Abstract: Techniques for training a language model for code switching content are disclosed. Such techniques include, in some embodiments, generating a dataset, which includes identifying one or more portions within textual content in a first language, the identified one or more portions each including one or more of offensive content or non-offensive content; translating the identified one or more salient portions to a second language; and reintegrating the translated one or more portions into the textual content to generate code-switched textual content. In some cases, the textual content in the first language includes offensive content and non-offensive content, the identified one or more portions include the offensive content, and the translated one or more portions include a translated version of the offensive content. In some embodiments, the code-switched textual content is at least part of a synthetic dataset usable to train a language model, such as a multilingual classification model.

Abstract: Systems and methods are provided for performing automated speech recognition. The systems and methods access a LM that includes a plurality of n-grams, each of the plurality of n-grams comprising a respective sequence of words and corresponding LM score and receive a list of words associated with a group classification, each word in the list of words being associated with a respective weight. The systems and method compute, based on the LM scores of the plurality of n-grams, a probability that a given word in the list of words associated with the group classification appears in an n-gram in the LM comprising an individual sequence of words and adds one or more new n-grams to the LM comprising one or more words in the list of words in combination with the individual sequence of words and associated with a particular LM score based on the computed probability.

Abstract: Methods and systems of displaying substring pairs where visual characteristics delineate adjacent substring pairs from each other, specifically in the case of at least one of the primary text string and the secondary text string, the placement of the first substring alternates position on an electronic display above and below the second substring. The method may comprise receiving a plurality of the primary substrings, a plurality of the secondary substrings, and a plurality of visual characteristics, displaying, on an electronic display, the primary substrings and the secondary substrings arranged into substring pairs, and one of the visual characteristics in each of the correspondence areas. Additional desired visual effects may be achieved through the use of specific demarcations, demarcation placements, and substring modifications.

Abstract: Aspects of the disclosure relate to evaluating sources of training data for model generation. A computing platform may receive, from one or more data sources, a labelled data set. The computing platform may apply, to the labelled data set, an unsupervised learning algorithm, resulting in a clustered data set. The computing platform may compare, for each data point in the labelled data set, corresponding clustering information and labelling information to identify discrepancies. The computing platform may flag, for data points with identified discrepancies between the clustering information and labelling information, a labelling error. The computing platform may grade, based on the flagged labelling errors, each of the one or more data sources. Using remaining data of the labelled data set, not flagged with labelling errors, the computing platform may train a supervised learning model by weighting the remaining data based on: a corresponding data source and its grade.

Abstract: A character input device according to one or more embodiments may include: an output unit configured to output a first character string to an application program having a suggestion function; a detector configured to detect selection of a second character string that is presented in correspondence with the first character string by the application program; and a registration unit configured to register, in a dictionary database, the second character string that is detected, by the detector, to have been selected.

Abstract: A method and/or system for query expansion may include: providing a set of training data in a given domain in the form of training question texts and training answer texts, identifying disjoint answer words in the training answer text that do not occur in the associated training question text, generating a graph of question word nodes and answer word nodes generated from the set of training data for the given domain in the form of the training question texts and the training answer texts, wherein edges are provided between a disjoint pair of a question word node for a question word in a training question and an answer word node for a disjoint answer word in an associated training answer, and applying spreading activation through the graph to result in a top n most highly activated nodes that are used as candidate words for expansion of a user query input.

Abstract: A computer implemented method is used for changing a password in a multi-domain environment. The method includes obtaining a private key and a public key from a security card at a user device in a user domain, transferring the public key to a controller in a secure domain, requesting a password change, receiving a public key encrypted new password from the secure domain, and decrypting the new password using the private key.

Abstract: Techniques are described for generating an encoded-string automaton for a regex pattern from a decoded-string automaton of the regex pattern. In an embodiment, the process obtains a decoded-string automaton of the regex pattern and applies unique decoded string value(s) from the dictionary of the encoding. When applied at a selected state in the decoded-string automaton, the application may yield a transition to at least one target state in the decoded-string automaton for a unique dictionary value. Such a transition generates a transition in the encoded-string automaton from an encoded state corresponding to the selected state in the decoded-string automaton to a target state in the encoded-string automaton corresponding to the target state in the decoded-string automaton. The generated transition in the encoded-string automaton is conditioned on the token of the unique decoded string value in the dictionary.

