like steps generated by a language model as it works toward an answer.

Notes Such a sequence can be elicited through chain-of-thought prompting, for example by providing demonstrations containing worked-out steps.

See also reasoning modelreasoning tokens

Context regular expressions

A regular-expression construct written in square brackets that matches exactly one character from the listed set or range, optionally negated with ^.

See also regular expressionmetacharacter

Context text encoding

A convention for representing the characters in a text as sequences of bytes so that the text can be stored, transmitted, and reconstructed. The same bytes may produce different characters when interpreted using different encodings, so the declared encoding must correspond to the encoding actually used to save the file. Character encoding should not be confused with a font, which controls how characters look, or with a character repertoire, which determines which characters are available.

See also UTF-8reserved charactereXtensible Markup Language

Context machine learningNLP

A task whose answer is one of a fixed set of categories: the system is asked which of k classes an input belongs to.

Notes Spam detection, sentiment analysis, language identification and part-of-speech tagging are all classification tasks. A classifier usually produces a probability for each category, and a threshold turns that into the decision.

other attested French terms: système de classification

See also regressionsupervised (machine) learningfeature

Context evaluationlanguage models

A distinction between two criteria for judging an output: coherence, whether its claims, reasoning and conclusion are consistent with one another, and correspondence, whether the output agrees with the facts outside it.

Notes An output may therefore be coherent but factually incorrect, or reach a correct answer through inconsistent or invalid reasoning. The pair is borrowed from research on human judgment, where the same two standards are applied to people's reasoning.

See also factuality

Context corporalinguistics

A combination of words that co-occur more often than chance would predict and form a habitual, semi-fixed pairing, such as heavy rain or commit a crime (rather than strong rain or do a crime).

Notes Corpus tools identify collocations statistically, by comparing how often two words appear together with how often they would be expected to by chance. Knowing the collocations of the target language is a large part of producing idiomatic text, which makes corpus-based collocation lookup directly useful in translation.

See also concordancen-gram

Context dataweb

A nonprofit organization that regularly collects publicly accessible web data and makes the resulting archives and datasets freely available.

Notes Filtered and processed versions of Common Crawl data have been used to train many large language models.

Context corporatranslation

A corpus containing texts in two or more languages (or language varieties) matched on subject, genre or communicative function, whose components are typically written originally in each language rather than translated: press articles written independently in French and in English on the same events, for instance.

Notes Comparability is a matter of degree rather than a yes-or-no property: documents can resemble one another in content and in form to varying extents and along several dimensions, with parallel corpora at one end of that spectrum and unrelated collections at the other (Sharoff et al., 2023). In practice the components are matched on a shared sampling frame: the same proportions of the same genres and domains, over the same period, in each language (McEnery & Xiao, 2007). Because those components are usually originals rather than translations, a comparable corpus shows how each language expresses similar content natively, which is what makes it useful in terminology work and for checking that a translation reads like an original; natural language processing puts the same corpora to a different use, as training data for machine translation and cross-lingual search. The term is not applied consistently: some authors also use it for two collections in one and the same language, one of translated and one of non-translated texts, the usual design for studying how translated language differs from original language, while McEnery and Xiao reserve it for components in different languages. Corpora of this kind are increasingly augmented or generated synthetically, which reopens the question of how comparability should be defined and measured (Sharoff et al., 2023).

See also parallel corpuscorpus

Context NLPlinguistics

The interdisciplinary study of human language from a computational perspective, including the development of computational models of linguistic phenomena.

Notes The scientific side of the domain also called natural language processing or language technology; the three are used more or less interchangeably.

See also natural language processinglanguage technology

Context agentsAI

The capability of AI systems to control computer interfaces through screenshots, mouse movements, and keyboard input, to carry out a task.

See also AI agenttool use

Context CAT toolstranslation

Software that assists a human translator during the translation process, commonly by dividing texts into segments and providing functions such as translation-memory retrieval, terminology management, consistency checks and file management.

Notes Unlike machine translation, the human translator remains responsible for producing and approving the translation, although a CAT tool may incorporate machine-generated suggestions.

See also translation memorytermbasemachine translation

Context corporaterminology

A display of all occurrences of a word or expression in a corpus, each shown with its surrounding context, typically one occurrence per line with the search word aligned in the middle (the key word in context, or KWIC, format).

Notes Concordances let an analyst move from counting words to observing how they are used: their typical constructions, collocations, register and senses. The tool that produces them is called a concordancer.

See also corpuscollocation

Context ethics & societyAI

The process by which an online service reviews and acts on content to enforce legal requirements or platform rules.

Notes Actions can include labelling content, reducing its visibility, restricting access to it, or removing it, and may be performed by human reviewers, automated systems, or both.

See also ghost work

Context CAT toolstranslation

A translation-memory match in which the stored source segment is identical to the segment being translated and the segments around it match as well, so the stored translation was produced in the same surroundings.

Notes It therefore depends on segmentation twice over: on the segment itself, and on where segmentation placed its neighbours. Tools treat it as more dependable than a plain exact match and label it separately, above 100 % or under a name of their own, because the commonest reason an identical sentence needs a different translation is that it sits somewhere else in the document.

other attested English terms: in-context exact match, 101 % match

other attested French terms: correspondance exacte avec contexte

See also translation memoryexact matchfuzzy matchsegmentsegmentation

Context language models

"The amount of input (like text) that an AI system can process at one time during a task. Think of it as a model's short-term memory. A model can only consider so much text when generating its next response, and once the conversation exceeds its context window, the model starts 'forgetting' prior responses." (Stanford HAI).

See also promptin-context learning

Context embeddingslanguage models

A numerical representation generated for a particular occurrence of a token by taking account of the other tokens available to the model.

Notes The representation can therefore change according to the token's context: bank can receive different representations in river bank and bank account. In a transformer, each layer progressively combines information from relevant token positions, producing increasingly contextualized representations. Depending on the model, the available context may include preceding text, following text, or both. Unlike a static embedding, a contextual embedding is computed for a particular occurrence rather than retrieved as a single permanent representation from a vocabulary. Contextual embeddings can encode information related to a token's meaning, grammatical role and relationships with other parts of the text; different layers of a model may capture different kinds of information, so a contextual embedding should not be understood as a complete or directly interpretable representation of meaning.

other attested English terms: contextualized word representation

See also static word embeddingword embedding

Context lawethics & society

"Copyright infringement occurs when a person uses content protected by copyright in a way that violates rights granted in the Copyright Act, for example by copying or communicating to the public a copyrighted work (all or a substantial part of it) without authorization from the owner of the copyright." (Government of Canada).

See also fair use

Context corporaNLP

A collection of written texts, speech recordings or transcribed language assembled for analysis or use in language technology (plural: corpora).

Notes In corpus linguistics, a corpus is typically machine-readable and selected and documented according to explicit criteria so that it can support research on a particular language or language variety. In NLP, the term is often used more broadly for a collection of language data used to train, develop or evaluate a system, even when the collection was not designed to be balanced or representative.

See also parallel corpuscomparable corpusconcordance

Context deep learninglanguage models

An attention mechanism in which the queries come from one sequence or component and the keys and values come from another.

Notes In an encoder-decoder model, cross-attention allows the decoder to select information from the encoder's representations while producing the output. Cross-attention can also connect text with image, audio or other representations in multimodal systems. In self-attention, queries, keys and values are derived from the same sequence; in cross-attention, they connect different sequences or components. Standard decoder-only language models have no separate encoder and therefore do not use the encoder-decoder form of cross-attention, although multimodal variants may add cross-attention modules.

See also self-attentionencoder-decoder modeldecodermultimodal model

Context prompting

User-provided preferences, background information, or requirements that an AI assistant stores and applies automatically across conversations, so that they do not need to be repeated in every prompt.

See also system promptpersistent memory

Context dataNLP

The process of adding labels or other structured information, such as categories, tags, transcriptions, or marked regions, to data so that it can be used to train or evaluate a machine learning system.

Notes Annotation is often carried out by people, sometimes with machine assistance, and may involve judgement when the appropriate labels are not unambiguous.

See also ground truthgold standardghost work

Context environmentcomputing

A facility that houses computing, storage, and networking equipment, together with supporting infrastructure such as power supplies, cooling systems, and security systems.

Notes Data centres provide the computing resources used to train and run AI models, although AI is only one of the workloads they support.

other attested French terms: centre de données

See also training emissionsinference emissions

Context evaluationmachine learning

A problem that occurs when some of the data used to evaluate a model also appears in its training data.

Notes Because the model may already have encountered the evaluation examples, its benchmark results can overstate how well it performs on genuinely unseen material.

other attested English terms: contamination

See also benchmarktest set

Context data

The operation of searching for data according to precise criteria, in a database or another source, and isolating them in order to remove them from the set they belong to, or to copy or migrate them to another database (translated from the OQLF's French definition).

See also web scraping

Context deep learning

A neural-network component or architecture used to generate an output sequence.

Notes During generation, a decoder processes the tokens already available and produces a probability distribution over possible next tokens. Its attention mechanism is normally restricted so that a prediction cannot use tokens that have not yet been generated. In an encoder-decoder model, the decoder also draws on a representation of the input produced by a separate encoder, an arrangement commonly used for tasks such as machine translation. In a decoder-only model there is no separate encoder: the prompt and the previously generated tokens are processed as a single sequence, and most contemporary general-purpose generative language models use this architecture. The decoder calculates the next-token probabilities; a decoding method, such as greedy decoding or top-p sampling, determines how an output token is selected from those probabilities.

See also encoderdecodingdecoder-only modelencoder-decoder model

Context language modelsdeep learning

A transformer language-model architecture that uses a sequence of decoder blocks without a separate encoder.

Notes At each token position, causal masking allows the model to use earlier token positions but prevents it from using later positions in the same sequence. During text generation, the model uses the prompt and the tokens already generated to predict a probability distribution for the next token, selects a token through a decoding method, adds it to the sequence and repeats the process. This architecture is widely used for general-purpose generative language models. "Earlier" refers to earlier token positions in the modelled sequence, not necessarily to words visually appearing on the left, and causal masking does not mean the prompt must be processed one token at a time: its token representations can be computed in parallel while respecting the restriction on access to later positions. Unlike a decoder inside an encoder-decoder model, a decoder-only model normally has no cross-attention layer for reading a separate encoder's output.

See also decoderdecodingautoregressive generationcausal language modelencoder-only modelencoder-decoder model

Context language models

The process used during text generation to turn a language model's predicted probabilities into an actual sequence of output tokens.

Notes The most common decoding methods are greedy decoding and sampling. Decoding is a different thing from the decoder, the part of some neural networks that generates an output sequence.

