Dave's PCF WIP: Terms

Term #2,272

Edit Term

cosine similarity

View List Terms Delete

Claim Term

Application

App Docket

Created Date

Full Desc

"Cosine similarity" refers to a mathematical measure used to quantify the similarity between two non-zero vectors by calculating the cosine of the angle between them in a multi-dimensional space. Synonyms for “cosine similarity” may include angular similarity, vector angle similarity, semantic vector comparison, directional similarity metric, or the like. In the context of natural language processing or artificial intelligence, cosine similarity may be applied to compare vectorized representations—such as sentence embeddings—of textual data, such as machine-generated editorial comments and human-generated editorial comments.

Cosine similarity may yield a value in the range from -1 to 1, where a value closer to 1 indicates greater similarity in direction (i.e., semantic or contextual alignment), a value near 0 indicates orthogonality or irrelevance, and a value near -1 indicates opposition. Cosine similarity may be used during model training to evaluate the alignment between generated and reference outputs or during inference to assess consistency, paraphrasing, or redundancy. The vectors being compared may be derived from neural network embeddings, sentence encoders, transformer-based language models, or the like. Cosine similarity may support training objectives, retrieval systems, clustering, or editorial quality evaluation. (Defined in conjunction with ChatGPT 4o Version, July 15, 2025.)