Dense retrieval employing progressive distillation training

A method comprising:

• receiving a query that is indicative of at least one of a computing context of a client computing device or user input set forth by way of the client computing device;
• identifying search results from a pool of potential results by performing acts comprising:
• computing an embedding of the query by way of a first encoder;
• selecting candidate search results from the pool of potential results based upon the embedding of the query and embeddings of the potential results, the candidate search results being a subset of the potential results;
• computing a similarity score for a first candidate search result in the candidate search results by way of a second encoder, wherein the second encoder that receives the query and the first candidate result as input, wherein the second encoder is trained based upon an order metric, and wherein the first encoder is trained based upon output of the second encoder;
• ranking the candidate search results based upon the similarity score;
• and
• identifying the search results based upon the ranking of the candidate search results;
• and
• outputting the identified search results to the client computing device.

A computer-readable storage medium storing instructions that, when executed by a processor, cause the processor to perform acts comprising:

• receiving a query that is indicative of at least one of a computing context of a client computing device or user input set forth by way of the client computing device;
• retrieving candidate search results from a pool of potential results by way of a dense retriever, wherein retrieving the candidate search results comprises:
• computing an embedding of the query by way of a first encoder;
• and
• retrieving the candidate search results from the pool of potential results based upon the embedding of the query and embeddings of the potential results, the candidate search results being a subset of the potential results;
• ranking the retrieved candidate search results by way of a ranker, wherein ranking the retrieved candidate search results comprises:
• computing a similarity score for a first candidate search result in the candidate search results by way of a second encoder, wherein the second encoder receives the query and the first candidate result as input, wherein the second encoder is trained based upon an order metric, and wherein the first encoder is trained based upon output of the second encoder;
• and
• ranking the candidate search results based upon the similarity score;
• and
• outputting a search result to the client computing device based upon the ranked candidate search results.