See also decodersamplinggreedy decodingbeam searchtemperature

Context machine learningAI

An approach to machine learning based on neural networks with many successive layers of computing units.

Notes The term deep refers to this layered architecture, in which each layer processes information received from the preceding layer.

Context machine learning

A training approach in which input data are deliberately corrupted and a model learns to reconstruct the original or otherwise clean version.

Notes For text, corruption may involve masking, deleting, replacing, inserting or reordering tokens; masked language modeling is one form of denoising. Noise need not be random visual or acoustic interference: in machine learning, any deliberate alteration that the model is trained to reverse may function as noise.

other attested French terms: apprentissage par débruitage

See also masked language modelself-supervised learning

Context machine learning

A vector in which most or all components are non-zero real numbers.

Notes In natural language processing, dense vectors are often relatively short, learned representations of words or other tokens. Their individual dimensions usually have no obvious interpretation, but the vector as a whole can represent useful similarities between items.

See also sparse vectorword embedding

Context computingAI

Putting a system into operation where it will be used: installing it with what it needs to run and making it available to its users. For an AI system, it means making a trained model available to people, for instance behind an application or a chat interface.

Notes For a language model, deployment changes nothing in its parameters: system prompts, output filters and access rules are set around the model, and they can be changed far more often than the model itself. Deployment tasks include piloting the system, checking compatibility with existing systems and ensuring regulatory compliance (Tabassi, 2023).

See also system promptpost-traininginference

Context computing

Describes a process, algorithm, or system that, under the same initial conditions and with the same input, always produces the same output.

Notes Chatbots based on large language models are generally not deterministic: the same prompt can give different outputs.

See also temperaturesampling

Context ethics & society

Inequalities between people, communities, organizations, or regions in their access to digital technologies, their ability to use those technologies effectively, and the benefits they obtain from their use.

See also linguistic hegemony

Context language modelsmachine learning

A preference-tuning method that trains a language model directly from pairs of preferred and rejected responses.

Notes Training increases the relative probability of preferred responses while limiting how far the resulting model moves from a reference model. Unlike common RLHF methods, DPO does not require a separately trained reward model or a reinforcement-learning stage that repeatedly samples from the model being updated. DPO eliminates the explicit reward model, not the assumptions underlying preference modelling: its results still depend on the quality, coverage and values represented in the preference data.

See also preference tuningreinforcement learning from human feedbackreward modelpost-training

Context machine learning

A way of representing an item as a pattern of values spread over many units or dimensions, a vector for example, where each unit also takes part in representing other items.

Notes The opposite is a local representation, where one unit stands for one item and for nothing else. A one-hot vector is exactly that, so it is not a distributed representation: one position carries the word, that position carries nothing else, and every other position is a zero. A word embedding is one: the word becomes a pattern across a few hundred values, each of which also takes part in the pattern of every other word, so no dimension can be read on its own and only whole vectors can be compared.

See also distributional hypothesisword embeddingone-hot vector

Context embeddingslinguistics

The linguistic hypothesis that patterns of word use provide evidence about meaning: words that occur with similar neighbouring words are likely to have related or similar meanings.

Notes Many methods for learning word embeddings from text are based on this principle.

See also word embeddingdistributed representation

Context XML

The building block of an XML document: an opening tag, the content it delimits and a closing tag, or a single empty-element tag when there is no content.

Notes Elements nest inside one another and form the document's tree.

See also eXtensible Markup Languagetagattribute

Context embeddings

The use of addition and subtraction on embedding vectors to test whether a semantic relationship is represented as an approximately consistent direction or offset in an embedding space.

Notes A familiar illustration is that the vector produced by king − man + woman may lie near the vector for queen.

See also word embeddingembedding space

Context embeddings

A vector space into which an embedding method maps items such as words, sentences, documents, or images.

Notes Each item is represented by a vector, and the geometry of the space is constructed or learned so that proximity or other geometric relationships reflect patterns useful for the intended task.

other attested French terms: espace de plongement vectoriel

See also latent spacevector spaceword embedding

Context language models

An ability is described as emergent when a model shows little or no measured success on a task at smaller scales but substantially better performance at a larger scale, in a way that is not predicted by simply extrapolating the smaller models' results.

Notes Apparent emergence can depend on how performance is measured and therefore does not necessarily demonstrate that an entirely new ability appeared suddenly.

See also scaling laws

Context deep learning

The component of a neural network that processes an input and converts it into internal representations that can be used by another component or by a task-specific output layer.

Notes In a transformer encoder, self-attention allows the representation of each input token to incorporate information from other token positions. An encoder may form part of an encoder-only model or supply representations to the decoder in an encoder-decoder model. Saying that an encoder "encodes the meaning" of the input can be convenient but overstates what is established; more precisely, it computes numerical representations learned to be useful for the model's training objective and subsequent tasks.

See also decoderencoder-only modelencoder-decoder modelcontextual embedding

Context deep learninglanguage models

A neural-network architecture in two connected parts: an encoder that turns an input sequence into representations, and a decoder that uses them, with what it has produced so far, to generate an output sequence.

Notes Because the output can differ from the input in length, word order and vocabulary, the architecture suits tasks that turn one sequence into another, such as machine translation, summarization and speech recognition. It is older than the transformer: the first neural translation systems built it from recurrent networks.

See also encoderdecodercross-attentionmachine translationsequence-to-sequence modeltransformer

Context language modelsdeep learning

A neural-network architecture that processes an input sequence and produces a contextual representation for each input token, without using a separate decoder to generate an output sequence.

Notes In the common bidirectional transformer form, each token can be represented using information from positions both before and after it in the input. Encoder-only models are therefore especially useful for analyzing or representing text, for example in classification, information extraction, semantic similarity and search; BERT is a well-known example. These models are generally not designed to generate extended text one token at a time, although encoder-based components can be incorporated into generative systems.

See also encoderdecoder-only modelencoder-decoder modelcontextual embeddingmasked language model

Context ethics & societyAI

A serious disruption of the shared practices, standards, and institutions through which a society produces, verifies, and recognizes knowledge.

Notes It is marked by deep disagreement about which sources and methods are trustworthy, weakening the common factual basis needed for public judgment and collective decision-making.

See also information pollutionknowledge collapse

Context CAT toolstranslation

A translation-memory match in which the stored source segment is identical to the segment being translated, character for character, so the tool can insert the stored translation as it is.

Notes The memory is searched segment by segment, so what counts as identical depends on the segmentation applied to both texts: the same sentence divided differently is not found. Identical source text does not guarantee that the same translation still fits: the same sentence may need to be rendered differently somewhere else in a document, which is what a context match adds and an exact match does not.

other attested English terms: full match, 100 % match

other attested French terms: correspondance à 100 %

See also translation memoryfuzzy matchcontext matchsegmentsegmentation

Context XMLmarkup

A markup language and W3C standard for storing structured data as plain text: elements delimited by tags, with attributes on them, nested in a hierarchy.

Notes The translation interchange formats TMX, TBX and XLIFF are all XML-based.

See also markup languagetagattribute

Context language modelsevaluation

Generated content that can be neither supported nor contradicted by the source or context against which the output is being evaluated.

Notes The content may be factually true or false; determining that requires evidence beyond the supplied source.

other attested French terms: hallucination extrinsèque

See also hallucinationintrinsic hallucination

Context evaluationNLP

The harmonic mean of precision and recall: a single number between 0 and 1 that is high only when both are high.

Notes Between two annotators of the same text, it also serves as a measure of their agreement.

other attested English terms: F-score, F-measure

other attested French terms: mesure F1

See also precisionrecallinter-annotator agreement

Context evaluationlanguage models

The extent to which the factual claims in generated content are correct according to reliable evidence about the world.

See also hallucinationcoherence-correspondence distinction

Context law

Related copyright doctrines that permit certain uses of copyright-protected material without the rightsholder's authorization.

Notes Although comparable, the two doctrines have different legal requirements and are not interchangeable.

other attested English terms: fair dealing

See also copyright infringement

Context evaluationNLP

An item that the gold standard contains but that a system did not return, or did not assign to the class it looks for: for example, a term that was not proposed, or spam that reaches the inbox.

Notes Taken together, the false negatives are also called silence. They lower recall.

See also silencetrue positivefalse positiverecall

Context evaluationNLP

An item that a system returned, or assigned to the class it looks for, but that the gold standard does not contain: for example, a candidate that is not a term, or a legitimate message stopped by a spam filter.

Notes Taken together, the false positives are also called noise. They lower precision.

See also noisetrue positivefalse negativeprecision

Context machine learning

A property or representation of an observation that a machine-learning model uses as input when making a prediction.

Notes In natural language processing, features may include word counts, sentence length, grammatical categories, or learned numerical representations.

See also feature engineeringrepresentation learning

Context machine learning

The process of selecting, creating, or transforming information from raw data into features that a machine-learning model can use effectively.

Notes Traditionally, much of this work was performed by people; representation learning enables models to learn many useful features automatically.

See also featurerepresentation learning

Context machine learning

A machine-learning approach in which a model learns or adapts to perform a new task, or to recognize new categories, using only a small number of relevant examples, typically by drawing on knowledge acquired from other data or tasks.

other attested French terms: apprentissage avec peu d'exemples

See also few-shot promptingzero-shot learningin-context learning

Context prompting

A prompting technique in which a small number of demonstrations, typically examples pairing an input with its desired output, are included in the prompt to show a language model how to perform a task.

Notes The model uses these examples at inference time without updating its parameters.

other attested French terms: requête avec quelques exemples

See also few-shot learningzero-shot promptingin-context learning

Context machine learninglanguage models

Further training of a pretrained model on new data to adapt it for a particular task, domain, or intended behaviour.

Notes Fine-tuning may adjust all the model's parameters or only a selected subset and generally requires much less data and computation than pretraining.

other attested French terms: réglage fin, peaufinage

See also instruction tuningsupervised fine-tuningpost-training

Context regular expressionscomputing

In programming, an option turned on or off for a whole operation, changing how it is executed.

Notes Flags are added to a regular expression as letters written after the pattern, in any order and any combination: g finds every match instead of only the first, i ignores case, m makes ^ and $ match at the start and end of each line rather than of the whole text, s lets the dot match a line break, u turns on full Unicode support.

other attested French terms: indicateur

See also regular expressionanchor

Context machine learningAI

A pre-trained model that serves as the starting point for multiple downstream applications.

Notes The term emphasizes the "train once, adapt many times" approach to AI development.

other attested French terms: modèle fondateur

See also base modelpre-training

Context agentslanguage models

A mechanism that allows a language model to request the use of an external function or tool by producing its name and required arguments in a structured form.