Abstract: A system for protecting an endpoint device of a user includes a web interface module that identifies a present URL visited by the user and target URLs to which navigation is available. A password management module installed on the endpoint device stores multiple entries. One entry includes a username, a password, and a login URL. The password management module selectively supplies credentials to the web interface module, including supplying the password to the web interface module in response to the web interface module identifying the login URL as the present URL. A URL analysis module evaluates the target URLs to classify each of the target URLs as either safe or suspicious and initiates a warning to the user in response to one of the target URLs being classified as suspicious. The URL analysis module performs the classification based in part on login URLs stored by the password management module.

Abstract: A method for processing formulae includes encoding a formula by: training, with a server, a model by using a machine learning algorithm with a data set that includes a plurality of formulae; transforming, with a processor, a first formula into a tree format using the trained model; converting, with the processor, the tree format of the first formula into a plurality of lists; and encoding, with the processor, the plurality of lists into a fixed dimension vector by leveraging a stacked attention module; and generating one or more formula candidates by: obtaining, with the processor, input information; and generating, with the processor, one or more second formula candidates based on input information by using the stacked attention module with a tree beam search algorithm.

Abstract: Disclosed is a method of correcting text information. The method can be performed by a computing device. The method includes obtaining the text information. The method includes determining problem text within the text information. The method includes generating alternative text to correct the problem text by utilizing expanded text associated with the problem text or non-text type information associated with the problem text. The method includes providing information about the alternative text for correcting the problem text.

Abstract: A computer implemented method for managing search queries. The method uses a number of processor units to receive data records. The number of processor units identify a set of data record pairs from the data records. The number of processor units generates a list of long data records based on frequencies of occurrences for long data records associated with each short data record in the set of data record pairs. The number of processor units receive a search query comprises a number of short data records in the set of data record pairs. The number of processor units identify a number of long data records for each short data record in the number of short data records using the lists of long data records for short data records. The number of processor units expand the search query by adding the number of long data records to the search query.

Abstract: A parameter learning apparatus 100 extracts one entity in a document and a related text representation as a one-term document fact, outputs a one-term partial predicate fact including only the one entity using a predicate fact that includes entities and a predicate, calculates a first one-term score indicating the degree of establishment of the one-term document fact using a one-term partial predicate feature vector, a one-term text representation feature vector, and a one-term entity feature vector that are calculated from parameters, calculates a second one-term score with respect to a combination of one entity and a predicate or a text representation that is not extracted as the one-term partial predicate fact, updates the parameters such that the first one-term score is higher than the second one-term score, and calculates a score indicating the degree of establishment of the predicate fact and a score indicating the degree of establishment of a combination of entities and a predicate that is not obtained

Abstract: Systems and methods of the present disclosure enable context-aware haptic error notifications. The systems and methods include a processor to receive input segments into a software application from a character input component and determine a destination. A context identification model predicts a context classification of the input segments based at least in part on the software application and the destination. Potential errors are determined in the input segments based on the context classification. An error characterization machine learning model determines an error type classification and an error severity score associated with each potential error and a haptic feedback pattern is determined for each potential error based on the error type classification and the error severity score of each potential error of the one or more potential errors. And a haptic event latency is determined based on the error type classification and the error severity score of each potential error.

Abstract: Implementations of the present disclosure relate to methods, devices, and computer program products for generating a destination vocabulary from a source vocabulary. In a method, a group of candidate vocabularies are determined from the source vocabulary based on a corpus, a size of a candidate vocabulary in the group of candidate vocabularies being different from a size of the source vocabulary. A group of marginal scores are obtained for the group of candidate vocabularies, respectively, a marginal score in the group of marginal scores being obtained for the candidate vocabulary based on a corpus entropy of the candidate vocabulary and a size of the candidate vocabulary. The destination vocabulary is selected from the group of candidate vocabularies based on the group of marginal scores.

Abstract: Polarity classifications of writing samples are obtained by sentiment analysis operations including embedding each word of a writing sample into a word vector based on surrounding words, extracting one or more features of the writing sample, applying a feature learning function to the one or more features, estimating a polarity of the writing sample based on output from the word learning function and output from the feature learning function, and training the word learning function and the feature learning function based on a loss function relating the estimated polarity to the word vector to produce a model for writing sample polarity classification.

