Incorporations de mots

A word embedding is a learned vector: a list of numerical coordinates associated with a word. Its dimensions are features learned by a training method, not a manually written dictionary definition. The word2vec research introduced efficient CBOW and skip-gram architectures. CBOW predicts a target from surrounding context, while skip-gram predicts surrounding words from a target. GloVe uses global word co-occurrence information in its training objective. Common released configurations use tens or hundreds of dimensions rather than one coordinate per possible word. Some vector offsets capture useful relationships, motivating illustrations such as king minus man plus woman landing near queen. Such an example is a property investigated in learned representations, not an arithmetic law that succeeds for every relation and dataset. Static word2vec and GloVe embeddings assign a word a fixed vector regardless of the sentence. Contextual representations, studied in work such as ELMo and used in transformer networks, depend on surrounding text. A transformer starts with input token embeddings and produces contextual hidden representations through its layers. That distinction matters for a word such as bank used for a river edge or a financial institution. Embeddings can support language applications, but word vectors and whole-document representations are not automatically interchangeable. A semantic-search system may use a separately trained or aggregated representation for queries and passages. Evaluate that representation for the intended task rather than assuming that any word-vector model supplies reliable document retrieval.