Notes The surrounding application validates and executes the function and may return its result to the model, which can then use it in generating a response.

See also tool usestructured outputapplication programming interface

Context CAT toolstranslation

A translation-memory match in which the stored source segment resembles the segment being translated without being identical to it, reported as a percentage below 100. The comparison is made segment by segment, so the percentage measures the distance between two segments as segmentation produced them, not between two texts.

Notes The tool proposes the stored translation and shows where the two source segments differ, and the translator adapts it. Tools apply a threshold below which no match is offered, on the reasoning that editing a distant match costs more than translating the segment afresh.

other attested English terms: partial match

other attested French terms: correspondance approximative, correspondance floue

See also translation memoryexact matchcontext matchsegmentsegmentation

Context computingdata

The principle that the quality and reliability of a computer system's outputs depend substantially on the quality and suitability of its inputs.

Notes In AI, inaccurate, incomplete, biased, outdated, irrelevant, or unrepresentative data can lead to unreliable or misleading results.

Context AI

An area of artificial intelligence concerned with systems that generate content such as text, code, images, audio, or video by learning patterns from training data and producing new outputs in response to input or prompts.

Context language models

A transformer-based generative model, originally developed by OpenAI for language processing, that is pretrained on large quantities of text to predict successive tokens and can subsequently be prompted or further trained to perform a wide range of tasks.

Notes GPT also names OpenAI's family of such models.

Context ethics & societyAI

Human labour that remains hidden behind digital services presented as automated.

Notes In AI, it includes work such as collecting, labelling, or classifying data, moderating content, evaluating model outputs, and completing tasks that automated systems cannot handle.

See also data annotationcontent moderation

Context evaluationdata

The best available reference dataset, annotation, measurement or procedure against which a system's results are evaluated.

Notes In natural language processing, usually a set of labels produced or adjudicated by people. A gold standard is not necessarily correct: annotations can contain mistakes, and a single adjudicated answer can hide real differences between annotators.

other attested French terms: test de référence, référence standard, jeu de référence, référence

See also ground truthdata annotation

Context computingdeep learning

A processor with a large number of cores that work in parallel, designed to render images, and now the standard hardware for training and running neural networks.

Notes The arithmetic behind rendering a scene and the arithmetic behind a neural network are much the same: many small multiplications carried out at once. A general-purpose processor does a few operations very fast in sequence, where a GPU does thousands at a time, which is what suits it to both tasks. Training a large model occupies thousands of them at once for weeks or months, and their number, their power draw and their price are what put the cost of a training run where it is.

other attested French terms: unité de traitement graphique

See also pre-trainingneural networkdata centre

Context language models

A decoding method that selects the single most probable next token at every generation step. The selected token is added to the sequence, and the procedure repeats until generation stops.

Notes Greedy decoding is simple and deterministic under fixed conditions, but its locally best choice at one step does not necessarily produce the most probable or best complete sequence. It can produce repetitive or generic text, and it differs from sampling in that it introduces no deliberate randomness.

other attested French terms: décodage avide

See also decodingsamplingbeam searchtop-k sampling

Context evaluationdata

The values, labels or observations treated as the correct description of the cases in a dataset.

Notes Often used as a synonym of gold standard, but the two describe different things: a gold standard is data that people produced, and a ground truth is answers that are verifiable facts. Annotated data are a ground truth only where the answers can be verified.

See also gold standarddata annotation

Context language modelsevaluation

Content generated by an AI model that is factually incorrect or fabricated, even though it may sound plausible.

Notes For example, a language model may invent people, events, quotations, references, or other details that do not exist. The term is often criticized as anthropomorphic: it suggests faulty perception, while a model perceives nothing. Hicks et al. (2024) argue that "bullshit", in Frankfurt's sense of speech produced with indifference to truth, describes it better. Hallucination is nonetheless the term the field adopted, and it is the one used here.

other attested French terms: hallucination de l'IA

See also intrinsic hallucinationextrinsic hallucinationfactuality

Context agentslanguage models

The software around a language model that turns it into a working system. It determines what the model receives and which tools it may use, carries out the actions the model requests, and passes the results back to it.

Notes The harness covers everything in such a system except the model: the system prompt, which tools are available and how they are described, memory, permissions, and checks on the output. The model chooses which of the available tools to call at each step (Schick et al., 2023). Two harnesses around the same model can produce very different results (Pan et al., 2026), and improving the harness, called harness engineering, has become a research area of its own. The term is used above all for AI agents, but a chat application also has a harness.

other attested English terms: agent harness

See also function callingagentic AIsystem prompt

Context ethics & societyAI

The tendency of an AI model to portray members of some social groups, particularly marginalized or socially subordinate groups, as more similar to one another and less individually varied than members of socially dominant groups.

See also training biasbias in AI

Context evaluationAI

A method for evaluating AI systems in which people compare two or more outputs and indicate which they prefer, usually in response to the same prompt and according to stated criteria.

Notes The resulting judgments can be aggregated to compare or rank responses, models, or systems according to perceived quality or usefulness.

See also benchmarkpreference tuning

Context machine learning

A setting that controls a model's design or training process but is not learned from the training examples in the same way as the model's parameters.

Notes Examples include the learning rate, batch size, number of layers and degree of regularization. Hyperparameter values are usually selected by comparing performance on a validation set. A hyperparameter need not be chosen manually: automated optimization methods can search for suitable values. What distinguishes it from an ordinary parameter is how its value is selected, not whether a person types it in directly.

See also parametervalidation set

Context markupweb

The markup language of web pages: a vocabulary of elements laid down by a standard, marking headings, paragraphs, links, images and the rest, which a browser renders.

Notes Its element names come from the standard, whereas an XML format defines its own.

other attested French terms: langage de balisage hypertexte

See also markup languageeXtensible Markup Languagetag

Context promptinglanguage models

The temporary adaptation of a language model's behaviour to information, instructions, or examples provided in its current prompt or context, without changing the model's parameters.

Notes This can help the model perform a new task or make better predictions during the current interaction, but the adaptation does not persist after the context is removed.

See also few-shot promptingcontext window

Context machine learning

The stage in which a trained AI model processes new input to produce a prediction, classification, decision, or generated output.

Notes In a language model, inference includes processing the prompt and generating an output, usually one token at a time.

Context environmentAI

The greenhouse gas emissions associated with using a trained AI model to process inputs and produce predictions or generated outputs.

See also training emissionscarbon footprintdata centre

Context ethics & societyAI

The degradation of an information environment through the accumulation and circulation of content that is false, misleading, irrelevant, redundant, unsolicited, or otherwise of low value.

See also AI slopepistemic crisis

Context CAT toolsmarkup

A marker inside a segment standing for formatting or a code of the original document, a bold run or a footnote reference for instance, rather than for text to translate.

Notes Each interchange format defines its own inline elements: bpt, ept and ph in TMX; pc, sc, ec and ph in XLIFF.

See also tagTranslation Memory eXchangeXML Localization Interchange File Format

Context language models

A form of supervised fine-tuning in which a pretrained language model is trained further on examples pairing instructions or questions with appropriate responses.

Notes This makes the model better able to follow instructions across a range of tasks, including tasks not represented exactly in the training examples.

other attested English terms: instruction fine-tuning

See also fine-tuningsupervised fine-tuningpost-training

Context evaluationdata

How far two or more people who annotate the same data independently give the same answers.

Notes A measure of how reliable a gold standard is. Low agreement means that the task, the guidelines or both leave room for judgement.

other attested English terms: inter-rater agreement, inter-coder agreement

other attested French terms: accord inter-évaluateurs, accord interannotateurs

See also data annotationgold standardF1 score

Context exchange formatsdata

A documented way of organizing and representing data so that it can be transferred between different systems or applications. In translation technology, TMX is primarily used to exchange translation-memory data, TBX to exchange terminological data, and XLIFF to move localizable content, translations, and related workflow information. A standardized format facilitates interoperability, but successful exchange may still depend on the versions, optional features, and metadata supported by each tool.

other attested English terms: data exchange format

other attested French terms: format d'échange de données

See also interoperabilityTranslation Memory eXchangeTermBase eXchangeXML Localization Interchange File Format

Context exchange formatstranslation

The ability of two or more systems, applications, or tools to exchange information and make use of what is exchanged.

Notes In translation technology, interoperability allows resources such as bilingual texts, translation memories, terminology records, and project data to move between tools, often through shared formats or interfaces.

See also Translation Memory eXchangeTermBase eXchangeXML Localization Interchange File Format

Context language modelsevaluation

Generated content that uses or combines information from the provided source or context in a way that misrepresents or contradicts that information.

Notes Unlike an extrinsic hallucination, it distorts available source information rather than introducing information that the source does not support.

other attested French terms: hallucination intrinsèque

See also hallucinationextrinsic hallucination

Context AI

The uneven and often difficult-to-see boundary between tasks that an AI system can perform well and tasks that it handles poorly.

Notes Tasks that appear similarly difficult to people may lie on opposite sides of this boundary, so AI assistance can improve performance on one task while harming it on another.

other attested French terms: frontière technologique en dents de scie

Context AI safetyAI

The deliberate use of specially crafted prompts or interaction strategies to make an AI model bypass restrictions on its outputs or behaviour, such as its refusal to provide prohibited content or sensitive information.

other attested French terms: débridage d'IA

See also prompt injectionred-teaming

Context ethics & societyAI

A hypothesized process in which widespread reliance on AI-generated or AI-mediated information concentrates attention on common, centrally represented ideas, while rarer, specialized, or unconventional knowledge is consulted, transmitted, and developed less often.

Notes Over time, this can narrow the range of knowledge that a community considers accessible or worth pursuing.

See also model collapseepistemic crisis

Context language modelsdata

The point in time up to which the text a model was trained on was gathered, and after which nothing reached it during training.

Notes Creators who do not release their training data often publish a cutoff date in its place. That reported date can differ from the effective cutoff, the point up to which the model actually shows knowledge of a given source, and the two often differ by a long way: a new web crawl carries a substantial amount of old material, so a resource can be years staler than the date advertised (Cheng et al., 2024). A cutoff limits what training put into the model, not what a system can answer, since text retrieved from the web or from a document you upload reaches the model through its context instead.

other attested English terms: knowledge cut-off

other attested French terms: date de coupure

See also training corpusCommon Crawlpre-training

Context language modelsNLP

A machine-learning model that estimates how likely different sequences of language are.

