Abstract: In an embodiment, a GAN model is trained based on an image dataset. A set of images of a first class is generated by the GAN model. Further, a first saliency map of a first generated image is determined by a neural network model. A second saliency map of a second image, belonging to the first class, from image dataset is determined by the neural network model. A first interpretability coefficient is determined, based on the first and second saliency maps. A first typicality score between the first generated image and a first set of images, belonging to the first class, from the image dataset, is determined. A second typicality score between a pair of generated images is determined. A second interpretability coefficient is determined basis the first and second typicality scores. An interpretability score associated with the GAN model is determined based on the first and second interpretability coefficients.

Abstract: In an embodiment, a textual description of a situation of a first user is received. A first set of vector embeddings is determined based on the textual description. A set of ethical texts is received based on an input from a second user. A second set of vector embeddings is determined based on the set of ethical texts. A set of antonym words and a set of synonym words are determined with respect to the first set of vector embeddings, based on the second set of vector embeddings. A set of sentences is determined based on the set of antonym words and the set of synonym words. A first sentence is selected from the set of sentences based on parts-of-speech in each sentence. By using a GAN model, an image is generated based on the first sentence. The image is rendered on a display device associated with the second user.

Abstract: According to an aspect of an embodiment, operations may include retrieving a first graph. The operations may further include identifying a set of node-types, determining a first count of each of the identified set of node-types, and determining first statistical information. The operations may further include identifying a set of edge-types, determining a second count of each of the identified set of edge-types and determining a two-dimensional (2D) distribution of each of the identified set of edge-types. The operations may further include determining second statistical information, identifying a set of combinations of edge-types connecting three node-types and determining a third count of each of a set of three node-type groups.

Abstract: In an example, a method may include obtaining a language model to be audited. The method may include providing one or more common sense tests to the language model. The common sense tests may include one or more complex problems having multiple parameters or multiple answers. The common sense tests may also provide an indication of the ability of the language model to reflect laymen understanding of the world in the processed responses. The method may include obtaining model results based on responses to the language model with respect to the one or more common sense tests. The method may include obtaining one or more proposed changes to the language model based on the model results. The method may include implementing the one or more proposed changes to the language model based on the model results.

Abstract: In an embodiment, multiple datasets related to multiple application domains are received. Further, feature dependency information associated with a first dataset is determined, based on a first user input. Also, feature difference information associated with the first dataset and a second dataset is determined, based on a second user input and a set of ethical requirements. A set of structural causal models (SCMs) associated with the first dataset are determined based on the feature dependency information and the feature difference information. A set of ethical coefficients associated with the set of ethical requirements are determined based on an application of a causal transportability model on the set of SCMs. A trust score associated with the first dataset is determined based on the set of ethical coefficients. The trust score is used to train a meta-learning model associated with the multiple application domains.

Abstract: In an embodiment, operations include extracting first information about a first set of features of a first candidate, from a document or profile information of the first candidate. Second information about a second set of features, corresponding to the first set of features, is extracted from one or more databases. The second set of features is associated with a population of candidates with at least one demographic parameter same as that of the first candidate. A third set of features is determined based on difference of corresponding features from the first set of features and the second set of features. A pre-trained neural network model is applied on the third set of features to determine a set of weights associated with the third set of features. An empathy score of the first candidate is determined based on the set of weights. The empathy score of the first candidate is rendered.

Abstract: A method may include obtaining a first result of a graph explainable artificial intelligence (GXAI) classification analysis of a dataset of graph-structured data and a second result of a graph analysis algorithm that represents relationships between elements of the dataset. The method may include determining a correlation between the first result and the second result and generating a display within a graphical user interface (GUI) that visualizes similarities between the first result and the second result based on the correlation. Determining the correlation between the first result and the second result may include generating a first vector of the first result of the classification analysis using GXAI techniques and a second vector of the second result of the graph analysis algorithm. A Pearson correlation coefficient or a cosine similarity coefficients may be computed based on the first vector and the second vector in which the computed coefficients are indicative of the correlation.

Abstract: In an embodiment, a textual description of a situation of a first user is received. A first set of vector embeddings is determined based on the textual description. A set of ethical texts is received based on an input from a second user. A second set of vector embeddings is determined based on the set of ethical texts. A set of antonym words and a set of synonym words are determined with respect to the first set of vector embeddings, based on the second set of vector embeddings. A set of sentences is determined based on the set of antonym words and the set of synonym words. A first sentence is selected from the set of sentences based on parts-of-speech in each sentence. By using a GAN model, an image is generated based on the first sentence. The image is rendered on a display device associated with the second user.

Abstract: In an embodiment, each of a set of subgraphs associating an entity from an entity graph with an item is extracted from a graph database. A label score, which is an importance of an item to a respective entity is computed for each subgraph. A training dataset including the set of subgraphs and the label score for each subgraph is generated. A set of ML regression models is trained on respective entity-specific subsets of the training dataset. An ML regression model associated with a second entity generates a prediction score for an unseen graph. From the set of subgraphs, one or more subgraphs associated with the second entity are determined based on the prediction score. A recommendation for one or more items is determined, based on the one or more subgraphs. The recommendation is displayed on a user device of the first entity.

Abstract: According to an aspect of an embodiment, operations may include obtaining multiple electronic documents and obtaining a theme text. The method may also include selecting a seed text based on a semantic similarity between the seed text and the theme text. The method may also include changing a seed weight included in a weight vector that is used in identification of topics of the multiple electronic documents. The changed seed weight may bias the identification of topics of the plurality of electronic documents in favor of the seed text as compared to one or more other text strings of the weight vector. The method may also include generating, a representation of a topic model for display to a user, the topic model may be based on the multiple electronic documents and the weight vector.

