Embeddings

An embedding is a numerical representation produced from an input for a particular model or method. Software can compare these representations while locating related text, images, or other material. The construction and comparison method affect the resulting relation.

For an interview passage, software produces a representation, stores its link to the permitted source, compares it with another question or account. A high similarity can identify a candidate for review. Staff still examine the actual passages because the comparison can obscure negation, context, disagreement, a specific unusual reference.

Staff can preserve several local representations instead of averaging all material from a person. Clusters and centroids examine that choice. A source correction, permission change, model change may require updating or removing derived representations. Reprocessing and withdrawal include these derived records.