Notes Given a sequence of tokens as context, it assigns probabilities to the tokens that could come next. These probabilities can be used to predict or generate text and to compare the likelihood of complete sequences.

other attested French terms: modèle de langage

See also large language modelcausal language modelmasked language model

Context NLPtranslation

The field that builds usable tools and services out of the computational processing of human language, such as machine translation systems, terminology tools, speech interfaces and writing aids. By extension, the term also names those tools themselves.

Notes The applied, product-facing side of the domain also called natural language processing or computational linguistics; the three are used more or less interchangeably. In Canada the plural also names an industry, and the federal research centre devoted to it is the Centre de recherche en technologies langagières.

See also natural language processingcomputational linguistics

Context language modelsAI

A neural-network language model, usually based on the transformer architecture, that is trained on very large amounts of text and contains many learned parameters.

Notes Most current generative LLMs process text as tokens and generate text one token at a time by repeatedly assigning probabilities to possible next tokens. When incorporated into an application, an LLM can support tasks such as answering questions, summarizing, translating, and conducting conversations.

See also language model

Context speechcomputing

The delay between an input and the system's response; in live speech recognition and interpreting, the time between the spoken words and their displayed transcription or translation.

other attested French terms: temps de latence

Context machine learning

A space in which a model represents data using latent, that is, not directly observed, features learned from the data.

Notes Each point or vector encodes underlying patterns identified by the model, often in a representation that is more compact than the original data. Related to, but not the same as, an embedding space.

See also embedding spacevector space

Context linguisticscorpora

The base or dictionary form under which the inflected forms of a word are grouped: go, goes, went and gone share the lemma go.

Notes Corpus tools use lemmatization to count or search all forms of a word at once; a frequency list of lemmas answers a different question than a frequency list of word forms.

other attested French terms: forme canonique

See also typetoken¹

Context text encoding

The character that ends a line in a file, stored in it like any other and invisible on screen. It is what divides a file into lines, so a file's line count is fixed: two of them make three lines, whatever width the window is. Windows writes it as two characters, a carriage return followed by a line feed (CR LF), where macOS and Linux use one (LF).

other attested English terms: hard return, newline

other attested French terms: retour à la ligne forcé, retour forcé, présentation de ligne

See also word wrapplain textcharacter encoding

Context ethics & societylinguistics

The dominance of one language or a small group of languages in institutions, communication, and knowledge production, such that their use is treated as normal or more valuable while other languages receive less recognition, resources, or support.

Notes In AI, linguistic hegemony can be reinforced when training data, benchmarks, development resources, and model performance are concentrated in dominant languages, especially English, disadvantaging speakers and knowledge expressed in underrepresented languages.

See also digital divide

Context computing

A method of reducing the amount of data needed to represent information by discarding some information, so that the original cannot be reconstructed exactly.

Notes LLM pretraining is sometimes described metaphorically as lossy compression of the training corpus because the model's parameters retain statistical patterns and associations from the text without preserving the corpus as a complete, exact archive.

Context machine learningAI

An approach to artificial intelligence in which a model is trained on data or experience so that its adjustable parameters come to reflect patterns relevant to a task.

Notes The trained model can then apply those patterns to new inputs, for example to make predictions or classifications. The task-specific patterns are learned from data rather than fully specified in advance as hand-written rules.

See also supervised (machine) learningunsupervised learningself-supervised learning

Context translationAI

The use of a computer system to produce content in a target language that conveys content expressed in a source language.

Notes The system produces the translation itself, unlike a computer-assisted translation tool, which supports a human translator and may include a machine translation system as one of its components. The output is used as is, or as a draft that a translator revises.

See also computer-assisted translationpost-editingencoder-decoder model

Context ethics & societyAI

The extent to which sales, output, or another relevant measure of activity in a defined market is accounted for by a relatively small number of firms.

Notes It is commonly assessed from firms' market shares using measures such as concentration ratios or the Herfindahl-Hirschman Index (HHI).

See also oligopoly

Context markupweb

A convention for annotating text with tags that mark its structure or presentation, keeping the annotations distinct from the content itself; HTML and XML are the best-known examples.

See also eXtensible Markup Languagetag

Context language models

A language model commonly built using a bidirectional transformer encoder and pretrained to reconstruct selected tokens that have been hidden or otherwise altered in a text.

Notes To predict an original token, the model can use the visible context both before and after its position. This self-supervised training process enables the model to learn contextual representations of words and passages without requiring every training example to be labelled by people; the pretrained model can then be adapted to tasks such as text classification, information extraction, semantic similarity and question answering. BERT and RoBERTa are well-known masked language models. Not every selected token is necessarily replaced by a visible [MASK] token: in the original BERT procedure, some are replaced by [MASK], some by a randomly chosen token and some are left unchanged, and in all three cases the model is trained to predict the original. A masked language model is not synonymous with an encoder-only model, and these models are generally used to produce contextual representations rather than to generate extended passages autoregressively, although they can fill masked positions and newer generative methods can build on masking.

other attested French terms: modèle de langage masqué

See also masked language modelingencoder-only modelbidirectional language modelcontextual embeddingself-supervised learningdenoising

Context language models

A self-supervised training objective in which selected tokens in a text are hidden or otherwise altered and a model is trained to reconstruct the original tokens from the remaining context.

Notes In the usual bidirectional form, the prediction can draw on visible tokens occurring both before and after each selected position.

See also masked language modelself-supervised learning

Context regular expressions

A character with a special meaning inside a regular expression (such as ., *, [ or ^) instead of matching itself literally; a backslash escape turns it back into a literal character.

See also regular expression

Context cognitive scienceAI

Awareness and monitoring of one's own cognitive processes, often combined with attempts to regulate them.

Notes It includes judging what one knows, assessing confidence, noticing possible errors, and changing strategies when necessary. In AI research, the term is sometimes used by analogy for mechanisms or behaviours through which a system monitors or regulates aspects of its own processing, such as estimating uncertainty, checking an answer, or changing strategy; such behaviour does not by itself demonstrate consciousness or human-like self-awareness.

Context data

Structured information that describes, identifies, or provides context for other data. In a marked-up document or translation format, metadata may indicate the language, author, creation date, file type, segment identifier, translation status, subject field, or origin of the content. Metadata may be stored in elements or attributes. It is generally distinguished from the main content to be translated, although some metadata values may themselves require localization.

See also attributeelementTermBase eXchangeXML Localization Interchange File Format

Context language modelsdata

A degenerative process in which successive generations of a generative model become less faithful to the original data because they are trained recursively on data produced by earlier models.

See also synthetic content

Context AIlanguage models

A detailed document in which a developer of AI models sets out how its models should behave, as behavioural guidelines and ethical principles for their responses.

Notes The best-known examples are OpenAI's Model Spec and Anthropic's constitution for Claude; the related term AI constitution is also used (Zhang et al., 2025c). A specification states the behaviour a developer intends, so it is one of the few ways to compare how a model actually behaves with what its designers wanted; how much effort a developer puts into making its models follow it cannot be known from outside (Lambert, 2026). Specifications studied in 2025 contained principles that contradicted each other in some situations and left other situations uncovered (Zhang et al., 2025c).

other attested English terms: model spec

See also alignment¹system promptpreference tuning

Context AImachine learning

The 37th move played by the Go program AlphaGo in the second game of its 2016 match against Lee Sedol, one of the strongest players in the world: a move so far outside human playing convention that one of the match's two commentators first took it for a mistake, and that proved central to winning the game.

Notes The name has become shorthand for two related ideas: a system trained by reinforcement learning discovering a valuable action no human would have chosen, and the moment an observer realizes that an ability they had considered inherently human is not exclusively human.

See also reinforcement learning

Context AI

An AI model that processes and relates information from two or more data modalities, such as text, images, audio, video, or sensor data.

Notes Depending on its design, it may accept multiple modalities as input, produce outputs in one or more modalities, or both.

Context corporaNLP

A sequence of n consecutive items from a text, usually words: bigrams (n = 2), trigrams (n = 3), and so on.

Notes Frequent n-grams reveal recurring multi-word expressions. Before neural networks, n-gram models were the standard way to estimate the probability of a word from the n − 1 words before it.

See also collocationlanguage model

Context NLPterminology

The task of finding the spans of a text that are proper names, and labelling the category, most often a person, a location or an organization.

Notes A named entity is, roughly, anything that can be referred to with a proper name, and the label is commonly extended to dates, times and amounts. Labelling the category is part of the task as it is standardly defined; finding the spans without typing them is a separate, narrower task. Two things make it hard: deciding where an entity starts and stops, and deciding which type it is, since the same name can be a person, an airport or a school.

See also featureclassificationcorpus

Context XML

A name, written as a URI, that says which vocabulary an element or attribute belongs to.

Notes Declared with an xmlns attribute, it lets one document mix vocabularies, a standard format and a tool's own extensions for instance, without name collisions.

other attested French terms: espace de nommage

See also eXtensible Markup Languageelementattribute

Context NLP

An interdisciplinary field that develops and studies computational methods for analyzing, processing, and generating human language.

Notes The engineering side of the domain also called computational linguistics or language technology; the three are used more or less interchangeably.

See also computational linguisticslanguage technology

Context machine learningAI

A machine-learning model composed of interconnected layers of small computational units.

Notes Each unit combines input values using adjustable numerical weights and passes the resulting value to other units. During training, the network adjusts its weights to improve its performance on a task. The name reflects the field's historical inspiration from biological neurons; modern neural networks are mathematical computing systems, not realistic models of how the human brain works.

other attested French terms: réseau de neurones artificiels

Context language models

The task of estimating the probabilities of the tokens that could follow a given sequence of tokens.

Notes It is a common training objective for autoregressive language models. During text generation, the model repeatedly produces this probability distribution, a decoding method selects a token, and the selected token is added to the context for the next prediction.

other attested English terms: next-word prediction

See also autoregressive generationprobability distributioncausal language model

Context evaluationNLP

The items a system returns that are not correct: in term extraction, the candidates that are not terms.

Notes Noise lowers precision. The word comes from information retrieval, where it names the irrelevant documents a search returns.

See also false positivesilenceprecisioncandidate term

Context NLPcorpora

The processing that reduces the variation in a text that is irrelevant to a task, for instance by putting every word in lowercase, removing accents, unifying quotation marks or grouping the forms of a word, so that fewer distinct forms remain to count or to match.

Notes Every normalization loses information, so the operations are chosen per task and stated with the results: a count made after lowercasing is not comparable to one made without.

other attested English terms: text normalization

See also tokenization¹typelemmastemmingstop word

Context ethics & society

A market structure in which a small number of firms account for most of the supply of a product or service.