Abstract: Systems, methods, and computer-readable media are disclosed for list attribute normalization and standardization for creation of a controlled vocabulary. A vocabulary set comprising a plurality of vocabulary term may be received. For each vocabulary term, semantic duplicates may be identified. The semantic duplicates may be identified by analyzing semantics, syntactics, or phonetics of the vocabulary terms. Semantic chains may be formed from each vocabulary term and the corresponding semantic duplicates. The terms in each semantic chain may be ranked to determine a most probable vocabulary term. The most probable vocabulary term may then replace the semantic chain. The most probable vocabulary term across all semantic chains from the vocabulary set may form the controlled vocabulary.

Abstract: A system may receive a data glossary comprising a list of terms. The system may then measure a usage dimension for a set of the terms from the list of terms. The system may select a candidate term from the set based on the usage dimension and perform a maintenance action on the candidate terms.

Abstract: There is disclosed a method and system for classifying a word as an obscene word, the method comprising, at a training phrase: acquiring a first word, the first word corresponding to a given obscene word; generating a first set of misspelled words, the first set of misspelled words comprising a plurality of misspelled variations of the first word; generating a training pairs, the training pairs comprising: a set of positive training pairs comprising the first word paired with each misspelled variations of the first word; training a machine learning algorithm, the training comprising: determining, for each training pairs, a set of features representative of a property of the training pairs; generating an inferred function based on the set of features, the inferred function being configured to assign, in use, an indecency score, the decency score being indicative of a likelihood of the word being obscene.

Abstract: Systems, apparatuses, methods, and computer program products are disclosed for processing electronic information indicative of natural language. An example method includes generating a natural language attribute data set based on a first word in a sequence of words provided by a user, a first natural language of the word, and one or more exogenous events. The example method further includes generating a natural language transliteration data set based on the natural language attribute data set. The example method further includes generating a translation of the first word in a second natural language based on the natural language transliteration data set. The example method further includes generating, using machine learning and based at least in part on the translation, a response signal for transmission to a client device.

Abstract: Techniques are disclosed for detecting distributed incompetence in text of a conversation using communicative discourse trees and then inserting an automatic response from an autonomous agent (chatbot) or other entity. For example, a computing system generates a communicative discourse tree from utterances from multiple agents to a user. The computing system obtains a prediction of whether the text includes distributed incompetence by applying a trained predictive model to the communicative discourse tree. Based on the detection, the computing system generates an updated response to a user device.

Abstract: A computer-implemented method according to one embodiment includes receiving a plurality of linguistic expressions (LEs); changing one or more conditions of the plurality of linguistic expressions to create an updated plurality of linguistic expressions, utilizing a visual exploration framework (VEF) that visually presents to a user each of the plurality of linguistic expressions; and including the updated plurality of linguistic expressions in a model used to classify input sentences. According to another embodiment, a computer-implemented method includes receiving (i) a set of linguistic expressions (LEs) and (ii) a set of labeled data as input, where the LEs are logical combinations of predicates learned from the labeled data, and each data point in the labeled data comprises a piece of text and ground-truth labels; presenting the LEs in a visual exploration framework; and allowing a user to sort, filter, subset, and select LEs based on different criteria, utilizing the framework.

Abstract: A learning method to be executed by a computer, the learning method includes when a first input sentence in which a predetermined target is represented by a first named entity is input to a first machine learning model, learning a first parameter of the first machine learning model such that a value output from the first machine learning model approaches correct answer information corresponding to the first input sentence; and when an intermediate representation generated when the first input sentence is input to the first machine learning model and a second input sentence in which the predetermined target is represented by a second named entity are input to a second machine learning model, learning the first parameter and a second parameter of the second machine learning model such that a value output from the second machine learning model approaches correct answer information corresponding to the second input sentence.

Abstract: Disclosed is a method and an apparatus NER-orientated Chinese clinical text data augmentation, and unannotated data and annotated data of label linearization processing through data preprocessing. A concealed part is predicted based on retained information by using the unannotated data and concealing part of information in text, and meanwhile an entity word-level discrimination task is introduced for pre-training of a span-based language model; and a plurality of decoding mechanisms are introduced in a fine-tune stage, a relationship between a text vector and text data is obtained based on the pre-trained span-based language model, linearized data with entity labels is converted into the text vector, and text generation is performed through forward decoding and reverse decoding in a prediction stage of a text generation model to obtain enhanced data with annotation information.

