Abstract: Techniques to apply machine learning to schedule events of interest for a device user. As described herein, a typical device user is inundated with information suggesting activities and other things to do. Using these techniques, the information is categorized according to the activity and those activities that are most likely to be engaged in by the device user are recommended to the device user (via their device) as events of interest. If the device user selects an event of interest, the device is updated to reflect that selection. For instance, a calendar application may be updated to include an event description at the event's date and time. Other embodiments are described and claimed.

Abstract: In non-limiting examples of the present disclosure, systems, methods and devices for determining a language of a text string are presented. A language detection model may be maintained. The language detection model may comprise identities and weights for initial and final consonants, identities and weights for prefixes and suffixes, and identities and weights for vowel sequences, where each identity is derived from a training corpus. The weights may correspond to a frequency of a text unit in the corpus. A text string may be received and a match score between the text string and the language of the language detection model may be determined. The match score may be based on initial and final consonant scores, prefix and suffix scores, and/or vowel sequence scores for each word in the text string. If the match score meets a threshold value a follow-up action associated with the language may be performed.

Abstract: Systems and methods to briefly deviate from and resume back to amending a section of a note are disclosed. Exemplary implementations may: obtain audio information representing sound captured by an audio section of a client computing platform, such sound including speech from a user associated with the client computing platform; effectuate presentation of a graphical user interface that includes sections of the note; analyze the audio information to determine which individual ones of the spoken inputs are the primary spoken input or the deviant spoken input; determine, based on analysis, which section of the note to which the deviant spoken input is related; alternately amend, based on the determination, sections of the note by deviating from one section to another section and returning back to the one section for continued population; and effectuate, via the user interface, presentation of the alternating amendments to the sections of the note.

Abstract: One embodiment provides a method, including: obtaining at least two documents, wherein one of the at least two documents comprises a revision different than another of the at least two documents; identifying, within each of the at least two documents, portions corresponding to groups of text containing a conceptual unit; assigning at least a subset of the identified portions to a category type corresponding to a topic of a given portion, wherein the assigning comprises (i) generating a semantic tag for the identified portions in the subset and (ii) tagging the identified portions in the subset with the semantic tag; and determining changes between the at least two documents, wherein the determining comprises (iii) aligning given portions across the at least two documents based upon a relationship between the given portions across the at least two documents, (iv) identifying semantic differences between the aligned portions, and (v) identifying any remaining unaligned portions.

Abstract: A method for extracting an event from a text including obtaining an input text; inputting the input text into a model for extracting trigger words to obtain a trigger word extraction result of the input text; inputting the input text and the trigger word extraction result into a model for extracting arguments to obtain an argument extraction result of the input text; and obtaining an event extraction result of the input text according to the trigger word extraction result and the argument extraction result.

Abstract: A system, computer program product, and method are provided for use with an intelligent computer platform to identify intent and convert the intent to one or more physical actions. The aspect of converting intent includes receiving content, identifying potential variants, and statistically analyzing the variants with a confidence assessment. The variants are sorted based on a protocol associated with the confidence assessment. A variant from the sort is selected and applied to a physical device, which performs a physical action and an associated hardware transformation based on the variant.

Abstract: With the proliferation of data and documents available on the internet and other information sources, analysis of adverse events poses a serious technical challenge on account of associated data volume and variety. This disclosure relates generally to identification and profiling of adverse events. By receiving a set of articles from a plurality of data sources and utilizing a series of Natural Language Processors, NLP techniques are employed to identify implicit and explicit adverse events. Entity statistics and sentiment extraction and analysis is performed. An ontology based adverse event identification framework is proposed for identification and profiling of implicit adverse event. An attention based bi-directional long short term memory network for adverse event identification and classification is proposed.

Abstract: A method starts with a processor receiving, at a computing system from a client device, descriptive text data from a descriptive text interface displayed on the client device, the descriptive text data associated with a media content item displayed on the client device, analyzing the descriptive text data to identify at least one data characteristic within the descriptive text data, and accessing a plurality of personalized media overlay icons each comprising at least one media overlay icon characteristic. The processor determines whether the identified data characteristic is associated with any of the at least one media overlay icon characteristics of each of the plurality of personalized media overlay icons, generates a personalized overlay icon interface which includes a selection of the plurality of personalized media overlay icons that include at least one media content icon characteristic associated with the data characteristic.