Notes Because the decisions of each firm can materially affect the others, firms in an oligopoly tend to make decisions with their competitors' likely responses in mind. In discussions of AI, the relevant market should be specified: oligopolistic conditions may occur in markets such as advanced AI chips, cloud computing, or foundation-model development, but it is generally too broad to describe the entire AI sector as a single oligopoly.

See also market concentration

Context machine learningNLP

A vector whose elements are all 0 except for a single 1. The position of the 1 identifies one item in a fixed set, such as a category or a token in a vocabulary.

Notes Unlike an embedding, a one-hot vector identifies an item but does not represent similarities or relationships between items.

Context AI

An AI model whose trained parameters, commonly called weights, are made available for others to download.

Notes What may be done with the weights depends on the licence they are released under. Open weights are not open-source AI, which also requires the training code and detailed information about the training data.

Context computingregular expressions

The rule fixing the order in which the operators of an expression are applied. In a regular expression that order runs from the quantifiers, which attach to the single element in front of them, through the elements written one after another, to the vertical bar, which has the lowest precedence of all and so separates everything on its left from everything on its right. Parentheses are applied before any of it, which is why /^(cat|dog)$/ and /^cat|dog$/ do not mean the same thing.

other attested English terms: precedence

other attested French terms: préséance des opérateurs, priorité

See also alternationcapturing groupquantifier

Context language modelsNLP

Describes a word that is absent from the vocabulary used by a particular language-processing system.

Notes In systems with a fixed word-level vocabulary, such a word cannot be represented directly and is often replaced by a special unknown-word symbol such as <UNK>. Modern subword or byte-level tokenizers can usually represent an unfamiliar word as a sequence of smaller known tokens, so a word may be out of vocabulary as a whole without being unprocessable.

See also vocabularysubword tokenization

Context corporatranslation

A corpus of texts in one language together with their translations into one or more other languages, usually aligned sentence by sentence so that each segment can be viewed next to its equivalent.

Notes Parallel corpora such as Europarl or the Canadian Hansard record real translation decisions, which makes them useful for observing equivalences, building translation memories and training machine-translation systems. A translation memory is in effect a parallel corpus built from one's own translations.

See also comparable corpuscorpustranslation memorymachine translation

Context computing

The performance of two or more computational operations at the same time, typically by distributing the work across multiple processor cores or computing devices.

Notes In AI, it can reduce computation time by allowing independent calculations or portions of data to be processed simultaneously rather than sequentially. During transformer training, many computations involving the tokens in an input sequence can be performed in parallel; the model's ability to represent relationships between distant tokens comes from its attention mechanism, not from parallel processing itself.

Context machine learning

A numerical value learned during a model's training and used in its computations.

Notes Parameters are made up of two kinds of number, weights and bias terms. Their total number is commonly used as one measure of model size.

See also weightbias termhyperparameter

Context NLPcorpora

The task of taking a sequence of tokenized words and a tagset, and assigning to each word one tag naming its grammatical class, such as noun or verb.

Notes The label attached to each word is the tag, and the list of available labels is the tagset. Many words belong to more than one class, so tagging is a disambiguation task: book is a verb in book that flight and a noun in hand me that book.

other attested English terms: morphosyntactic tagging, grammatical tagging

other attested French terms: étiquetage morpho-syntaxique, étiquetage grammatical

See also lemmadata annotationcorpus

Context promptingAI

An AI assistant's ability to retain information from earlier interactions and retrieve relevant information in later conversations, allowing it to maintain continuity and personalize its responses across sessions.

Notes Persistent memory is usually implemented by storing and retrieving information outside the language model itself; it does not ordinarily involve changing the model's trained parameters, and persistent does not necessarily mean permanent.

See also custom instructionscontext window

Context dataethics & society

Information that can identify, distinguish, or trace a person, either on its own or when combined with other information that can be linked to that person.

Notes It includes direct identifiers, such as passport numbers or biometric records, as well as information that may become identifying in combination, such as birth dates, locations, or demographic characteristics.

Context text encodingcomputing

Text represented as a sequence of characters alone, with no formatting or layout of its own. A plain-text file opens in any text editor, whatever produced it, which is why interchange formats are written this way: markup languages record structure inside plain text, by writing that structure as characters in the text itself.

other attested English terms: plaintext

other attested French terms: texte pur

See also character encodingUTF-8markup language

Context linguistics

The property of a word or lemma having two or more related senses.

Notes The intended sense is ordinarily determined from the context in which the word occurs. For example, mouse can refer to an animal or, through a related metaphorical sense, a computer input device.

Context deep learninglanguage models

A numerical representation of the position of each token in a sequence.

Notes In a transformer, positional information is combined with token representations so that the model can take token order into account even though computations for different sequence positions can be performed in parallel. In the original transformer architecture, a positional encoding is added to each token embedding; other transformer architectures may represent absolute or relative position in different ways.

See also transformer

Context machine translationtranslation

The human correction of raw machine translation output; the counterpart, for machine translation, of revision in human translation.

See also machine translation

Context language models

Additional training applied to a pretrained model to adapt its capabilities or behaviour for intended uses.

Notes For modern language models, post-training may include instruction tuning, preference tuning and reinforcement learning with verifiable rewards. These stages can make a base model better at following instructions, producing preferred responses or solving particular kinds of problems. Post-training is an umbrella term, and organizations do not necessarily use the same stages or sequence. In-context learning and ordinary prompting are not post-training, because they do not update the model's parameters. In some engineering contexts, post-training is used more broadly for processes applied after training, such as quantization; this entry concerns post-training as a stage in developing language models.

See also base modelinstruction tuningfine-tuningpreference tuningreinforcement learning from human feedbackreinforcement learning with verifiable rewards

Context language modelsmachine learning

An initial stage in which a model learns general patterns and reusable representations from a large dataset, often through self-supervised learning, before any later adaptation for particular tasks or uses.

Notes For autoregressive language models, pretraining typically involves learning to predict the next token from its context.

See also base modelpost-trainingself-supervised learning

Context evaluationNLP

The share of the items a system returned that are correct according to the gold standard.

Notes Noise lowers precision. It says nothing about the items the system missed, which recall measures.

other attested French terms: taux de précision

See also recallF1 scorenoisegold standard

Context language modelsmachine learning

A form of fine-tuning, applied in post-training, that adjusts a language model so that it becomes more likely to produce responses judged preferable and less likely to produce responses judged undesirable.

Notes The preferences usually come as rankings of candidate responses, made by people or by another model. Reinforcement learning from human feedback (RLHF) and direct preference optimization (DPO) are its two most common methods. The result reflects the preferences of whoever did the ranking; the stage is also called preference alignment or preference fine-tuning.

other attested English terms: preference alignment, preference fine-tuning

other attested French terms: alignement sur les préférences

See also fine-tuninginstruction tuningreward modelreinforcement learning from human feedbackdirect preference optimizationalignment¹human preference evaluation

Context machine learning

A description of how probability is allocated among the possible values of a random variable.

Notes For a discrete set of outcomes, each outcome is assigned a probability between 0 and 1, and the probabilities of all possible outcomes sum to 1. A language model produces a probability distribution over the possible next tokens in its vocabulary; to generate text, a decoding procedure selects or samples a token using this distribution.

See also decodingsampling

Context promptinglanguage models

Input given to a generative AI system to specify a task, provide context, or otherwise guide its output.

Notes For a language model, a prompt may contain instructions, questions, examples, data, or text that the model is asked to continue. The prompt becomes part of the context on which the model conditions its subsequent token generation.

See also prompt engineeringsystem promptcontext window

Context prompting

The systematic design, testing, and refinement of prompts to help a generative AI system produce outputs that better satisfy the requirements of a particular task.

other attested French terms: ingénierie de requêtes

See also promptautomatic prompt optimization

Context AI safetyAI

An attack in which untrusted input is incorporated into a prompt and interpreted by a generative AI system as instructions, causing the system to disregard or conflict with its intended instructions or to produce unintended outputs or actions.

Notes The injected instructions may be supplied directly by a user or indirectly through data that the system retrieves or processes.

other attested French terms: attaque par infiltration de requête, infiltration d'invite, injection de requête

See also jailbreakingsystem prompt

Context CAT toolstranslation

A planned, systematic set of checks carried out to give reasonable confidence that a product or service satisfies the requirements set for it.

Notes In translation, and inside a computer-assisted translation tool in particular, it takes the form of an automatic check of formal properties of a translation: numbers that differ from the source, tags present in the source and missing from the target, empty or untranslated segments, double spaces, punctuation, terms from the termbase whose equivalent does not appear in the target. The check reports where the target departs from the source or from the project's rules, and says nothing about whether the translation is any good. Which checks run and how strictly is set per project, because a software-localisation job and a marketing job tolerate different things.

other attested English terms: QA check

other attested French terms: assurance qualité, contrôle qualité

See also computer-assisted translationtermbaseinline tagsegment

Context regular expressions

A regular-expression operator stating how many times the preceding element may repeat ({n}, {n,m}, *, +, ?); greedy by default, matching as much as possible, and made lazy with a following ?

See also regular expression

Context language modelsAI

The process of using available information to draw conclusions or work toward the solution of a problem.

Notes In AI, reasoning may involve formal rules, probabilistic inference, learned patterns or representations, or a combination of these methods. In discussions of language models, reasoning is often used operationally to describe performance on tasks requiring multiple connected steps; successful answers or generated rationales do not necessarily demonstrate human-like reasoning or reveal the model's actual internal computations.

See also reasoning modelchain-of-thought

Context language modelsprompting

A setting that controls how much computation a model spends on an answer, which in practice means how many tokens it generates.

Notes The setting is separate from the model: the same model can be run quickly and cheaply or slowly and thoroughly. Raising it spends more tokens, takes longer and reaches a usage limit sooner, and it moves one model along its own range rather than making a small model behave like a large one. Every provider exposes it under its own name, as a named scale in the products and as a parameter in the APIs, and the names change from one release to the next. The labels borrow from human effort, like the rest of the reasoning vocabulary. What a higher setting changes is how many tokens the model may generate before producing its answer; nothing in the model tries harder, and the extra tokens help on some tasks and not on others.

other attested English terms: effort level, thinking level, thinking budget

other attested French terms: niveau d'effort, budget de réflexion

See also reasoning modelreasoning tokenstest-time computationtoken-based billing

Context language models

A language model designed to improve its performance on complex, multi-step problems by using additional computation during inference.

