Abstract: A computing system generates a plurality of training data sets for generating the NLP model. The computing system trains a teacher network to extract and classify tokens from a document. The training includes a pre-training stage where the teacher network is trained to classify generic data in the plurality of training data sets and a fine-tuning stage where the teacher network is trained to classify targeted data in the plurality of training data sets. The computing system trains a student network to extract and classify tokens from a document by distilling knowledge learned by the teacher network during the fine-tuning stage from the teacher network to the student network. The computing system outputs the NLP model based on the training. The computing system causes the NLP model to be deployed in a remote computing environment.

Abstract: A method including building a graph data structure storing network data from a relational data structure that stores sequential data describing object identifiers and relationships between the object identifiers. The method also includes generating, from the sequential data, a features matrix for the object identifiers. The method also includes building a machine learning model layer including a long short-term memory neural network (LSTM) programmed to take, as input, the features matrix and to generate, as output, a prediction vector. The method also includes building machine learning model layers including graph convolutional neural network (GCN) layers. The machine learning model layers is programmed to take, as input, the graph data structure and the prediction vector, and generate, as output, a future prediction regarding the sequential data. The method also includes combining, into a machine learning model ensemble, the machine learning model layer and the machine learning model layers.

Abstract: A method including detecting, in a written electronic communication, an input sentence satisfying a readability metric threshold. The method also includes transforming, by a sentence transformer model, the input sentence to output suggested sentences. The method also includes evaluating the suggested sentences along a set of acceptability criteria. The method also includes determining, based on the evaluating, that the set of acceptability criteria is satisfied. The method also includes modifying, based on determining that the set of acceptability criteria is satisfied, the written electronic communication with the suggested sentences to obtain a modified written electronic communication. The method also includes returning the modified written electronic communication.

Abstract: A method implements efficient real time serving of ensemble models. The method includes receiving an input and processing the input with an abridged model to generate a set of component scores and an abridged score. The method further includes processing the set of component scores with a deviation threshold to select one of the abridged score and an ensemble score as an output and presenting the output.

Abstract: Systems and methods or determining tax recommendations for a taxpayer by using a tax calculation graph to identify tax variables that a taxpayer can control and modify, including a recommendation engine configured to analyze a tax calculation graph which is calculated using tax data of the taxpayer. An identified tax variable can be analyzed by determining nodes of the graph affecting a value of the identified tax variable, providing a user interface enabling at least one modification to the nodes, and determining an effect on the identified tax variable due to the at least one modification.

Abstract: Methods and computer systems are provided for generating synthetic data. A real vector is generated representing real data. Using a classification model, a first output vector that represents a first class is generated from the real vector. the real vector is mutated to generate a counterfactual vector. using the classification model, the second output vector that represents a second class is generated from the counterfactual vector. the counterfactual vector is then mutated to generate a synthetic vector. Using the classification model, a third output vector that corresponds to the first class is generated from the synthetic vector, synthetic data is generated from the synthetic vector.

Abstract: A method including receiving an input and embedding the input into a first data structure that defines first relationships among images and texts. The method also includes comparing the first data structure to an index including a second data structure that defines second relationships among pre-determined texts and pre-determined images. The method also includes returning a subset of images from the pre-determined images. The subset includes those images in the pre-determined images for which matches exist between the first relationships and the second relationships.

Abstract: Systems and methods of the present disclosure provide processes for determining how much to adjust machine-learning parameter values in a direction of a gradient for gradient-descent steps in training processes for machine-learning models. Current parameter values of a machine-learning model are vector components that define an initial estimate for a local extremum of a cost function used to measure how well the machine-learning model performs. The initial estimate and the gradient of the cost function for the initial estimate are used to define an auxiliary function. A root estimate is determined for the auxiliary function of the gradient. The parameters are adjusted in the direction of the gradient by an amount specified by the root estimate.

Abstract: A method including receiving a command to display a modal dialog. The modal dialog is displayed using both first and second scrolling frames. The first scrolling frame permits scrolling when a modal dialog height exceeds a first scrolling frame constraint. The second scrolling frame permits scrolling of a content section when a content section height exceeds a second scrolling frame constraint. The first scrolling frame constraint has a first and second priorities. The second scrolling frame constraint has a third priority. An orientation of the display screen is determined as being either in a portrait orientation or a landscape orientation. Responsive to determining the physical orientation, an applicable priority that is applicable to the first scrolling frame constraint is assigned. The applicable priority is the first priority in the portrait orientation, and is the second priority in the landscape orientation. After assigning the applicable priority, the modal dialog is displayed.

Abstract: Systems and methods are disclosed for model based document image enhancement. Instead of requiring paired dirty and clean images for training a model to clean document images (which may cause privacy concerns), two models are trained on the unpaired images such that only the dirty images are accessed or only the clean images are accessed at one time. One model is a first implicit model to translate the dirty images from a source space to a latent space, and the other model is a second implicit model to translate the images from the latent space to clean images in a target space. The second implicit model is trained based on translating electronic document images in the target space to the latent space. In some implementations, the implicit models are diffusion models, such as denoising diffusion implicit models based on solving ordinary differential equations.