Abstract: This disclosure relates to methods, non-transitory computer readable media, and systems apply machine-learning techniques and computational sentiment analysis to summarize sentences extracted from a group of textual responses or to select representative-textual responses from the group of textual responses. By using a response-extraction-neural network to extract (and sometimes paraphrase) sentences from textual responses, the disclosed methods, non-transitory computer readable media, and systems can generate a response summary of textual responses based on sentiment indicators corresponding to the textual responses. By applying a machine-learning classifier to generate textual quality scores for textual responses, the disclosed methods, non-transitory computer readable media, and systems select representative-textual responses from a group of textual responses based on relevancy parameters and sentiment indicators corresponding to the textual responses.

Abstract: An apparatus comprises at least one processing device configured to receive a query to generate a summary of a document, to perform two or more iterations of filtering the document to produce a current version of the summary of the document, wherein each of the iterations comprises determining similarity between a first vector representation of the current version of the summary and second vector representations of respective ones of two or more portions of the unstructured text data of the document not yet added to the current version of the summary. The processing device is also configured to generate, following identification of one or more designated stopping criteria in a given iteration, a final version of the summary based at least in part on the current version of the summary produced in the given iteration, and to provide a response to the query comprising the final version of the summary.

Abstract: An electronic apparatus, and a method for the electronic apparatus to control to execute a function according to a voice command uttered by a user are provided. The electronic apparatus includes a memory including a plurality of databases divided according to predicates, a communication circuit that receives voice data corresponding to a voice command from an external apparatus, and a processor operatively connected with the memory and the communication circuit. The processor is configured to identify the type of a predicate included in the voice data, and based on information included in a database corresponding to the identified predicate, determine at least one apparatus to execute the voice command and a function that the at least one apparatus will perform according to the voice command, and control the at least one apparatus to execute the function according to the voice command.

Abstract: A method of enhancing an automated speech recognition confidence classifier includes receiving a set of baseline confidence features from one or more decoded words, deriving word embedding confidence features from the baseline confidence features, joining the baseline confidence features with word embedding confidence features to create a feature vector, and executing the confidence classifier to generate a confidence score, wherein the confidence classifier is trained with a set of training examples having labeled features corresponding to the feature vector.

Abstract: Techniques to enable a chat bot system to use metaphors during an interaction with a user are provided. Identification information for the user can be received and can be used to determine a customer group assigned to the user. An initial state of a sequence for resolving an issue of the user can be determined, with the sequence for resolving the issue of the user including a final state wherein the issue of the user is resolved. A metaphor can be selected to include in a question based on the determined customer group assigned to the customer and the determined initial state of the sequence for resolving the issue of the customer. The question can request information from the user relating to the issue of the user. The question can be generated to include the selected metaphor and then provided to the user with a prompt to provide a response.

Abstract: Systems and methods for sentence fusion are described. Embodiments receive coreference information for a first sentence and a second sentence, wherein the coreference information identifies entities associated with both a term of the first sentence and a term of the second sentence, apply an entity constraint to an attention head of a sentence fusion network, wherein the entity constraint limits attention weights of the attention head to terms that correspond to a same entity of the coreference information, and predict a fused sentence using the sentence fusion network based on the entity constraint, wherein the fused sentence combines information from the first sentence and the second sentence.

Abstract: A configurable user interface architecture to be used in high integrity safety critical avionic systems and is adaptable to multiple cockpit displays has a protocol-agnostic layer that handles entry and display update behaviors, and a protocol-specific layer that connects the entry and display behaviors to protocol-specific input/output messages using configuration tables. The configuration tables are built using a specially designed configuration tool and the configuration table built by the tool can be embedded in the target user interface software or loaded externally into a pre-compiled application.

Abstract: The present disclosure relates to frameworks and methodologies for enabling categorisation and/or searching of digitised information, including clinical report data. Embodiments of the invention have been particularly developed to assist categorisation of digitised information, such as clinical report data, in a streamlined manner based on a pre existing set of classification codes. This, in some embodiments, enables the discovery and extraction of meaningful patterns from unstructured clinical reports. Further embodiments of the invention have been particularly developed to assist in the discovery and extraction of meaningful patterns from an unstructured set of digitised information such as unstructured clinical reports. While some embodiments will be described herein with particular reference to those applications, it will be appreciated that the invention is not limited to such a field of use, and is applicable in broader contexts.