Notes Such models are often post-trained on tasks whose answers can be checked automatically and may generate intermediate steps before producing a final response. In this term, reasoning is an operational label for a model's training, inference process, and problem-solving performance: it does not establish that the model reasons in the same way as a person, and a generated rationale is not necessarily a faithful account of how the answer was produced.

other attested French terms: modèle génératif de résolution de problèmes, modèle de résolution de problèmes, modèle de résolution

See also reasoningtest-time computationreasoning tokensreasoning effort

Context language models

Tokens generated during an intermediate reasoning process before a model produces its final answer.

Notes They allow the model to perform additional step-by-step computation at inference time. Depending on the model and interface, these tokens may be displayed as a reasoning trace, hidden from the user, or represented only by a summary; they may nevertheless affect response time, token limits, and cost. The term is not completely standardized across providers. Reasoning tokens should not be interpreted literally as a model's "thoughts", nor assumed to provide a faithful explanation of how it reached its answer.

other attested English terms: thinking tokens

See also reasoning modelchain-of-thoughttoken-based billingreasoning effort

Context evaluationNLP

The share of the items in the gold standard that a system returned.

Notes Silence lowers recall. A system that returns everything has perfect recall and very low precision.

other attested French terms: taux de rappel

See also precisionF1 scoresilencegold standard

Context AI safetyAI

A structured testing process in which testers adopt an adversarial perspective and deliberately try to expose vulnerabilities, harmful or unexpected behaviour, and possible ways of misusing an AI model or system.

Notes The findings can be used to evaluate and improve the system's safeguards.

See also jailbreaking

Context machine learning

A task whose answer is a value on a scale: the system is asked to predict a number from an input.

Notes Predicting a price, a duration, a score. The training loop is the same as for classification, and only the shape of the answer differs.

See also classificationsupervised (machine) learningfeature

Context regular expressions

A pattern describing a set of character strings, built from literal characters and metacharacters with special meanings; used to search, filter and transform text by structure rather than by exact wording.

other attested French terms: expression régulière

See also metacharactercharacter classquantifierword boundary

Context machine learning

A type of machine learning in which an agent learns how to act by receiving rewards or penalties in response to its actions, with the aim of maximizing its overall reward.

Notes In language-model post-training, generated responses are evaluated by a reward signal, and the model is adjusted to make responses receiving higher rewards more likely.

See also reinforcement learning from human feedbackreward model

Context language modelsmachine learning

A method for post-training a language model in which human evaluations of possible responses are used to create a reward signal, often by training a separate reward model.

Notes Reinforcement learning then adjusts the language model to make responses receiving higher predicted rewards more likely.

See also reinforcement learningreward modelpreference tuning

Context language modelsmachine learning

A method for post-training a language model in which rewards are based on whether its responses satisfy criteria that can be checked automatically.

Notes For example, a mathematical answer may be compared with a known solution, or generated code may be tested to determine whether it runs correctly. Reinforcement learning adjusts the model to make responses receiving higher rewards more likely.

other attested French terms: apprentissage par renforcement avec récompenses vérifiables, apprentissage par renforcement à partir de résultats vérifiables

See also reinforcement learningreinforcement learning from human feedbackpost-training

Context machine learning

A machine-learning approach in which a system learns useful internal representations of its input from data rather than relying mainly on features designed by people.

Notes In a neural network, these representations are numerical vectors or activation patterns that make information useful for prediction, classification or another training objective. Representation learning can be supervised, self-supervised or unsupervised; it is not a synonym for unsupervised learning. A learned representation need not have dimensions that people can interpret individually, and calling it a representation of meaning should not imply human-like semantic understanding.

See also feature engineeringfeatureself-supervised learningword embedding

Context markuptext encoding

A character with a special syntactic function in a programming or markup language. Where it should appear as ordinary text instead, it usually has to be written another way, as a short code standing in for it: in XML, "&" and "<" cannot generally appear literally in text and are written "&amp;" and "&lt;". Whether it counts as reserved depends on the context; it is not necessarily prohibited everywhere in the document.

See also markup languageeXtensible Markup Languagetagcharacter encoding

Context language modelsagents

A technique for improving the answers of a large language model by retrieving passages from an external source, such as a document collection or a web search, and adding them to the model's context before it generates its answer.

Notes The model itself is unchanged: what it knows from training is supplemented, for one request, by text found at the moment of asking. That is how a chat product answers from your uploaded files and why its answer is only as good as the passages retrieved. Introduced by Lewis et al. (2020b), whose models combine "pre-trained parametric and non-parametric memory for language generation".

other attested French terms: génération améliorée par récupération d'information, génération de contenu améliorée par récupération d'information

See also context windowharnessfunction callinghallucination

Context language modelsmachine learning

A model trained to assign a numerical score to a candidate output, usually by learning from human judgements of which outputs are better.

Notes The judgements can also come from another model applying written principles, as in Constitutional AI (Bai et al., 2022a). The scores can then guide reinforcement learning or help select among several candidate responses.

See also reinforcement learning from human feedbackpreference tuning

Context AINLP

A system whose behaviour is fixed by rules that people wrote, most often of the form "if this condition, then this action", so that every output can be traced to the rule that produced it.

Notes In natural language processing, and everywhere on this site, rule-based means that the rules were written by people: the term names the first of four paradigms, beside machine learning, deep learning and large language models. Strictly, a system can also run on rules it did not get from a person (the ISO/IEC vocabulary defines the term by the form of the rules alone, and a model learned from data can take the form of rules, as a decision tree does); those cases are not covered here, and a model whose rules were learned belongs to machine learning.

other attested French terms: système à base de règles de production, système à règles de production

See also machine learningdeep learningregular expression

Context language models

A decoding method that uses an element of randomness to select the next token generated by a language model.

Notes A likely token is more likely to be chosen but is not certain to be, so the same prompt can give different outputs. Settings such as temperature, top-k and top-p adjust or limit the probabilities that sampling chooses from.

See also decodingtemperaturetop-k samplingtop-p sampling

Context language modelsmachine learning

Empirically observed mathematical relationships describing how a model's measured performance or error tends to change as factors such as the number of parameters, the amount of training data, and the computational resources used for training increase.

Notes These relationships can help researchers estimate the effects of increasing scale and allocate a fixed training budget between model size and data.

See also emergent ability

Context CAT toolstranslation

A portion of translatable content that a CAT tool treats as a discrete unit for translation, alignment, storage and match retrieval.

Notes A segment often corresponds to a sentence, but it may also be a title, a menu item, a paragraph, or part of a sentence, depending on the document structure and the segmentation rules applied.

See also segmentationtranslation unittranslation memorycomputer-assisted translation

Context CAT toolstranslation

The operation of dividing translatable content into segments, which correspond roughly to sentences, titles and other separate text elements.

Notes A CAT tool may use the structure of the document together with language-specific rules and exceptions for detecting sentence boundaries: it has to distinguish a period that ends a sentence from one used in an abbreviation, for example. Segmentation affects what is stored in a translation memory and which matches can later be retrieved. SRX is a format designed for exchanging segmentation rules between tools.

See also segmenttranslation unittranslation memorytokenization¹

Context deep learninglanguage models

An attention mechanism that allows each token in a sequence to assign different weights to other relevant tokens in the same sequence and combine information from them.

Notes This helps the model construct a context-sensitive representation of each token and capture relationships between tokens, including relationships across long distances in the sequence.

See also attention mechanismcross-attentiontransformer

Context machine learning

A machine-learning approach in which the targets used for training are created automatically from the structure or content of the training data, rather than being supplied as separate human-written labels.

Notes The model is typically given one part or modified version of an example and trained to predict another part. For example, an autoregressive language model is trained to predict the next token from the preceding tokens; in masked-language-model training, selected tokens are hidden and the model learns to reconstruct them from the surrounding text. In both cases, the text itself supplies the expected answers. Self-supervised learning avoids the need to label every training example manually. It does not mean that the model independently chooses what or how to learn: people and organizations still select the data, design the learning objective and determine how the resulting model will be evaluated.

See also supervised (machine) learningunsupervised learningpre-trainingmasked language modeling

Context embeddingsNLP

The degree to which two words, expressions, or concepts resemble one another in meaning.

Notes In embedding-based NLP, semantic similarity is often estimated by comparing their vectors, commonly with cosine similarity. Vector proximity is an estimate of semantic similarity, not its definition. Similarity is also distinct from broader semantic relatedness: doctor and hospital are related, but not similar in the way doctor and physician are.

Context NLP

The task of dividing a text into sentences, which comes down to deciding, for each full stop or other end mark, whether it ends a sentence.

Notes A full stop also marks an abbreviation, sits inside a number or an address, and can do two of these at once, so rule-based splitters use a dictionary of abbreviations. CAT tools do the same operation under the name segmentation, with rule sets of their own.

other attested English terms: sentence splitting

See also segmentationtokenization¹segment

Context deep learninglanguage models

A model that maps an input sequence to an output sequence, which may have a different length or structure.

Notes Most current sequence-to-sequence models are encoder-decoder models: the first term names what the model does, the second how it is built, and the two are often used almost interchangeably.

See also encoder-decoder model

Context evaluationNLP

A public evaluation exercise in which every participating team is given the same task, the same data and the same scoring, so that their systems can be compared directly.

Notes Shared tasks are how a research community measures progress on a problem from one year to the next: an organiser releases training and test data, teams submit their systems' output by a deadline, and the results are published together in a findings paper. The best-established examples are the annual WMT campaign for machine translation and IWSLT for speech translation. The same arrangement is also called a challenge, a benchmark or an evaluation campaign, depending on the community.

See also benchmarkgold standardtest set

Context evaluationNLP

The correct items a system fails to return: in term extraction, the terms of the gold standard that were not proposed.

Notes Silence lowers recall. Like noise, the word comes from information retrieval, where it names the relevant documents a search misses.

See also false negativenoiserecall

Context machine learning

"A vector is sparse if most of the values are zero" (Jurafsky & Martin).

Notes Sparse representations can be stored and processed efficiently by recording only their non-zero values and positions.

See also dense vector

Context speech

The automatic segmentation of an audio recording by speaker: determining who speaks when, so that each stretch of speech is attributed to the right person.

other attested French terms: diarisation des locuteurs

Context speech

The generation of artificial speech from written text (text-to-speech), combining linguistic processing to derive a pronounceable phonetic form with signal processing to render it as audio; the inverse of automatic speech recognition.

other attested French terms: synthèse vocale, synthèse texte-parole

Context speechtranslation

The conversion of spoken input in one language into spoken output in another language, either by chaining speech recognition, machine translation and speech synthesis (cascade) or by a single end-to-end model.

Context embeddingsNLP

A numerical representation that assigns each word type, meaning each distinct word in a vocabulary, a single fixed vector learned from patterns in a corpus.

