Multilabel learning via supervised joint embedding of documents and labels

1. A method implemented by at least one server computer, comprising:

receiving a plurality of training documents, each training document being defined by a sequence of words, each training document having one or more labels associated therewith;

embedding the plurality of training documents, the words, and the labels in a vector space, wherein the embedding is configured to locate a given training document and its associated labels in proximity to each other in the vector space;

embedding a new document in the vector space;

performing a proximity search in the vector space to identify a set of nearest labels to the new document in the vector space; and

associating the set of nearest labels to the new document, wherein embedding the new document is configured to predict a target word in the new document using context words in the new document and identification of the new document, and wherein the embedding is configured to minimize a loss function that includes a component configured to approximate a conditional probability of the target word based on the context words and the identification of the new document.

2. The method of claim 1 , wherein the method solves a multi-label learning problem, such that a number of the set of nearest labels to the new document is not predefined prior to performing the proximity search.

3. The method of claim 1 , wherein the embedding is configured to simultaneously learn document vectors corresponding to the plurality of training documents and label vectors corresponding to the labels directly from the words of the plurality of training documents.

4. The method of claim 3 , wherein the embedding is configured to use the document vectors to learn the label vectors by solving a multiclass classification problem.

5. The method of claim 4 ,

wherein the embedding is configured to predict the labels of the given training document using the identification of the given training document.

6. The method of claim 1 ,

wherein the loss function includes a component configured to approximate a conditional probability of the labels of the given training document based on the identification of the given training document.

7. The method of claim 1 , wherein performing the proximity search includes performing a k-nearest neighbor search.

8. The method of claim 1 , wherein the new document is not one of the plurality of training documents and does not have labels already associated therewith, and wherein the embedding of the new document is independent of the embedding of the labels in the vector space.

9. The method of claim 1 , further comprising:

receiving a request to access documents associated with one label of the set of nearest labels to the new document; and

in response to the request, providing access to the new document in association with the one label of the set of nearest labels.

10. The method of claim 9 , wherein the documents define one or more of articles, product descriptions, and social media posts.

11. The method of claim 9 , wherein the request is defined from a search query, a social media access request, a product search, a category request, a topic request, or a community access request.

12. A computer readable medium, being non-transitory, having program instructions embodied thereon, the program instructions being configured, when executed by a computing device, to cause the computing device to perform the following operations:

receive a plurality of training documents, each training document being defined by a sequence of words, each training document having one or more labels associated therewith;

embed the plurality of training documents, the words, and the labels in a vector space, wherein the embedding of the plurality of training documents, the words, and the labels is configured to locate a given training document and its associated labels in proximity to each other in the vector space;

embed a new document in the vector space;

perform a proximity search in the vector space to identify a set of nearest labels to the new document in the vector space; and

associate the set of nearest labels to the new document, wherein the new document that is embedded is configured to predict a target word in the new document using context words in the new document and identification of the new document, and the embedding is configured to minimize a loss function that includes a component configured to approximate a conditional probability of the target word based on the context words and the identification of the new document.

13. The computer readable medium of claim 12 , wherein the operations solve a multi-label learning problem, such that a number of the set of nearest labels to the new document is not predefined prior to performing the proximity search.

14. The computer readable medium of claim 12 , wherein the embedding is configured to simultaneously learn document vectors corresponding to the plurality of training documents and label vectors corresponding to the labels directly from the words of the plurality of training documents; and

wherein the embedding is configured to use the document vectors to learn the label vectors by solving a multiclass classification problem.

15. A server computer, comprising:

training logic, the training logic configured to,

receive a plurality of training documents, each training document being defined by a sequence of words, each training document having one or more labels associated therewith, and

embed the plurality of training documents, the words, and the labels in a vector space, wherein the embedding is configured to locate a given training document and its associated labels in proximity to each other in the vector space; and

inference logic, the inference logic configured to,

embed a new document in the vector space,

perform a proximity search in the vector space to identify a set of nearest labels to the new document in the vector space, and

associate the set of nearest labels to the new document, wherein the new document that is embedded is configured to predict a target word in the new document using context words in the new document and identification of the new document, and the embedding is configured to minimize a loss function that includes a component configured to approximate a conditional probability of the target word based on the context words and the identification of the new document.

16. The server computer of claim 15 , wherein operations solve a multi-label learning problem, such that a number of the set of nearest labels to the new document is not predefined prior to performing the proximity search.

17. The server computer of claim 15 , wherein the embedding is configured to simultaneously learn document vectors corresponding to the plurality of training documents and label vectors corresponding to the labels directly from the words of the plurality of training documents; and

wherein the embedding is configured to use the document vectors to learn the label vectors by solving a multiclass classification problem.