Abstract: Aspects of the present disclosure relate to systems and methods for performing targeted searching based on a user profile. In examples, a user profile including a user embedding may be retrieved based on the receipt of a user indication. The user embedding may be created based on one or more user interest. A plurality of document embeddings may be identified based on the user embedding, where each document embedding of the plurality of document embeddings is determined to be within a first distance of the user embedding. In examples, a ranking for each document embedding of the plurality of document embeddings may be generated, where the ranking for each document embedding of the plurality of document embeddings is based on the user embedding. At least one document may be recommend based on a ranking associated with a document embedding.

Abstract: Aspects of the present disclosure relate to systems and methods for performing targeted searching based on a user profile. In examples, a user profile including a user embedding may be retrieved based on the receipt of a user indication. The user embedding may be created based on one or more user interest. A plurality of document embeddings may be identified based on the user embedding, where each document embedding of the plurality of document embeddings is determined to be within a first distance of the user embedding. In examples, a ranking for each document embedding of the plurality of document embeddings may be generated, where the ranking for each document embedding of the plurality of document embeddings is based on the user embedding. At least one document may be recommend based on a ranking associated with a document embedding.

Abstract: Aspects of the present disclosure relate to systems and methods for performing targeted searching based on a user profile. In examples, a user profile including a user embedding may be retrieved based on the receipt of a user indication. The user embedding may be created based on one or more user interest. A plurality of document embeddings may be identified based on the user embedding, where each document embedding of the plurality of document embeddings is determined to be within a first distance of the user embedding. In examples, a ranking for each document embedding of the plurality of document embeddings may be generated, where the ranking for each document embedding of the plurality of document embeddings is based on the user embedding. At least one document may be recommend based on a ranking associated with a document embedding.

Abstract: A system for extracting key phrase candidates from a corpus of documents, including a processor, a memory, and a program executing on the processor. The system is configured to run a key phrase model to extract one or more key phrase candidates from each document in the corpus and convert each extracted key phrase candidate into a feature vector. The key phrase model also filters the feature vectors to remove duplicates using a classifier that was trained on a set of key phrase pairs with manual labels indicating whether two key phrases are duplicates of each other, to produce remaining key phrase candidates. The system uses the remaining key phrase candidates in a computer-implemented application.

Abstract: 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.

Abstract: To provide automated categorization of structured textual content individual nodes of textual content, from a document object model encapsulation of the structured textual content, have a multidimensional vector associated with them, where the values of the various dimensions of the multidimensional vector are based on the textual content in the corresponding node, the visual features applied or associated with the textual content of the corresponding node, and positional information of the textual content of the corresponding node. The multidimensional vectors are input to a neighbor-imbuing neural network. The enhanced multidimensional vectors output by the neighbor-imbuing neural network are then be provided to a categorization neural network. The resulting output can be in the form of multidimensional vectors whose dimensionality is proportional to categories into which the structured textual content is to be categorized. A weighted merge takes into account multiple nodes that are grouped together.

Abstract: A system for extracting a key phrase from a document includes a neural key phrase extraction model (“BLING-KPE”) having a first layer to extract a word sequence from the document, a second layer to represent each word in the word sequence by ELMo embedding, position embedding, and visual features, and a third layer to concatenate the ELMo embedding, the position embedding, and the visual features to produce hybrid word embeddings. A convolutional transformer models the hybrid word embeddings to n-gram embeddings, and a feedforward layer converts the n-gram embeddings into a probability distribution over a set of n-grams and calculates a key phrase score of each n-gram. The neural key phrase extraction model is trained on annotated data based on a labeled loss function to compute cross entropy loss of the key phrase score of each n-gram as compared with a label from the annotated dataset.

Abstract: Technologies pertaining to electronic document understanding are described herein. A document is received, wherein the document includes a section of a type. An image of the document is generated, and a candidate region is identified in the image of the document, wherein the candidate region encompasses the section. A label is assigned to the candidate region based upon text of the section, wherein the label identifies the type of the section. An electronic document understanding task is performed based upon the label assigned to the candidate region.

Abstract: Aspects of the present disclosure relate to systems and methods for performing targeted searching based on a user profile. In examples, a user profile including a user embedding may be retrieved based on the receipt of a user indication. The user embedding may be created based on one or more user interest. A plurality of document embeddings may be identified based on the user embedding, where each document embedding of the plurality of document embeddings is determined to be within a first distance of the user embedding. In examples, a ranking for each document embedding of the plurality of document embeddings may be generated, where the ranking for each document embedding of the plurality of document embeddings is based on the user embedding. At least one document may be recommend based on a ranking associated with a document embedding.

Abstract: A technique produces a graph data structure based on at least partially unstructured information dispersed over web documents. The technique involves applying a machine-trained model to a set of documents (or, more generally “document units”) to identify topics in the documents. The technique then generates count information by counting the occurrences of the single topics and co-occurrences of parings of topics in the documents. The technique generates conditional probability information based on the count information. An instance of conditional probability information describes a probability that a first topic will occur, given an appearance of a second topic, and a probability that the second topic will occur, given an appearance of the first topic. The technique then formulates the conditional probability information in a graph data structure. The technique also provides an application system that utilizes the graph data structure to provide any kind of computer-implemented service to a user.