Notes Words that occur in similar linguistic contexts generally receive similar vectors. Once learned, the vector for a word remains the same wherever that word appears: bank receives the same representation in river bank and bank account. Static word embeddings are efficient and can capture broad semantic and grammatical relationships, but they cannot directly distinguish between different meanings of a polysemous word. Word2vec and GloVe are well-known methods for creating static word embeddings.

other attested French terms: plongement statique

See also word embeddingcontextual embedding

Context NLP

A rough form of normalization that cuts the endings off words by rule, so that the forms of one word usually share a stem, which need not be a word of the language.

Notes A lemmatizer returns a dictionary form, a stemmer what is left once the ending is cut: chercher and cherches become cherch. Stemming needs no dictionary and errs in both directions, grouping forms that should stay apart and separating forms that belong together.

other attested French terms: indexation par radicaux

See also lemmanormalization

Context NLPcorpora

A very frequent word, such as an article, a preposition or a pronoun, that a system is set to ignore because it says little about the content of a text.

Notes Stop-word lists come from information retrieval and are used in corpus statistics. The list is a choice, since a word is only empty with respect to a task: the function words a search engine drops are exactly what a tagger or a parser needs.

other attested English terms: stopword

other attested French terms: mot-vide

See also normalizationtypepart-of-speech tagging

Context promptingcomputing

Model output that conforms to a predefined, machine-readable structure, such as a JSON object with specified fields and data types.

Notes The structure may be requested in a prompt or enforced during generation through constrained decoding.

See also function calling

Context language modelsNLP

A data-driven method of dividing text into tokens that are often smaller than words, such as frequent word parts, characters, or byte sequences.

Notes It allows an unfamiliar word to be represented as a sequence of known units rather than as a single unknown token.

See also tokenization²byte-pair encodingout-of-vocabulary

Context machine learning

A form of machine learning in which a model learns from labelled examples.

Notes Each training example pairs an input with its correct output, allowing the trained model to predict an appropriate output for new inputs.

See also unsupervised learningself-supervised learningtraining set

Context language modelsmachine learning

Additional training of a pretrained model on labelled examples that pair an input with a desired output, with the model's parameters adjusted to make the desired outputs more likely.

Notes For LLMs, SFT often uses instruction-and-response examples, in which case it is also called instruction tuning. Instruction tuning is a common subtype of SFT, but SFT can also be used for classification, domain adaptation, and other labelled tasks.

other attested French terms: affinage par apprentissage supervisé, peaufinage par apprentissage supervisé

See also fine-tuninginstruction tuningpost-training

Context ethics & society

An economic logic in which organizations claim and analyze data about people's activities and experiences in order to predict behaviour and, increasingly, to influence it for commercial gain.

See also personally identifiable information

Context AI safetylanguage models

The tendency of a language model trained on human preferences to agree with the user, revise a correct answer under challenge, or flatter, partly because agreement was rated well during training.

Notes Sharma et al. (2024) find it in five assistants tested in 2023 and show that preference models sometimes rate a convincingly written sycophantic answer above a correct one. For a translator, a model's agreement after pushback is therefore evidence about its training, not about who was right.

other attested English terms: AI sycophancy

other attested French terms: flagornerie de l'IA, obséquiosité de l'IA

See also alignment¹preference tuningreinforcement learning from human feedbackalignment tax

Context ethics & societyAI

Digital text, images, audio, video, or other media generated or substantially modified by an algorithm, especially by a generative AI system.

See also AI slopmodel collapse

Context cognitive science

Fast, automatic, and intuitive cognitive processing that occurs with little deliberate effort.

Notes In discussions of language models, System 1 is often used as a useful, though imperfect, analogy for rapid, direct responses generated primarily from learned patterns. The term originates in theories of human cognition, and its application to AI is metaphorical: it does not imply that language models possess human cognitive systems, conscious effort, intuition, or self-awareness.

See also System 2 thinking

Context cognitive science

Slow, deliberate, and analytical cognitive processing that requires sustained attention and effort.

Notes In discussions of language models, System 2 is often used as a useful, though imperfect, analogy for processes that devote additional computation to multistep reasoning, evaluation, or verification before producing an answer. The term originates in theories of human cognition, and its application to AI is metaphorical: it does not imply that language models possess human cognitive systems, conscious effort, intuition, or self-awareness.

See also System 1 thinkingreasoning model

Context promptinglanguage models

An instruction supplied to a language model before or alongside the user's input that establishes the model's role, task, behavioural rules, tone, or other context for the interaction.

Notes Applications commonly insert it automatically at the beginning of the model's effective context. A system prompt is often, though not always, hidden; its priority and persistence depend on the application and model interface, and it is not an absolute security boundary.

See also promptcustom instructions

Context markupXML

A marker that delimits and identifies an element in a markup document, usually working as a pair: an opening tag and a closing tag around the content they label.

See also markup languageattribute

Context NLPcorpora

The set of labels available to a part-of-speech tagger or to an annotation task, from which exactly one is given to each word.

Notes Tagsets range from 17 tags in Universal Dependencies to hundreds in sets that also record number, gender or tense.

other attested English terms: tag set

other attested French terms: jeu d'étiquettes morphosyntaxiques

See also part-of-speech taggingannotation

Context language models

A setting that influences how predictable or varied a language model's output will be.

Notes A low temperature makes the most probable tokens even more likely, so the output is more predictable; a high temperature gives less probable tokens more chance, so the output is more varied but can become less coherent. Even at zero, it does not guarantee deterministic output.

See also samplingdecodingdeterministic

Context terminology

A word or group of words that designates a concept in a specialized domain.

Notes A working definition, for translation and term extraction; terminology theory defines the notion more strictly. In practice many terms are several words long, so recognizing one includes deciding where it starts and ends.

other attested French terms: unité terminologique

See also automatic term extractiontermhoodunithoodtermbase

Context terminologyCAT tools

A database of terminological entries organized by concept, grouping the terms of several languages with metadata such as definitions, domains and usage status.

See also TermBase eXchangecomputer-assisted translation

Context exchange formatsterminology

The XML-based standard (ISO 30042) for exchanging termbases: each concept entry can gather the terms of several languages plus rich metadata such as definitions, part of speech, domain and usage notes.

See also termbaseeXtensible Markup Languageinteroperability

Context terminologyNLP

The degree to which a unit of language is related to the concepts of a specialized domain.

Notes Usually estimated by comparing a word's frequency in a specialized corpus with its frequency in a general one.

other attested French terms: termitude, termicité, potentiel terminologique

See also termunithoodautomatic term extraction

Context machine learningdata

A portion of a dataset held back from both model training and model selection and used to evaluate the final model.

Notes Its purpose is to provide an estimate of how well the completed model will perform on new data from the intended task or population. A test set should not overlap with the training data and should not guide choices about model architecture, hyperparameters or prompts: repeatedly changing a system in response to test results gradually turns the test set into a validation set and can make the reported performance overly optimistic. Training-data contamination can produce the same problem if test examples appear in the model's training data. A test set is not automatically representative merely because it was kept separate; it must also reflect the kinds of inputs, languages, populations and conditions in which the model is intended to operate.

other attested French terms: jeu de test, données de test

See also training setvalidation setdata contamination

Context language models

Computation performed when a trained model is used to answer a particular input, beyond a single ordinary forward generation.

Notes It may include producing longer intermediate reasoning, generating and comparing several candidates, searching, retrieving information, or verifying an answer.

other attested French terms: calcul à l'inférence

See also reasoning modelinference

Context AIcomputing

A usage-pricing method in which the charge for a model request is calculated from the number and category of tokens processed.

Notes Providers may set different rates for input, cached input, visible output, and internal reasoning tokens.

See also application programming interfacetoken²reasoning tokens

Context linguistics

An individual occurrence of a word or another countable element in a text.

Notes Each occurrence counts separately, even when the same form recurs: cats chase cats contains three tokens but two types. Jurafsky and Martin (2026) call this sense a word instance, to keep it apart from the tokens of language models (token²).

other attested English terms: word token, word instance, instance

other attested French terms: occurrence

See also typetoken²tokenization¹

Context language models

A unit of text that a language model processes or generates.

Notes A token may correspond to an entire word, part of a word, a punctuation mark, or a smaller character- or byte-based sequence. Each token belongs to the model's vocabulary and is represented by a numerical identifier. Tokens are therefore not the same as words: one word may be represented by a single token or divided into several tokens, and different models may divide the same text differently.

other attested French terms: jeton

See also token¹tokenization²vocabulary

Context linguistics

The process of dividing running text into individual units that can be counted, annotated or analyzed.

Notes These units are commonly words and punctuation marks, but their boundaries depend on the language, the purpose of the analysis and the conventions adopted. For example, an analyst must decide whether punctuation should be counted separately, whether a contraction such as I'm represents one written word or two grammatical units, and how to identify word boundaries in languages that do not regularly separate words with spaces.

other attested French terms: segmentation lexicale, segmentation

See also tokenization²token¹

Context language models

The process of converting text into an ordered sequence of tokens that a language model can process.

Notes A tokenizer applies a model-specific vocabulary and segmentation procedure to the text. In modern language models, commonly used methods such as byte-pair encoding learn frequently occurring units from training data: common words or expressions may consequently be represented by single tokens, while rarer or unfamiliar words are divided into smaller subword, character or byte-based units. Because tokenizers and their vocabularies differ, the same text may require different numbers of tokens in different models or languages.

other attested French terms: segmentation en jetons textuels, conversion en jetons

See also tokenization¹token²subword tokenizationbyte-pair encoding

Context agentsAI

The ability of a language model or AI system to request and use the results of external resources or software in order to complete a task.

Notes Tools may include a search engine, calculator, database, code-execution environment, calendar or external application programming interface. In a typical tool-use process, the model selects an available tool and generates a structured request containing the necessary arguments; software surrounding the model validates and executes that request, then returns the result to the model, which can incorporate it into its response or use it to decide on another action. Tool use can therefore give a model access to current information, specialized computations and actions that cannot be performed using the model's learned parameters alone. Tool use is broader than function calling, which is a common technical mechanism for expressing a tool request. A single tool call does not necessarily make a system an agent: agentic systems usually plan and coordinate multiple tool calls or actions over time. The language model commonly proposes the action rather than executing it directly, a distinction that matters for accountability and safety because the surrounding system controls permissions, validates arguments and performs actions that may have external effects.

See also AI agentapplication programming interfacefunction calling

Context language models

A sampling method that, at each generation step, retains only the k tokens with the highest probabilities, renormalizes their probabilities and randomly selects the next token from that group. A smaller value of k restricts the choice more strongly, while a larger value permits greater variety.

