US 11,886,815 B2Granted

Self-supervised document representation learning

AssigneeAdobe Inc.

InventorsJiuxiang Gu, Vlad Morariu, Varun Manjunatha, Tong Sun, Rajiv Jain, Peizhao Li, Jason Kuen, Handong Zhao

Life of this patent

Estimating the term…

Loading patent…

Abstract

One example method involves operations for a processing device that include receiving, by a machine learning model trained to generate a search result, a search query for a text input. The machine learning model is trained by receiving pre-training data that includes multiple documents. Pre-training the machine learning model by generating, using an encoder, feature embeddings for each of the documents included in the pre-training data. The feature embeddings are generated by applying a masking function to visual and textual features in the documents. Training the machine learning model also includes generating, using the feature embeddings, output features for the documents by concatenating the feature embeddings and applying a non-linear mapping to the feature embeddings. Training the machine learning model further includes applying a linear classifier to the output features. Additionally, operations include generating, for display, a search result using the machine learning model based on the input.

Claims19

Dependent claims

A method in which one or more processing devices perform operations comprising:

receiving, by a machine learning model trained to generate a search result comprising a document, a search query comprising a text input, wherein the machine learning model is trained by:
receiving pre-training data comprising a plurality of documents;
and
pre-training the machine learning model by:
generating, using an encoder, feature embeddings for each of the plurality of documents by applying a masking function to visual features and textual features in each of the plurality of documents, wherein generating the feature embeddings comprises:
generating weighted visual features by applying a first modality-specific attention weight to the visual features;
generating weighted textual features by applying a second modality-specific attention weight to the textual features;
and
fusing the weighted visual features and the weighted textual features, wherein fusing the weighted visual features and the weighted textual features comprises multiplying the weighted visual features and the weighted textual features;
generating, using the feature embeddings, output features for each of the plurality of documents by (i) concatenating the feature embeddings and (ii) applying a non-linear mapping to the feature embeddings;
and
applying a linear classifier to the output features;
and
generating, for display, the search result using the machine learning model by retrieving the search result based on the text input.

11 more elements

A system comprising:

one or more processing devices;
and
a non-transitory computer-readable medium communicatively coupled to the one or more processing devices and storing instructions, wherein the one or more processing devices are configured to execute the instructions and thereby perform operations comprising:
receiving pre-training data comprising a plurality of documents;
pre-training a machine learning model by:
obtaining the pre-training data comprising the plurality of documents;
extracting visual features from each of the plurality of documents;
extracting textual features from each of the plurality of documents;
and
generating, using an encoder, feature embeddings for each of the plurality of documents by applying a masking function to the visual features and the textual features, wherein generating the feature embeddings comprises:
generating weighted visual features by applying a first modality-specific attention weight to the visual features;
generating weighted textual features by applying a second modality-specific attention weight to the textual features;
and
fusing the weighted visual features and the weighted textual features, wherein fusing the weighted visual features and the weighted textual features comprises multiplying the weighted visual features and the weighted textual features;
generating, using the feature embeddings, output features for each of the plurality of documents by (i) concatenating the feature embeddings and (ii) applying a non-linear mapping to the feature embeddings;
and
training the machine learning model to generate a search result in response to a search query by applying a linear classifier to the output features.

14 more elements

A non-transitory computer-readable medium communicatively coupled to one or more processing devices and storing instructions, wherein the one or more processing devices are configured to execute the instructions and thereby perform operations comprising:

receiving pre-training data comprising a plurality of documents;
pre-training a machine learning model by:
extracting visual features and textual features from each of the plurality of documents in the pre-training data;
and
generating, using an encoder, feature embeddings for each of the plurality of documents by applying a masking function to the visual features and the textual features wherein generating the feature embeddings comprises:
generating weighted visual features by applying a first modality-specific attention weight to the visual features;
generating weighted textual features by applying a second modality-specific attention weight to the textual features;
and
fusing the weighted visual features and the weighted textual features, wherein fusing the weighted visual features and the weighted textual features comprises multiplying the weighted visual features and the weighted textual features;
generating output features for each of the plurality of documents by applying a non-linear mapping to the feature embeddings;
and
outputting a trained machine learning model configured to generate, in response to a search query, a search result comprising a semantically-relevant document.

9 more elements

Dependencies are inferred from the claim text.