Embedding
A representation of a word, text, image or company as a vector of numbers, built so that objects close in meaning are close in space, the basis of semantic search and retrieval-augmented generation systems.
An embedding turns a complex object, word, sentence, document, image, or even a company described by its characteristics, into a list of a few hundred or thousand numbers. These vectors are learned by a model so that the distance between two vectors reflects how close in meaning the objects they represent are: the text of a war exclusion and of a cyber war clause will be close, those of a war clause and a cooking recipe far apart. Popularized for words in the early 2010s, the technique became widespread with language models, which produce embeddings of entire texts. Its uses in insurance are many: searching for similar clauses in a corpus of policies, matching a business description to an industry classification, detecting claims whose descriptions are suspiciously similar, and above all feeding retrieval-augmented generation systems, which find the relevant passages of documentation before answering. Its limits lie in its opacity, since the dimensions have no readable meaning, and its dependence on the model that produced it, whose biases carry over into the computed proximities.
Textbook case: an insurer turns the 12,000 clauses in its policy corpus into vectors. When a new cyber exclusion text arrives from a broker, proximity search finds within a second the twenty closest clauses already accepted, with their date and the underwriter who approved them.
Follow the research
One letter a week. First, the week's underwriting note, when one is published: a dated, argued decision, signed by Paul Bernardski. Then what the observatory published on cyber risk, political risk and insurance. No account, no tracking.
No spam, one-click unsubscribe, no data ever sold.