Notes When k = 1, top-k sampling is equivalent to greedy decoding. Unlike top-p sampling, top-k always retains the same number of candidates, even when the model's probability distribution is unusually concentrated or unusually flat.

See also samplingtop-p samplingtemperaturegreedy decoding

Context language models

A method for choosing the next token in generated text.

Notes At each step, the model ranks the possible tokens from most to least likely and keeps the smallest group whose combined probability reaches a chosen threshold, called p. It then adjusts the probabilities within this group so that they add up to 100% and randomly selects the next token from it.

See also samplingtop-k samplingtemperature

Context ethics & societymachine learning

An unwanted systematic pattern, distortion, or gap in the data used to train an AI model that influences what the model learns.

Notes It is one source of bias in AI among several, beside choices about what to measure or predict, how a model is designed and evaluated, and the context in which it is used.

other attested English terms: training data bias

other attested French terms: biais des données d'apprentissage

See also bias in AIhomogeneity biasamplification

Context datalanguage models

A collection of text or other language data used to estimate or update a model's parameters during training or pretraining.

See also training setcorpus

Context environmentAI

Greenhouse-gas emissions attributable to training a model. Depending on the assessment boundary, these may include emissions from electricity used for computation and cooling, supporting data-centre operations, and the manufacture of the hardware used.

other attested French terms: émissions liées à l'apprentissage du modèle

See also inference emissionscarbon footprintdata centre

Context machine learningdata

The portion of a dataset used to train a machine-learning model.

Notes The learning algorithm repeatedly processes these data and adjusts the model's parameters to reduce its errors or improve its performance on the training objective. In supervised learning, the training set generally contains inputs paired with expected outputs or labels; in self-supervised learning, the training targets are derived automatically from the data. For a language model, the training set may consist of a very large corpus of tokenized text from which the model learns to predict tokens.

other attested French terms: jeu d'entraînement, données d'entraînement

See also validation settest settraining corpus

Context deep learninglanguage models

A neural-network architecture for processing sequences, such as text.

Notes It consists of repeated blocks that combine self-attention with feedforward neural networks. Self-attention allows the model to determine which other tokens in the available context are most relevant to interpreting each token and to combine information from those positions. Positional information tells the model where the tokens occur, since attention by itself does not represent sequence order. By building increasingly rich contextual representations through successive layers, transformers can model relationships between parts of a text even when they are far apart. Transformers form the basis of most modern large language models.

See also self-attentionpositional encodingencoder-decoder model

Context CAT toolstranslation

A bilingual database of previously translated source segments and their corresponding target segments, which a translation-memory system searches to propose exact or similar matches for new source text.

See also Translation Memory eXchangecomputer-assisted translation

Context exchange formatsCAT tools

The XML-based standard format for exchanging translation memories between CAT tools: a TMX file contains translation units whose language variants can cover any number of languages.

See also translation memoryeXtensible Markup Languageinteroperability

Context CAT toolstranslation memories

A record that associates a source-language segment with its target-language equivalent, usually together with relevant metadata.

Notes The boundaries of these segments are determined during segmentation. In TMX, a tu element groups language-specific variants represented by tuv elements, each of which contains a seg element. A segment therefore contains text in one particular language, whereas a translation unit records the relationship between equivalent segments and may contain variants in more than two languages. The term is used here in the technical sense associated with translation memories; it can have broader and somewhat different meanings in translation studies.

See also segmentsegmentationtranslation memoryTranslation Memory eXchange

Context evaluationNLP

An item that the gold standard does not contain and that a system did not return, or did not assign to the class it looks for: for example, a legitimate message that a spam filter lets through.

Notes For a system that finds things in a text, such as a term extractor, the true negatives cannot be counted in a useful way, which is why precision and recall leave them out.

See also true positivefalse positivefalse negativeaccuracy

Context evaluationNLP

An item that a system returned, or assigned to the class it looks for, and that the gold standard also contains: for example, a candidate that is a term, or spam stopped by a spam filter.

Notes True positives are the correct answers that precision and recall count.

See also false positivefalse negativetrue negativeprecisionrecall

Context linguistics

A distinct linguistic form found in a text or corpus, counted only once regardless of how many times it occurs.

Notes For example, cats chase cats contains three word tokens but only two word types: cats and chase. What counts as the same type depends on the conventions used: capitalized and lowercase forms may be treated as the same type or as different types, and inflected forms may be counted separately unless they have been grouped under a common lemma.

See also token¹

Context terminologyNLP

The degree to which a sequence of words forms a stable unit.

Notes It only concerns terms of more than one word.

other attested French terms: figement, cohésibilité, cohésion syntagmatique

See also termtermhoodcollocation

Context machine learning

A machine-learning approach that seeks patterns or structure in data without being given a target label or expected output for each training example.

Notes It is used for tasks such as grouping similar examples, detecting anomalies, reducing dimensionality and estimating the distribution of data. This does not mean that no people are involved: people still select and prepare the data, choose the algorithm and decide how to interpret and validate the resulting patterns. Self-supervised learning is sometimes grouped under unsupervised learning because it starts with unlabelled data; it is useful to distinguish them, since self-supervised learning automatically constructs prediction targets while unsupervised learning in the narrower sense looks for structure without such input-target pairs.

See also self-supervised learningsupervised (machine) learningrepresentation learning

Context text encoding

A variable-length character encoding for Unicode. It represents each valid Unicode code point using a sequence of one to four bytes. Characters in the ASCII repertoire use one byte and retain the same byte values as in ASCII, which helps with compatibility. UTF-8 can represent the same Unicode repertoire as UTF-16 and UTF-32, but it does not imply that every character occupies a single byte. It is widely recommended for multilingual Web content and data exchange.

See also character encodingreserved character

Context machine learningdata

A portion of a dataset kept separate from the data used to fit a model's parameters and used during development to guide choices about the model.

Notes Developers may use validation results to compare model versions, select hyperparameters, set decision thresholds or decide when to stop training. Although the model does not ordinarily learn its parameters directly from the validation examples, repeated development decisions based on validation performance indirectly adapt the final system to that dataset; validation results should therefore not be presented as a fully independent estimate of final performance. Terminology is not completely consistent across disciplines, and some publications use test set for data that are actually being used for validation; this glossary maintains the stricter three-way distinction.

other attested English terms: development set, dev set

other attested French terms: jeu de validation, données de validation, jeu de développement

See also training settest sethyperparameter

Context machine learning

An ordered list of numbers.

Notes In machine learning, a vector can represent the features of an item or its location in a multidimensional space.

Context machine learning

A space in which each point is given by a list of numbers, every list in the space having the same length: two numbers place a point on a flat surface, three place it inside a volume, and a longer list places it in a space with as many directions as the list has numbers. Adding two of these lists together, or multiplying one of them by a single number, always gives another point of the same space.

Notes In NLP, vectors representing linguistic items can be treated as points in such a space and compared geometrically. A vector space is the general mathematical structure; an embedding space is a particular learned representation within such a structure, and a latent space is the broader space of hidden variables or representations in a model.

See also embedding spacelatent space

Context language models

The fixed set of tokens that a tokenizer and language model can recognize and represent.

Notes Depending on the tokenization method, the vocabulary may contain complete words, parts of words, individual characters, punctuation marks, bytes and special tokens. Each vocabulary item is assigned a numerical identifier: the tokenizer converts input text into these identifiers, and the language model selects from the same vocabulary when predicting tokens. The vocabulary is normally created during tokenizer training and then kept fixed while the model is trained and used.

other attested French terms: vocabulaire de jetons

See also token²tokenization²byte-pair encoding

Context speechethics & society

The use of AI models to build a digital copy of a person's voice and generate new speech in that voice; unlike generic speech synthesis, the goal is to reproduce a specific speaker.

Context dataweb

The automated process of systematically discovering and retrieving pages or other resources from the Web.

Notes A crawler typically begins with one or more known web addresses, downloads the corresponding pages, and follows links found in those pages to discover additional resources. Crawling is commonly used to build search-engine indexes, create web archives, monitor websites, or assemble collections of online material. Although a crawler may examine page contents and links, its main purpose is to find and collect resources rather than to extract particular data from them.

See also web scrapingCommon Crawl

Context dataweb

The automated process of retrieving online content and extracting selected information from it so that the information can be stored, analyzed, or reused in another context.

Notes A scraper may collect elements such as text, prices, tables, links, images, or metadata, often by using the structure of a web page to locate the desired information. Scraping may be applied to a known set of pages or combined with web crawling to discover additional pages; its main purpose is data extraction rather than page discovery.

See also web crawlingdata extraction

Context machine learning

A real number attached to one input of a computing unit, learned during training, that sets how much that input counts towards the unit's output.

Notes A weight can be positive or negative, so an input can count for or against a given answer. Weights and bias terms are the two kinds of number that make up a model's parameters, and the set of all of a network's weights specifies the network: that is what an open-weight model releases and a proprietary one withholds.

other attested French terms: pondération

See also parameterbias termneural networkopen-weight model

Context regular expressionscomputing

In a search, a character that stands for other characters rather than for itself, so the parts of a string that do not matter can be left unspecified.

Notes File search and word processors use * and ? this way. In a regular expression the wildcard is the dot: it matches any single character except a line break, and line breaks as well when the s flag is on.

other attested French terms: caractère joker, joker, caractère de remplacement, caractère de substitution

See also metacharacterregular expressionflag

Context regular expressions

The zero-width position between a word character and a non-word character, written \b in a regular expression; based on the ASCII \w set, so accented letters do not count as word characters.

other attested French terms: frontière de mot

See also regular expressioncharacter class

Context embeddingsNLP

A learned numerical representation of a word as a short, dense vector, meaning a list of numbers.

Notes The numbers are learned from patterns in language data rather than assigned manually, and their individual meanings are usually not interpretable. Words that occur in similar linguistic contexts generally receive similar vectors, allowing a system to compare words mathematically and recognize relationships in meaning or grammatical use. Word embeddings can therefore serve as inputs to tasks such as text classification, translation and language generation. In traditional static embedding methods such as word2vec and GloVe, each word type has a single fixed vector, so different senses of a polysemous word share the same representation. Modern language models instead create contextual embeddings for individual token occurrences, so the representation can change according to the surrounding text.

other attested French terms: plongement de mots, vecteur-mot, vectorisation de mots

See also static word embeddingcontextual embeddingembedding space

Context speechevaluation

The standard metric of speech recognition quality: the proportion of substituted, deleted and inserted words in the automatic transcript relative to the number of words in a human reference; it measures how frequent errors are, not how serious.