Training a learning-to-rank model using a linear difference vector

A computer-implemented method, comprising:

• serializing training data, the training data comprising training data entries, a training data entry comprising a query, one or more candidate documents, and labels corresponding to the one or more candidate documents, wherein serializing the training data avoids truncating or padding the one or more candidate documents;
• providing the serialized training data to a document recommendation model;
• obtaining one or more document prediction scores from the document recommendation model, the one or more document prediction scores indicative of a likelihood that the one or more candidate documents are responses to the query;
• generating a pairwise score difference matrix using the one or more document prediction scores;
• transforming the pairwise score difference matrix into a score difference vector using a correct document vector;
• calculating loss data of the document recommendation model using the score difference vector;
• and
• adjusting the document recommendation model using the loss data.

Computer storage media having computer-executable instructions that, upon execution by a processor, cause the processor to:

• serialize training data, the training data comprising training data entries, a training data entry comprising a query, one or more candidate documents, and labels corresponding to the one or more candidate documents, wherein serializing the training data avoids truncating or padding the one or more candidate documents;
• provide the serialized training data to a document recommendation model;
• obtain one or more document prediction scores from the document recommendation model, the one or more document prediction scores indicative of a likelihood that the one or more candidate documents are responses to the query;
• generate a pairwise score difference matrix using the one or more document prediction scores;
• transform the pairwise score difference matrix into a score difference vector using a correct document vector;
• calculate loss data of the document recommendation model using the score difference vector;
• and
• adjust the document recommendation model using the loss data.