Identifying content items in response to a text-based request

1. A non-transitory computer-readable medium bearing computer executable instructions which, when executed on a computing system comprising at least a processor executing the computer executable instructions, carries out a method, comprising:

maintaining a corpus of non-text content items comprising non-text content items, wherein each non-text content item of the corpus is associated with an embedding vector that projects the non-text content item into a non-text content item embedding space;

receiving a text-based request for content items of the corpus of non-text content items;

processing the text-based request to generate a request embedding vector for the text-based request that projects the request embedding vector into the non-text content item embedding space;

determining a first non-text content item of the corpus of non-text content items according to a projection of the request embedding vector into the non-text content item embedding space; and

returning the first non-text content item in response to the text-based request for content items.

2. The non-transitory computer-readable medium of claim 1 , wherein processing the text-based request to generate the request embedding vector is based at least in part on a plurality of word pieces determined from the text-based request.

3. The non-transitory computer-readable medium of claim 2 , wherein the plurality of word pieces are determined based at least in part on at least one of a morphological analysis of the text-based request or a byte pair encoding technique analysis of the text-based request.

4. The non-transitory computer-readable medium of claim 2 , wherein processing the text-based request to generate the request embedding vector includes:

determining a plurality of word piece embedding vectors corresponding to the plurality of word pieces, each of the plurality of word piece embedding vectors projecting a respective word piece of the plurality of word pieces into the non-text content embedding space; and

combining the plurality of word piece embedding vectors to generate the request embedding vector.

5. The non-transitory computer-readable medium of claim 1 , wherein determining the first non-text content item of the corpus of non-text content items is based at least in part on a distance between the request embedding vector and the embedding vector associated with the first non-text content item in the non-text content item embedding space.

6. A computer-implemented method, comprising:

generating, for a text-based request for content items, a request embedding vector that is representative of the text-based request and projects the request embedding vector into a non-text content item embedding space;

determining, based at least in part on the projection of the request embedding vector into the non-text content embedding space, a non-text content item from a plurality of non-text content items, each of the plurality of non-text content items associated with a respective content item embedding vector that projects the plurality of non-text content items into the non-text content item embedding space; and

providing the non-text content item in response to the text-based request for content items.

7. The computer-implemented method of claim 6 , wherein generating the request embedding vector is based at least in part on a plurality of word pieces determined from the text-based request.

8. The computer-implemented method of claim 7 , wherein the plurality of word pieces are determined based at least in part on at least one of a morphological analysis of the text-based request or a byte pair encoding technique analysis of the text-based request.

9. The computer-implemented method of claim 7 , wherein:

determining the plurality of word pieces includes:

processing the text-based request to determine a set of text terms included in the text-based request; and

performing a morphological analysis of the set of text terms to determine the plurality of word pieces; and

each of the plurality of word pieces includes a respective morpheme.

10. The computer implemented method of claim 7 , wherein each of the plurality of word pieces corresponds to one more word parts of text-based terms of the text-based request and includes one of a prefix of one of the text-based term, a suffix of one of the text-based term, or a root of one of the text-based term.

11. The computer-implemented method of claim 7 , wherein generating the request embedding vector includes:

determining a plurality of word piece embedding vectors corresponding to the plurality of word pieces, each of the plurality of word piece embedding vectors projecting a respective word piece of the plurality of word pieces into the non-text content embedding space;

determining a plurality of weightings corresponding to the plurality of word pieces; and

combining the plurality of word piece embedding vectors in accordance with the plurality of weightings to generate the request embedding vector.

12. The computer-implemented method of claim 11 , further comprising:

conducting a semantic analysis of the text-based request to determine at least one of a topic associated with the text-based request or an intent associated with the text-based request,

wherein the plurality of weightings is determined at least in part on at least one of the topic or the intent associated with the text-based request.

13. The computer-implemented method of claim 6 , wherein determining the non-text content item from the plurality of non-text content items is based at least in part on a distance between the request embedding vector and the respective content item embedding vector associated with the non-text content item in the non-text content item embedding space based on a cosine similarity of the request embedding vector and the respective content item embedding vector associated with the non-text content item.

14. The computer-implemented method of claim 6 , further comprising:

identifying a closest content item from the plurality of non-text content items that is closest to the request embedding vector in the non-text content item embedding space;

conducting a random walk originating from the closest content item in a content item graph representing relationships between the plurality of non-text content items;

determining, based at least in part on the random walk, a second plurality of non-text content items from the plurality of non-text content items; and

providing at least a portion of the second plurality of non-text content items in response to the text-based request for content items.

15. A computing system, comprising:

one or more processors; and

a memory including program instructions that, when executed by the one or more processors, cause the one or more processors to at least:

maintain a content item graph for a corpus of non-text content items, wherein each non-text content item of the corpus of non-text content items is associated with a respective embedding vector that projects the corpus of non-text content items into a non-text content item embedding space;

generate, for a text-based request for content items using a trained machine learning model configured to generate output embedding vectors that project text-based inputs into the non-text content item embedding space, a request embedding vector that is representative of the text-based request and projects the request embedding vector into the non-text content item embedding space;

determine, based at least in part on the projection of the request embedding vector into the non-text content embedding space, a non-text content item from the corpus of non-text content items; and

provide the non-text content item in response to the text-based request for content items.

16. The computing system of claim 15 , wherein the program instructions that, when executed by the one or more processors, further cause the one or more processors to at least:

process the text-based request to determine a plurality of word pieces;

generate a plurality of word piece embedding vectors corresponding to the plurality of word pieces, each of the plurality of word piece embedding vectors projecting a respective word piece of the plurality of word pieces into the non-text content embedding space; and

combine the plurality of word piece embedding vectors to generate the request embedding vector.

17. The computing system of claim 16 , wherein:

the program instructions that, when executed by the one or more processors, further cause the one or more processors to at least determine, for each of the plurality of word pieces, a plurality of weightings corresponding to the plurality of word pieces based at least in part on at least one of an importance or a frequency associated with each of the plurality of word pieces; and

combining the plurality of word piece embedding vectors is based at least in part on the plurality of weightings.

18. The computing system of claim 16 , wherein each of the plurality of word pieces includes a respective morpheme determined based at least in part on at least one of a morphological analysis of the text-based request or a byte pair encoding technique analysis of the text-based request.

19. The computing system of claim 15 , wherein the program instructions that, when executed by the one or more processors, further cause the one or more processors to at least:

identify a closest content item from the corpus of non-text content items that is closest to the request embedding vector in the non-text content item embedding space;

conduct a random walk originating from the closest content item in the content item graph;

determine, based at least in part on the random walk, a plurality of non-text content items from the corpus of non-text content items; and

provide at least a portion of the plurality of non-text content items in response to the text-based request for content items.

20. The computing system of claim 15 , wherein the program instructions that, when executed by the one or more processors, further cause the one or more processors to at least:

aggregate a corpus of request logs to generate training data for training the trained machine learning model, wherein each request log of the corpus of request logs includes a respective request and a respective content items to generate; and

train a machine learning model according to the generated training data to generate the trained machine learning model.