Abstract: Certain aspects of the present disclosure provide techniques for generating a replacement sentence with the same or similar meaning but a different sentiment than an input sentence. The method generally includes receiving a request for a replacement sentence and iteratively determining a next word of the replacement sentence word-by-word based on an input sentence. Iteratively determining the next word generally includes evaluating a set of words of the input sentence using a language model configured to output candidate sentences and evaluating the candidate sentences using a sentiment model configured to output sentiment scores for the candidates sentences. Iteratively determining the next word further includes calculating convex combinations for the candidate sentences and selecting an ending word of one of the candidate sentences as the next word of the replacement sentence. The method further includes transmitting the replacement sentence in response to the request for the replacement sentence.

Abstract: Various techniques for providing perspective annotation to numerical representations are disclosed herein. For example, a method includes detecting a numerical representation in an original content and retrieving one or more perspectives from a database based on the detected numerical representation. The one or more perspectives individually include a restatement of information contained in the numerical representation. The method can also include annotating the original content with the retrieved one or more perspectives to form an annotated content.

Abstract: A computer-implemented method for data augmentation is provided according an embodiment of the present disclosure. In the method, a first feature vector for input data may be obtained based on a first model. The input data may be clustered to a plurality of clusters. For each of the clusters, a second feature vector may be obtained based on the first model. Then, a similarity between the first feature vector and the second feature vector may be estimated for each of the clusters. At least one cluster of the plurality of clusters for which the similarity is lower than a threshold may be determined. Moreover, data augmentation may be performed to the at least one cluster.

Abstract: A method and a system for detecting harmful content on a network are provided. The method comprises: receiving a URL; obtaining, from the URL, an HTML document associated therewith; converting the HTML document into a text; normalizing the text associated with the HTML document, thereby generating a plurality of tokens associated therewith; aggregating, each one of the plurality of tokens into a token vector associated with the HTML document; and applying, one or more classifiers to the token vector associated with the HTML document to determine a likelihood parameter indicative of the URL being associated with the harmful content; in response to the likelihood parameter being equal to or greater than a predetermined likelihood parameter threshold: identifying, the URL as being associated with the harmful content; and storing, the URL in a database of harmful URLs.

Abstract: A computer-implemented method is programmed to support efficient and rapid generation of machine translation suggestions on client devices. Network latency is substantially reduced or eliminated by separating certain aspects of the translation workload across multiple classes of tasks, including final neural network output, between a client device and server device. The client device and server device may be connected such that a decoder portion of a machine translation system may be downloaded onto the client device, along with an initial translation suggestion and encoder outputs associated with a document, which document is in a source language to be translated into a target language. The initial translation suggestion may be replaced by an updated machine translation suggestion as a user inputs text in the target language called a prefix. This updated machine translation is generated on the client-side decoder using the previously-downloaded encoder outputs as input and the prefix as constraint.

Abstract: Computer implemented methods and systems are provided for generating diverse key phrases while maintaining competitive output quality. A system for training a sequence to sequence (S2S) machine learning model is proposed where neural unlikelihood objective approaches are used at (1) a target token level to discourage the generation of repeating tokens, and (2) a copy token level to avoid copying repetitive tokens from the source text. K-step ahead token prediction approaches are also proposed as an additional mechanism to augment the approach to further enhance the overall diversity of key phrase outputs.

Abstract: A classification code parser and method can include: reading a classification code having a description; reading a required keyword, and a total number of keywords associated with the classification code; reading text of a note; tokenizing the text of the note to create a note token stream, the note token stream having a note token and a position of the note token within the note token stream; creating a keyword map including a total number of matched keywords; determining a match ratio from the total number of the matched keywords and the total number of the keywords; determining a proximity factor based on a shortest span of tokens within the note token stream containing all the matched keywords; and determining a strength of a match between the classification code and the note based on the match ratio being multiplied by the proximity factor.

Abstract: A method manages bot dialogue. A user input is converted to a phrase vector. A set of identified tokens are identified by a token identification engine from the phrase vector. An unsupervised token is selected from the set of identified tokens. A supervised token is selected from the set of identified tokens. A voted token selected from the unsupervised token and the supervised token. A next token is identified based on a set of recent tokens that includes the voted token. The next token is presented as one of a voice communication and an email communication.

