Context: I am interested in using the potential of using embeddings to consolidate texts/documents with very different surface forms into one searchable database - in other words, to produce a sort of multimodal semantic search function.

Aim: For instance, a retail business may hold plain text product descriptions, salesperson chat logs in a dialogue format, clickstream data in a JSON format, etc. Once encoded into embeddings, this produces a searchable vector index. Queries can be encoded and indexed, and relevant (i.e. semantically-similar) texts/documents are returned regardless of whether they are plain text, chat logs, clickstreams, etc.

Question: What research has been done on this problem of semantic search over texts with very different surface forms? Are there any pitfalls or limitations that are unique to semantic search with this sort of document database? Are there any approaches that have been shown to mitigate those limitations in this context?



Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge that you have read and understand our privacy policy and code of conduct.