This disclosure describes techniques of information retrieval from text-based corpuses by combining dense and sparse embeddings into a single, composite, dense embedding. Documents from the corpus most relevant to a user query can be found by using the composite dense embedding to natively run nearest-neighbor searches using existing tools. The techniques retain information from both sparse and dense embeddings and provide a straightforward, mathematically sound way of combining them. The composite dense embedding, when used in retrieving information in response to a user query, performs well even when the query includes a proper name or a rare term, and can process the subtleties of natural language while executing natively on existing embedding matching tools.

Creative Commons License

Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.