Abstract: One example method includes, for each document in a group of annotated documents, extracting a set of words from the annotated document, and each of the words is positioned in a respective field of the annotated document. The method further includes using an aggregation function to determine, for one of the fields, a similarity of each one of the annotated documents to all of the other annotated documents, creating a document layout graph with nodes that each correspond to a respective annotated document, and each node is connected to all other nodes for which a similarity threshold for the one field has been met, and running an algorithm on the document layout graph to identify a clique of the annotated documents, and each annotated document in the clique has a similar layout to respective layouts of the other annotated documents in the clique.

Abstract: Techniques are disclosed for identifying hypocrisy in text. A computer system creates, from fragments of text, a syntactic tree that represents syntactic relationships between words in the fragments. The system identifies, in the syntactic tree, a first entity and a second entity. The system further determines that the first entity is opposite to the second entity. The system further determines a first sentiment score for a first fragment comprising the first entity and a second sentiment score for a second fragment comprising the second entity. The system, responsive to determining that the first sentiment score and the second sentiment score indicate opposite emotions, identifies the text as comprising hypocrisy and providing the text to an external device.

Abstract: The present invention may be a system for creating indexes for information retrieval comprises a processor and a memory. The memory has program instructions embodied therewith. The program instructions are executable by the processor to cause the system to read a document having hinting information into a memory, where the hinting information is associated with each unique expression in an original document. The program instructions are further executable to create the indexes from the document, where a first analysis method for generating a contiguous sequence of items from a text in the document is used for creating the indexes for each sequence in the unique expression with which the hinting information is associated and a second analysis method for dividing the text into meaningful units is used for creating the indexes for each word in the text other than the unique expression.

Abstract: Identification of content gaps based on relative user-selection rates between multiple discrete content sources. A system analyzes search log activity to determine whether users that are conducting particular types of search activities are ultimately selecting and relying upon content resources from a predefined content source of interest or, alternatively, whether such users are unsatisfied with the predefined content source of interest and are instead relying upon other third-party content sources. This particular type of analysis provides valuable insights into whether content gaps exist within the predefined content source of interest.

Abstract: A system and process for tagging electronic documents or other electronic content with concepts mentioned, contained, or otherwise described in that content. Once tagged, the content may be searchable, indexable, and retrievable in order to provide that content to an end user or another recipient. The system may be configured to handle a considerable number of asset files and a large number of users, workflows, and access applications simultaneously. The system may auto-tag the content and also may include a user interface for confirming and updating those tags and for manually creating new or additional tags. Content may include documents such as medical documents relating to procedures, diagnoses, medications or other domains. Alternatively, the content may include information about various care providers, in order to allow a user to locate a physician meeting one or more desired criteria.

Abstract: A process receives, with a processor, audio corresponding to media content. Further, the process converts, with the processor, the audio to text. In addition, the process concatenates, with the processor, the text with one or more time codes. The process also parses, with the processor, the concatenated text into one or more text chunks according to one or more subtitle parameters. Further, the process automatically translates, with the processor, the parsed text from a first spoken language to a second spoken language. Moreover, the process determines, with the processor, if the language translation complies with the one or more subtitle parameters. Additionally, the process outputs, with the processor, the language translation to a display device for display of the one or more text chunks as one or more subtitles at one or more times corresponding to the one or more time codes.

Abstract: There is a need for more effective and efficient natural language processing (NLP) solutions. This need can be addressed by, for example, solutions for performing NLP-based document prioritization by utilizing joint sentiment-topic (JST) modeling.

Abstract: Techniques for interpreting a text classifier model are described. An exemplary method includes receiving a request to interpret the text classifier; receiving input text to be used to interpret the text classifier; interpreting the text classifier using the input text and masked input text to determine two or more of a counterfactual score for the received input text or an aspect thereof, an importance score for the received input text or an aspect thereof, and a bias score for the received input text or an aspect thereof as requested by the request, and providing the determined one or more scores is provided to a requester.

Abstract: Provided is at least one processor, and the processor is configured to analyze an image to derive property information indicating a property of a structure of interest included in the image, generate a sentence related to the image based on the property information, analyze the sentence to specify a term representing the property related to the structure of interest included in the sentence, and collate the property information with the term.