Abstract: A method includes training, using first real data objects, a generative adversarial network having a generator model and a discriminator model to create a trained generator model that generates realistic data, and training, using adversarial data objects and second real data objects, the discriminator model to output an authenticity binary class for the adversarial data objects and the second real data objects. The method further includes deploying the discriminator model to a production system. In the production system, the discriminator model outputs the authenticity binary class to a system classifier model.

Abstract: Systems and methods for inferring recommendations and experiences for anonymous users of an online website are disclosed. Anonymous users of the online website are assigned anonymous user identifiers, and the browsing activity of the anonymous users is converted into features and aggregated over time. The anonymous users' interactions are monitored and used to generate labels that are combined with the feature dataset to produce a training dataset which is used to train a machine learning model. The browsing activity of an anonymous user may be converted into features and aggregated over time and fed into the trained machine learning model from which personalized experiences and recommendations may be generated and provided to the anonymous user.

Abstract: A method including receiving a selected domain from a set of domains. The method also includes selecting, based on the selected domain, a selected machine learning model from among a set of machine learning models. Each of the machine learning models is configured to receive, as input, a dataset of past time-dependent data and generate, as output, a corresponding predicted quality measure for each of a number of time periods. The selected machine learning model is trained using training data generated for an entity corresponding to the domain. The method also includes executing the selected machine learning model on the dataset to generate predicted quality measures for the time periods. The method also includes generating, using the predicted quality measures, a schedule for executing a computer process. The method also includes presenting the schedule.

Abstract: A method implements efficient counterfactual search. The method includes receiving a request corresponding to an input vector, processing the input vector with a model to generate an output vector that does not correspond to a selected class, and processing the input vector using a component, of a plurality of components, to generate a counterfactual vector to the selected class. The plurality of components includes a number of dimensions that is less than a number of features of the input vector. The method further includes processing the counterfactual vector to generate a recommendation and presenting the recommendation.

Abstract: Disambiguity in large language models (LLMs) includes receiving an original query in a user interface, generating an ambiguity query from the original query, and sending, via an application programming interface (API) of an LLM, the ambiguity query to the LLM. The ambiguity query includes the original query and training the LLM to recognize ambiguities. The method further includes receiving, via the API and responsive to the ambiguity query, a binary response and detecting, based at least in part on the binary response, the original query as ambiguous. Disambiguity may include detecting an ambiguity location in the original query using perturbed queries and the LLM.

Abstract: A method and system assists train a classifier model with a machine learning process. The method and system trains the classifier with a labeled training set and with an unlabeled training set. The method and system trains the classifier model to correctly classify data items that fall within a distribution of the labeled training set. The method and system trains the classifier to indicate a lack of confidence in classification for data items that do not fall within the distribution of the labeled training set.

Abstract: Aspects of the present disclosure provide techniques for behavior prediction. Embodiments include receiving activity data of a user, identifying user sessions comprising sets of time-stamped actions in the activity data, and segmenting the activity data into subsets corresponding to the user sessions. Embodiments include providing the subsets as inputs to a hierarchical attention time-series (HAT) model comprising: a first layer that determines attention scores for respective time-stamped actions in the subsets; and a second layer that determines attention scores for the subsets based on aggregations of the attention scores for the respective time-stamped actions.

Abstract: Certain aspects of the present disclosure provide techniques for generating a metric, include receiving a rule defining one or more text strings; determining a set of transactions based on a user attribute; determining a first subset of transactions; determining a second subset of transactions; generating a first categorical distribution based on each transaction of the first subset of transactions being associated with a transaction description containing at least one text string of the one or more text strings; calculating a first unity metric based on the first categorical distribution; generating a second categorical distribution based on each transaction of the second subset of transactions being associated with a transaction description that does not contain a text string of the one or more text strings; calculating a second unity metric based on the second categorical distribution; determining a reliability metric for the rule; and providing the reliability metric.

Abstract: A method including receiving, at a platform and from a first user using a first user device, selection of a uniform resource indicator (URI) unique to a second user using a second user device. The method also includes generating, automatically by the platform in response to receiving the URI, a conference session unique to the first user and the second user. The method also includes transmitting, automatically by the platform, a message to the second user, the message indicating that the conference session is initiated. The method also includes receiving, by the platform, an indication from the second user device that the second user joins the conference session. The method also includes joining, automatically by the platform, the first user device and the second user device in the conference session.

Abstract: A method may include executing a baseline classifier on unreviewed transaction features of an unreviewed transaction record to obtain a baseline account identifier, and executing a comparison model on (i) an unreviewed transaction vector of the unreviewed transaction record and (ii) reviewed transaction vectors to obtain comparison scores. The reviewed transaction vectors may correspond to reviewed transaction records each having a user-approved account identifier. The method may further include selecting, using the comparison scores, a reviewed transaction record. The reviewed transaction record may correspond to a comparison score. The comparison score may correspond to a user-approved account identifier of the reviewed transaction record.