Abstract: 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.

Abstract: A system for extracting a key phrase from a document includes a neural key phrase extraction model (“BLING-KPE”) having a first layer to extract a word sequence from the document, a second layer to represent each word in the word sequence by ELMo embedding, position embedding, and visual features, and a third layer to concatenate the ELMo embedding, the position embedding, and the visual features to produce hybrid word embeddings. A convolutional transformer models the hybrid word embeddings to n-gram embeddings, and a feedforward layer converts the n-gram embeddings into a probability distribution over a set of n-grams and calculates a key phrase score of each n-gram. The neural key phrase extraction model is trained on annotated data based on a labeled loss function to compute cross entropy loss of the key phrase score of each n-gram as compared with a label from the annotated dataset.

Abstract: Technologies pertaining to ranking webpages in response to receipt of a query are described. A search engine receives a query and identifies webpages that are germane to the query. The search engine ranks the identified webpages to form a ranked list, wherein a first webpage is positioned in the ranked list based upon a static score assigned to the first webpage. The static score is based upon a weight assigned to a hyperlink in a second webpage, wherein the hyperlink points to the first webpage, and further wherein the weight is based upon a value of a feature of the hyperlink, such as a location of the hyperlink on the second webpage when the second webpage is rendered. Further, the second webpage includes several hyperlinks that point to different webpages, wherein each of the several hyperlinks has a different weight assigned thereto.

Abstract: A system for extracting a key phrase from a document includes a neural key phrase extraction model (“BLING-KPE”) having a first layer to extract a word sequence from the document, a second layer to represent each word in the word sequence by ELMo embedding, position embedding, and visual features, and a third layer to concatenate the ELMo embedding, the position embedding, and the visual features to produce hybrid word embeddings. A convolutional transformer models the hybrid word embeddings to n-gram embeddings, and a feedforward layer converts the n-gram embeddings into a probability distribution over a set of n-grams and calculates a key phrase score of each n-gram. The neural key phrase extraction model is trained on annotated data based on a labeled loss function to compute cross entropy loss of the key phrase score of each n-gram as compared with a label from the annotated dataset.

Abstract: To provide automated categorization of structured textual content individual nodes of textual content, from a document object model encapsulation of the structured textual content, have a multidimensional vector associated with them, where the values of the various dimensions of the multidimensional vector are based on the textual content in the corresponding node, the visual features applied or associated with the textual content of the corresponding node, and positional information of the textual content of the corresponding node. The multidimensional vectors are input to a neighbor-imbuing neural network. The enhanced multidimensional vectors output by the neighbor-imbuing neural network are then be provided to a categorization neural network. The resulting output can be in the form of multidimensional vectors whose dimensionality is proportional to categories into which the structured textual content is to be categorized. A weighted merge takes into account multiple nodes that are grouped together.

Abstract: A technique produces a graph data structure based on at least partially unstructured information dispersed over web documents. The technique involves applying a machine-trained model to a set of documents (or, more generally “document units”) to identify topics in the documents. The technique then generates count information by counting the occurrences of the single topics and co-occurrences of parings of topics in the documents. The technique generates conditional probability information based on the count information. An instance of conditional probability information describes a probability that a first topic will occur, given an appearance of a second topic, and a probability that the second topic will occur, given an appearance of the first topic. The technique then formulates the conditional probability information in a graph data structure. The technique also provides an application system that utilizes the graph data structure to provide any kind of computer-implemented service to a user.

Abstract: Technologies pertaining to electronic document understanding are described herein. A document is received, wherein the document includes a section of a type. An image of the document is generated, and a candidate region is identified in the image of the document, wherein the candidate region encompasses the section. A label is assigned to the candidate region based upon text of the section, wherein the label identifies the type of the section. An electronic document understanding task is performed based upon the label assigned to the candidate region.

Abstract: Technologies pertaining to ranking webpages in response to receipt of a query are described. A search engine receives a query and identifies webpages that are germane to the query. The search engine ranks the identified webpages to form a ranked list, wherein a first webpage is positioned in the ranked list based upon a static score assigned to the first webpage. The static score is based upon a weight assigned to a hyperlink in a second webpage, wherein the hyperlink points to the first webpage, and further wherein the weight is based upon a value of a feature of the hyperlink, such as a location of the hyperlink on the second webpage when the second webpage is rendered. Further, the second webpage includes several hyperlinks that point to different webpages, wherein each of the several hyperlinks has a different weight assigned thereto.

Abstract: A technique is described herein for processing network-accessible documents in a scalable and resource-efficient manner. A model-generating process provided by the technique includes three-phases. A first phase generates a set of sample documents associated with a particular class of documents, a second phase applies labels to the sample documents to produce a set of labeled documents, and a third phase generates at least one data-extraction model based on the set of labeled documents. The data-extraction model includes data-extracting logic for extracting at least one specified data item from new documents that match the class of documents. In a data-extracting process, the technique identifies a data-extraction model that applies to the new document and then applies that model.

Abstract: Configurations herein comprise a multi-layer framework to extract document structural data. The framework extracts structural data from raw, unstructured, electronic documents, for example, .pdf documents. Structural data refers to the semantic elements, for example, paragraphs, lists, tables, titles etc. that may be visible in the displayed document but not described in electronic data.