Abstract: Systems and methods of the present disclosure enable context-aware haptic error notifications. The systems and methods include a processor to receive input segments into a software application from a character input component and determine a destination. A context identification model predicts a context classification of the input segments based at least in part on the software application and the destination. Potential errors are determined in the input segments based on the context classification. An error characterization machine learning model determines an error type classification and an error severity score associated with each potential error and a haptic feedback pattern is determined for each potential error based on the error type classification and the error severity score of each potential error of the one or more potential errors. And a haptic event latency is determined based on the error type classification and the error severity score of each potential error.

Abstract: The present invention provides text-based news significance evaluation methods, apparatuses, and electronic devices for improving efficiency and accuracy of news significance evaluation, and implementing real-time dynamic evaluation on text news. The method comprises: reading text news; preprocessing the text news to obtain original data; extracting feature values from the original data, which comprises metadata, a keyword, and a probability model feature value; and obtaining a score of each feature value according to a weight ratio corresponding to each feature value. The apparatus comprises: a text news reading module, a text news preprocessing module, a feature value extraction module, a feature value weight determining module, and a text news significance evaluation module. The electronic device comprises a memory and a processor. The memory stores a computer program that can run on the processor.

Abstract: The present invention may be a method, a computer system, and a computer program product for suggesting an output candidate. The method comprises receiving a user input; selecting a corpus containing an expression similar to the user input among a plurality of corpuses; finding, in the user input, a seed word that may be present in a definition statement of an entry in a dictionary; identifying, in the dictionary, an entry of a definition statement containing the seed word or within a threshold similarity to the seed word with reference to the selected corpus; and suggesting the identified entry as an output candidate.

Abstract: Systems and methods are provided for providing a navigation interface to access or otherwise use electronic content items. In one embodiment, an augmentation application identifies at least one entity referenced in a document. The entity can be referenced in at least two portions of the document by at least two different words or phrases. The augmentation application associates the at least one entity with at least one multimedia asset. The augmentation application generates a layout including at least some content of the document referencing the at least one entity and the at least one multimedia asset associated with the at least one entity. The augmentation application renders the layout for display.

Abstract: Various aspects of this disclosure provide digital data processing systems for using encrypted variant data objects to facilitate queries of sensitive data. In one example, a digital data processing system can receive sensitive data about an entity. The digital data processing system can create, in an identity data repository and from the sensitive data, a searchable secure entity data object for the entity. The searchable secure entity data object is usable for servicing a query regarding the entity. For instance, a transformed query parameter can be generated from a query parameter in the query. The query can be serviced by matching the transformed query parameter to tokenized variant data in the searchable secure entity data object and retrieving tokenized sensitive data from the searchable secure entity data object.

Abstract: NLP techniques are disclosed that apply computer technology to sentence data for performing entity referencing. For example, a processor can parse sentence data in a defined window of sentence data into a list of entity terms and a plurality of classifications associated with the listed entity terms. A processor can also a plurality of context saliency scores for a plurality of the listed entity terms based on the classifications associated with the listed entity terms as well as maintain a list of referring terms corresponding to the listed entity terms. For new sentence data that includes a referring term from the referring term list, a processor can (i) select a corresponding entity term on the entity term list based on the context saliency scores for the entity terms, and (ii) infer that the referring term in the new sentence data refers to the selected corresponding entity term.

Abstract: A hierarchical embedding model is used to obtain respective language-agnostic embeddings of entity records of a cross-language data set. A plurality of record representation pairs is prepared based at least in part on the language-agnostic embeddings. A machine learning model is trained using the record representations pairs to generate similarity scores for pairs of entity records whose text attributes are expressed in different languages.

Abstract: Speech processing techniques are disclosed that enable determining a text representation of named entities in captured audio data. Various implementations include determining the location of a carrier phrase in a word lattice representation of the captured audio data, where the carrier phrase provides an indication of a named entity. Additional or alternative implementations include matching a candidate named entity with the portion of the word lattice, and augmenting the word lattice with the matched candidate named entity.

Abstract: Systems, methods, and other techniques for extracting data from obituaries are provided. In some embodiments, an obituary containing a plurality of words is received. Using a machine learning model, an entity tag from a set of entity tags may be assigned to each of one or more words of the plurality of words. Each particular tag from the set of entity tags may include a relationship component and a category component. The relationship component may indicate a relationship between a particular word and the deceased individual. The category component may indicate a categorization of the particular word to a particular category from a set of categories. The extracted data may be stored in a genealogical database.