Abstract: Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for training and using machine learning models. In some aspects, a method includes identifying a first set of data for users of multiple user groups. For each user, a first party user identifier is obtained that identifies the individual user to a first party content provider. A second set of data describing activity of the user with respect to content of the first party content provider is identified. For each user, a contextual analysis of the first set and the second set of data is performed to generate one or more labels indicating user interest. A training dataset is generated based on the first set and the second set of data and a label. The training dataset is then used to train one or more machine learning models to predict user interest.

Abstract: A first multi-party computation (MPC) system of an MPC cluster can receive, from an application on a client device, an inference request comprising a first share of a given user profile for a user of the application and a performance threshold. A set of nearest neighbors to the user profile can be identified by performing a secure MPC process using a trained machine learning model in collaboration with one or more second MPC systems. One or more nearest neighbors having a performance measure that satisfies the performance threshold can be selected from the set of nearest neighbors. The first MPC system can transmit data derived from the one or more nearest neighbors to the application.

Abstract: Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for selecting and distributing digital components based on predicted user attributes of users are described. In one aspect, a method includes obtaining data indicating content categories of content of the content pages accessed by the user during the user visits. A determination is made for an aggregate measure of each content category based on a quantity of user visits to content pages of the electronic resource of the publisher that included content classified as belonging to the content category. User attribute prediction data indicating previously predicted user attributes of the user is obtained. User attributes are predicted for the current visit of the user to the electronic resource of the publisher that is further used to select digital components for display with the electronic resource on a client device during the current visit.

Abstract: Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for selecting and distributing digital components based on predicted user attributes of users are described. In one aspect, a method includes obtaining data indicating content categories of content of the content pages accessed by the user during the user visits. A determination is made for an aggregate measure of each content category based on a quantity of user visits to content pages of the electronic resource of the publisher that included content classified as belonging to the content category. User attribute prediction data indicating previously predicted user attributes of the user is obtained. User attributes are predicted for the current visit of the user to the electronic resource of the publisher that is further used to select digital components for display with the electronic resource on a client device during the current visit.

Abstract: Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for selecting and distributing digital components based on predicted user attributes of users are described. In one aspect, a method includes obtaining data indicating content categories of content of the content pages accessed by the user during the user visits. A determination is made for an aggregate measure of each content category based on a quantity of user visits to content pages of the electronic resource of the publisher that included content classified as belonging to the content category. User attribute prediction data indicating previously predicted user attributes of the user is obtained. User attributes are predicted for the current visit of the user to the electronic resource of the publisher that is further used to select digital components for display with the electronic resource on a client device during the current visit.

Abstract: A first multi-party computation (MPC) system of an MPC cluster can receive, from an application on a client device, an inference request comprising a first share of a given user profile for a user of the application and a performance threshold. A set of nearest neighbors to the user profile can be identified by performing a secure MPC process using a trained machine learning model in collaboration with one or more second MPC systems. One or more nearest neighbors having a performance measure that satisfies the performance threshold can be selected from the set of nearest neighbors. The first MPC system can transmit data derived from the one or more nearest neighbors to the application.

Abstract: Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for training and using machine learning models. In some aspects, a method includes identifying a first set of data for users of multiple user groups. For each user, a first party user identifier is obtained that identifies the individual user to a first party content provider. A second set of data describing activity of the user with respect to content of the first party content provider is identified. For each user, a contextual analysis of the first set and the second set of data is performed to generate one or more labels indicating user interest. A training dataset is generated based on the first set and the second set of data and a label. The training dataset is then used to train one or more machine learning models to predict user interest.

Abstract: A computer-implemented technique can include receiving, at a server from a mobile computing device, the server having one or more processors, an image including a text. The technique can include obtaining, at the server, optical character recognition (OCR) text corresponding to the text, the OCR text having been obtained by performing OCR on the image. The technique can include identifying, at the server, non-textual context information from the image, the non-textual context information (i) representing context information other than the text itself and (ii) being indicative of a context of the image. The technique can include based on the non-textual context information, obtaining, at the server, a translation of the OCR text to a target language to obtain a translated OCR text. The technique can include outputting, from the server to the mobile computing device, the translated OCR text.

Abstract: A computer-implemented technique can include receiving, at a server from a mobile computing device, the server having one or more processors, an image including a text. The technique can include obtaining, at the server, optical character recognition (OCR) text corresponding to the text, the OCR text having been obtained by performing OCR on the image. The technique can include identifying, at the server, non-textual context information from the image, the non-textual context information (i) representing context information other than the text itself and (ii) being indicative of a context of the image. The technique can include based on the non-textual context information, obtaining, at the server, a translation of the OCR text to a target language to obtain a translated OCR text. The technique can include outputting, from the server to the mobile computing device, the translated OCR text.

Abstract: A computer-implemented technique can include receiving, at a server from a mobile computing device, the server having one or more processors, an image including a text. The technique can include obtaining, at the server, optical character recognition (OCR) text corresponding to the text, the OCR text having been obtained by performing OCR on the image. The technique can include identifying, at the server, non-textual context information from the image, the non-textual context information (i) representing context information other than the text itself and (ii) being indicative of a context of the image. The technique can include based on the non-textual context information, obtaining, at the server, a translation of the OCR text to a target language to obtain a translated OCR text. The technique can include outputting, from the server to the mobile computing device, the translated OCR text.

Abstract: Methods are disclosed for performing proper word alignment that satisfy constraints of coverage and transitive closure. Initially, a translation matrix which defines word association measures between source and target words of a corpus of bilingual translations of source and target sentences is computed. Subsequently, in a first method, the association measures in the translation matrix are factorized and orthogonalized to produce cepts for the source and target words, which resulting matrix factors may then be, optionally, multiplied to produce an alignment matrix. In a second method, the association measures in the translation matrix are thresholded, and then closed by transitivity, to produce an alignment matrix, which may then be, optionally, factorized to produce cepts. The resulting cepts or alignment matrices may then be used by any number of natural language applications for identifying words that are properly aligned.

Abstract: Methods are disclosed for performing proper word alignment that satisfy constraints of coverage and transitive closure. Initially, a translation matrix which defines word association measures between source and target words of a corpus of bilingual translations of source and target sentences is computed. Subsequently, in a first method, the association measures in the translation matrix are factorized and orthogonalized to produce cepts for the source and target words, which resulting matrix factors may then be, optionally, multiplied to produce an alignment matrix. In a second method, the association measures in the translation matrix are thresholded, and then closed by transitivity, to produce an alignment matrix, which may then be, optionally, factorized to produce cepts. The resulting cepts or alignment matrices may then be used by any number of natural language applications for identifying words that are properly aligned.