Identifying content items in response to a text-based request
Systems and methods for responding to a subscriber's text-based request for content items are presented. In response to a request from a subscriber, word pieces are generated from the text-based terms of the request. A request embedding vector of the word pieces is obtained from a trained machine learning model. Using the request embedding vector, a set of content items, from a corpus of content items, is identified. At least some content items of the set of content items are returned to the subscriber in response to the text-based request for content items.
1. A computing system, comprising:
one or more processors; and
a memory storing program instructions that, when executed by the one or more processors, cause the one or more processors to at least:
maintain a plurality of image content items, wherein the plurality of image content items are associated with a plurality of image content item embedding vectors that project the plurality of image content items into an image content item embedding space;
in response to a text-based query, project a request embedding vector generated from the text-based query into the image content item embedding space;
determine, based at least on the plurality of image content item embedding vectors and the request embedding vector, an image content item from the plurality of image content items; and
provide the image content item in response to the text-based request.
2. The computing system of claim 1 , wherein the program instructions include further instructions that, when executed by the one or more processors, further cause the one or more processors to at least generate, using a trained machine learning model, the request embedding vector.
3. The computing system of claim 2 , wherein generating the request embedding vector includes:
determining a plurality of word pieces from the text-based query;
determining a plurality of word piece embedding vectors for the plurality of word pieces; and
combining the plurality of word piece embedding vectors to generate the request embedding vector.
4. The computing system of claim 3 , wherein combing the plurality of word piece embedding vectors includes:
determining weights for at least one word piece embedding vector of the plurality of word piece embedding vectors; and
combining the plurality of word piece embedding vectors according to the determined weights.
5. The computing system of claim 2 , wherein the trained machine learning model is trained using training data generated using request-content item logs that includes a plurality of positive request-content item pairs.
6. A computer-implemented method, comprising:
projecting a first plurality of embedding vectors corresponding to a plurality of images into an image embedding space;
generating a second embedding vector representative of a text input, wherein generating the second embedding vector includes:
determining a plurality of word pieces based on the text input;
determining a plurality of word piece embedding vectors for the plurality of word pieces; and
combining the plurality of word piece embedding vectors to generate the second embedding vector;
projecting the second embedding vector into the image embedding space;
determining, based at least in part on the projection of the second embedding vector and the projection of the first plurality of embedding vectors, a first image from the plurality of images; and
providing the first image in response to the text input.
7. The computer-implemented method of claim 6 , wherein generating the second embedding vector further includes:
performing a semantic analysis of the text input, the performing the semantic analysis including determining an intent of the text input.
8. The computer-implemented method of claim 6 , wherein determining the plurality of word pieces includes:
performing a morphological analysis of a set of terms derived from the text input, the performing the morphological analysis including determining a plurality of morphemes as the plurality of word-pieces.
9. The computer-implemented method of claim 8 , wherein the set of terms is determined using a lexical analysis of the text input.
10. The computer-implemented method of claim 8 , further comprising:
storing the plurality of morphemes in association with a plurality of corresponding words;
receiving a second text input;
generating a third embedding vector representative of the second text input, wherein generating the third embedding vector includes determining a second plurality of word pieces for the second text input based at least in part on the stored plurality of morphemes;
projecting the third embedding vector into the image embedding space;
determining, based at least in part on the projection of the third embedding vector and the projection of the first plurality of embedding vectors, a second image from the plurality of images; and
providing the second image in response to the second text input.
11. The computer-implemented method of claim 6 , wherein combining the plurality of word piece embedding vectors includes:
determining weights for at least one word piece embedding vector of the plurality of word piece embedding vectors; and
combining the plurality of word piece embedding vectors according to the determined weights.
12. The computer-implemented method of claim 11 , wherein the weights are based at least in part on at least one of:
an intent of the text input;
a topic of the text input;
an importance of a first corresponding word piece; or
a frequency of a second corresponding word piece.
13. The computer-implemented method of claim 6 , wherein determining the first image from the plurality of images is based at least in part on a distance between the second embedding vector and a respective embedding vector of the first plurality of embedding vectors that is associated with the first image in the image embedding space based on a cosine similarity of the second embedding vector and the respective embedding vector associated with the first image.
14. The computer-implemented method of claim 6 , wherein determining the first image from the plurality of images includes:
identifying a closest image from the plurality of images that is closest to the request embedding vector in the image embedding space;
conducting a random walk originating from the closest image in a content item graph representing relationships between the plurality of images;
determining, based at least in part on the random walk, a second plurality of images from the plurality of images; and
determining the first image from the second plurality of images.
15. The computer-implemented method of claim 6 , further comprising: storing the second embedding vector in association with the text input; receiving a second text input;
determining a third embedding vector for the second text input based at least in part on the stored second embedding vector;
projecting the third embedding vector into the image embedding space;
determining, based at least in part on the projection of the third embedding vector and the projection of the first plurality of embedding vectors, a second image from the plurality of images; and
providing the second image in response to the second text input.
16. A computer-implemented method, comprising:
maintaining a plurality of image content items in a content item graph, wherein each image content item of the plurality of image content items is associated with a respective embedding vector that projects each image content item into an image content item embedding space;
in response to a text-based input, generating, using a trained machine learning model, a first embedding vector representative of the—text-based input;
projecting the first embedding vector into the image content item embedding space;
determining, based at least in part on the first embedding vector and a second embedding vector associated with a first image content item of the plurality of image content items, that the first image content item is responsive to the text-based input; and
providing the first image content item in response to the text-based input.
17. The computer-implemented method of claim 16 , wherein the trained machine learning model is trained using training data generated using request-content item logs that includes a plurality of positive request-content item pairs.
18. The computer-implemented method of claim 16 , wherein:
the first image content item corresponds to a first node in the content item graph; and
the first image content item is determined based at least in part on conducting a random walk originating from a second node of the content item graph that corresponds to a second image content item that is determined to be closest to the text-based input in the image content item embedding space.
19. The computer-implemented method of claim 16 , wherein the first embedding vector is based at least in part on a combination of weighted word piece embedding vectors derived from the text-based input.