Document body vectorization and noise-contrastive training
Document embedding vectors for each document of a corpus may be generated by combining embedding vectors for document subparts, thereby yielding a final embedding vector for the document. A machine learning model is trained using a query corpus and the document corpus, where the model generates a ranking score for a given (query, document) pair. During training, rankings scores are generated using the model, such that the training dataset is further refined using the generated ranking scores. For example, top documents and a negative document may be determined for a given query and subsequently used as training data. Multiple negative documents may therefore be determined for a given query. A negative document for a given query may be determined from the negative documents using noise-contrastive estimation. Such determined negative documents may be evaluated using a loss function during model training, thereby yielding a more robust model for search processing.
1. A system comprising:
at least one processor; and
memory storing instructions that, when executed by the at least one processor, causes the system to perform a set of operations, the set of operations comprising:
training a machine learning model based on a training dataset comprising a search query corpus and the document corpus, wherein training the machine learning model comprises:
generating, for each document of the document corpus, a document embedding vector based on a plurality of embedding vectors that are each associated with a document subpart of a given document;
generating, using the machine learning model, a set of ranking scores for documents of the document corpus based on a first search query of the search query corpus and a document embedding vector for each of the documents;
refining the training dataset based on the generated set of ranking scores;
determining a first negative document from a set of negative documents for the first search query;
evaluating a loss function using the first negative document to train the machine learning model, thereby yielding an updated machine learning model; and
in a subsequent training iteration:
further refining the training dataset based on the updated machine learning model to generate a second training dataset; and
further training the updated machine learning model based on the second data set;
obtaining a request comprising a second search query;
generating, using the trained machine learning model, a set of documents from the document corpus that is responsive to the second search query; and
providing, in response to the request, the set of documents that is responsive to the second search query.
2. The system of claim 1 , wherein refining the training dataset comprises:
retaining, for the first search query, a subset of documents of the document corpus in the training dataset based on the set of ranking scores; and
determining a second negative document for the first search query from the document corpus, wherein the second negative document is part of the set of negative documents for the first search query.
3. The system of claim 2 , wherein the second negative document is randomly determined.
4. The system of claim 1 , wherein the first negative document is determined from the set of negative documents for the first search query using noise-contrastive estimation.
5. The system of claim 1 , wherein the loss function evaluates a first cosine similarity between a query embedding vector for the first search query and a first document embedding vector for the first negative document.
6. The system of claim 5 , wherein the loss function further evaluates a second cosine similarity between the query embedding vector and a second document embedding vector for a positive document associated with the first search query.
7. The system of claim 1 , wherein generating the set of documents that is responsive to the second search query comprises:
performing an approximate nearest neighbor search using a query embedding vector for the second search query and document embedding vectors for documents of the document corpus to generate the set of documents; and
ranking the set of documents according to associated ranking scores.
8. A method for generating a set of documents responsive to a search query, comprising:
obtaining a request comprising a search query;
generating a query embedding vector for the search query;
generating, based on the query embedding vector and document embedding vectors for documents of a document corpus, a set of documents responsive to the search query, wherein a document embedding vector of the document embedding vectors is generated based on a plurality of embedding vectors that are each associated with a document subpart of a given document;
evaluating, using a loss function, a first cosine similarity between the query embedding vector for a first search query and a first document embedding vector for a first negative document and a second cosine similarity between the query embedding vector for a second search query and a second document embedding vector for a positive document associated with the first search query;
generating, the set of documents responsive to the second search query; and
ranking the set of documents according to associated ranking scores; and providing, in response to the request, the ranked set of documents that is responsive to the search query.
9. The method of claim 8 , wherein generating the set of documents responsive to the search query comprises processing the query embedding vector and the document embedding vectors using an approximate nearest neighbor search.
10. The method of claim 8 , wherein a document embedding vector for a document of the document corpus is a pre-generated document embedding vector based on a plurality of embedding vectors, wherein each embedding vector of the plurality of embedding vectors is associated with a subpart of the document.
11. The method of claim 8 , wherein a document embedding vector for a document of the document corpus is associated with a body of the document.
12. The method of claim 8 , wherein providing the ranked set of documents comprises providing a subpart of a document in the ranked set of documents.
13. A method for machine learning model-based search processing, comprising:
training a machine learning model based on a training dataset comprising a search query corpus and the document corpus, wherein training the machine learning model comprises:
generating, using the machine learning model, a set of ranking scores for documents of the document corpus based on a first search query of the search query corpus;
refining the training dataset based on the generated set of ranking scores;
determining a first negative document from a set of negative documents for the first search query;
evaluating a loss function using the first negative document to train the machine learning model, thereby yielding an updated machine learning model; and
in a subsequent training iteration:
further refining the training dataset based on the updated machine learning model to generate a second training data set; and
further training the updated machine learning model based on the second data set;
obtaining a request comprising a second search query;
generating, using the trained machine learning model, a set of documents from the document corpus that is responsive to the second search query; and
providing, in response to the request, the set of documents that is responsive to the second search query.
14. The method of claim 13 , further comprising:
generating, for each document of the document corpus, a document embedding vector based on an embedding vector of a plurality of subparts of the given document.
15. The method of claim 13 , wherein refining the training dataset comprises:
retaining, for the first search query, a subset of documents of the document corpus in the training dataset based on the set of ranking scores; and
determining a second negative document for the first search query from the document corpus, wherein the second negative document is part of the set of negative documents for the first search query.
16. The method of claim 13 , wherein the first negative document is determined from the set of negative documents for the first search query using noise-contrastive estimation.
17. The method of claim 13 , wherein the loss function evaluates a first cosine similarity between a query embedding vector for the first search query and a first document embedding vector for the first negative document.
18. The method of claim 17 , wherein the loss function further evaluates a second cosine similarity between the query embedding vector and a second document embedding vector for a positive document associated with the first search query.
19. The method of claim 13 , wherein generating the set of documents that is responsive to the second search query comprises:
performing an approximate nearest neighbor search using a query embedding vector for the second search query and document embedding vectors for documents of the document corpus to generate the set of documents; and
ranking the set of documents according to associated ranking scores.
20. The system of claim 1 , wherein training the machine learning model further comprises refining the training dataset in a subsequent training iteration based on a current iteration of the machine learning model.