Abstract: In an embodiment, a GAN model is trained based on an image dataset. A set of images of a first class is generated by the GAN model. Further, a first saliency map of a first generated image is determined by a neural network model. A second saliency map of a second image, belonging to the first class, from image dataset is determined by the neural network model. A first interpretability coefficient is determined, based on the first and second saliency maps. A first typicality score between the first generated image and a first set of images, belonging to the first class, from the image dataset, is determined. A second typicality score between a pair of generated images is determined. A second interpretability coefficient is determined basis the first and second typicality scores. An interpretability score associated with the GAN model is determined based on the first and second interpretability coefficients.

Abstract: In an embodiment, operations include receiving first information associated with a first person and a first request of the first person to one or more institutions. A set of attributes of the first person is extracted and used to construct a causal model. The causal model represents causal relationships amongst attributes of the set of attributes. For the first person, a utility function associated with each of a plurality of AI models associated with the one or more institutions is determined. The utility function is determined based on the causal model, a first set of empathy criteria associated with the first person, and a second set of empathy criteria associated with each of the one or more institutions. For the utility function, optimin-point information is determined to reconcile a plurality of decisions taken by the plurality of AI models for the first request. The reconciled decision is rendered.

Abstract: Operations include obtaining a user graph of users of a social network, obtaining a content graph that indicates links between the users and content items interacted with by the users, and obtaining a resource graph that indicates links between the content items and external resources. The operations include generating first user representations, first content representations, and first resource representations. The operations include generating second resource representations based on the first content representations, the first resource representations, and the resource graph; generating second content representations based on the first content representations, the first user representations, the content graph, and the second resource representations; and generating second user representations based on the first user representations, the user graph, the first content representations, and the content graph.

Abstract: In an embodiment, operations include extracting first information about a first set of features of a first candidate, from a document or profile information of the first candidate. Second information about a second set of features, corresponding to the first set of features, is extracted from one or more databases. The second set of features is associated with a population of candidates with at least one demographic parameter same as that of the first candidate. A third set of features is determined based on difference of corresponding features from the first set of features and the second set of features. A pre-trained neural network model is applied on the third set of features to determine a set of weights associated with the third set of features. An empathy score of the first candidate is determined based on the set of weights. The empathy score of the first candidate is rendered.

Abstract: A method may include obtaining multiple electronic documents and multiple topics associated with the electronic documents. The method may further include determining a similarity between a first topic and a second topic. The first topic may be associated with a first set of electronic documents. The method may further include refining the multiple topics based on the similarity between the first topic and the second topic by associating the first set of the electronic documents with the second topic and removing the first topic from the multiple topics. The method may further include building a document-classifier model by applying machine learning to at least one electronic document associated with each of the refined topics. The method may further include obtaining an electronic document and classifying the electronic document into one of the refined topics using the document-classifier model.

Abstract: A method includes constructing a hierarchal graph associated with a document. The hierarchal graph includes a document node, a set of paragraph nodes, a set of sentence nodes, and a set of token nodes. The method further includes determining, based on a language attention model, a set of weights associated with a set of edges between a first node and each connected second set of nodes. The method further includes applying a GNN model on the hierarchal graph based on a set of first features associated with each token node, and the set of weights. The method further includes updating a set of features associated with each node based on the application, and generating a document vector for an NLP task, based on the updated set of features. The method further includes displaying an output of the NLP task for the document, based on the document vector.

Abstract: Operations may include identifying a plurality of graphs as ground truth graphs in response to each ground truth graph having a heuristic characteristic and being categorized as a first- or second-class graph based on labeling of the graphs with respect to the heuristic characteristic. The operations may include identifying a graph as an unlabeled graph, the graph being unlabeled with respect to the heuristic characteristic. The operations may include comparing the unlabeled graph to the first- and second-class graphs, the comparing being based on the heuristic characteristic and including one or more operations selected from a group of operations including performing similarity matching, model-based heuristics operations, or query analysis operations. The operations may include categorizing the unlabeled graph as a first- or second-class graph based on the comparing. The operations may include training a machine learning model using the ground truth graphs and the previously unlabeled graph.

Abstract: A method may include obtaining a machine-learning model trained with respect to a subject. The machine-learning model may be based on a plurality of factors that correspond to the subject. The method may include obtaining human provided information regarding the subject. The expert information may indicate relationships between the plurality of factors with respect to how the plurality of factors affect each other. The method may include generating a structural causal model that represents the relationships between the plurality of factors based on the expert information. The method may include identifying, as a confounding factor and based on the structural causal model, a factor of the plurality of factors that causes a confounding bias in the machine-learning model. The method may include estimating the confounding bias based on the identified confounding factor.

Abstract: A method includes storing a research paper and a set of candidate resources including media content. The method further includes encoding each of one or more first content fields in the research paper into a first vector based on a first field type associated with each of the one or more first content fields. The method further includes encoding each of one or more second content fields in each of the parsed set of candidate resources into a second vector, based on a second field type associated with each of the one or more second content fields. The method further includes comparing the first vector with the second vector to determine a final set of resources based on the comparison. The method further includes controlling a display screen to output the determined final set of resources and the research paper.

Abstract: A method of automatically estimating the difficulty level of online content. The method includes receiving, by one or more processors, electronic content which is accessible via an online network and which is at least temporarily stored in at least one non-transitory storage media, extracting, by the one or more processors, structural and non-linguistic information from the electronic content. The method further includes extracting, by the one or more processors, linguistic information from the electronic content. The method further includes generating, by the one or more processors, a difficulty estimation for the content based on the structural and non-linguistic information in addition to the linguistic information. The method further includes generating, by the one or more processors, a recommendation of a subset of the electronic content to a user based on the difficulty estimation.