Abstract: Disclosed embodiments relate to natural language processing. Techniques can include receiving input text, extracting, from the input text, at least one modifier and aspect pair, receiving data from a knowledgebase, based on the at least one modifier and aspect pair and commonsense data, generate one or more premise embeddings, convert the input text into tokens, generating at least one vector for one or more of the tokens based on an analysis of the tokens, combine the at least one vector with the one or more premise embeddings to create at least one combined vector, and analyze the at least one combined vector wherein the analysis generates an output indicative of a feature of the input text.

Abstract: Examples of artificial intelligence-based reasoning explanation are described. In an example implementation, a knowledge model having a plurality of ontologies and a plurality of inferencing rules is generated. Once the knowledge model is generated, based on a real-world problem, a knowledge model from amongst various knowledge models is selected to be used for resolving a real-world problem. The data procured from the real-world problem is clustered and classified into an ontology of the determined knowledge model. Inferencing rules to be used for deconstructing the real-world problem are identified, and a machine reasoning is generated to provide a hypothesis for the problem and an explanation to accompany the hypothesis.

Abstract: Embodiments of the present invention include a computer program product, a computer system, and a computer-implemented method for generating a data model that machine learns from social media postings to determine whether social media postings will gain traction.

Abstract: As described, an artificial intelligence (AI) design application exposes various tools to a user for generating, analyzing, evaluating, and describing neural networks. The AI design application includes a network generator that generates and/or updates program code that defines a neural network based on user interactions with a graphical depiction of the network architecture. The AI design application also includes a network analyzer that analyzes the behavior of the neural network at the layer level, neuron level, and weight level in response to test inputs. The AI design application further includes a network evaluator that performs a comprehensive evaluation of the neural network across a range of sample of training data. Finally, the AI design application includes a network descriptor that articulates the behavior of the neural network in natural language and constrains that behavior according to a set of rules.

Abstract: A transcript of a conference (e.g., a video conference, an audio conference, or a telephone call with two or more participants) is processed to extract a conference summary. A graphical user interface is presented that enables a user to edit highlighting to select portions of a conference recording for inclusion in the conference summary. The user can edit highlighting on strings of a transcript and/or on a playback slide bar and the highlighting edits are propagated to both of these interfaces. Video clips selected for inclusion in a summary based on timestamps of strings in a transcript may be extended forward in time to incorporate a nearest I-frame of video. The short video or audio summary may be presented to users along with a text summary of the conference to enable efficient storage and transmission of information from the conference within a unified communications system.

Abstract: A personal electronic device is adapted to construct data tables from user input received over time relating to translations of translatable items from a first language to a second language. Entries in a data table for a user are dynamic and may indicate likelihoods of the user correctly translating translatable items as a function of time. Each translatable item may have a different time-dependent likelihood of a correct translation. Operation of the personal electronic device for a user may be based in part on the acquired, time-dependent likelihoods for that user, so that information may be presented to the user in a more efficient manner.

Abstract: Systems and methods to improve the optical character recognition of records, and in particular health records, are provided. An image of a medical record is received, and an initial optical image recognition (OCR) on the image is performed to identify text information. The OCR signal quality may be measured, and areas of insufficient OCR signal quality may be isolated. The signal quality is determined by a weighted average of semantic analysis of the resulting text, and/or OCR accuracy measures. The OCR process may be repeated on the isolated regions of lower signal quality, each time using a different OCR transform, until all regions are completed with a desired degree of signal quality (accuracy). All the regions of the document may then be recompiled into a single document for outputting.

Abstract: A method of operating a health tracking system is disclosed. The method comprises: receiving a first data record comprising at least a first descriptive string regarding a consumable item, the first descriptive string having at least one word thereof incorrectly spelled; generating a vector using the first descriptive string using a machine learning model; identifying a second descriptive string which corresponds to the consumable item and which has a correct spelling of the at least one incorrectly spelled word by applying the machine learning model to the generated vector; calculating a confidence factor regarding the identified second descriptive string using the machine learning model; and when it is determined that the confidence factor exceeds a predetermined threshold, (i) modifying the first data record by replacing the first descriptive string with the second descriptive string, and (ii) storing the modified first data record in the database.

Abstract: Various embodiments relate generally to data science and data analysis, computer software and systems, including a subset of intermediary executable instructions constituting an communication interface between various software and/or hardware platforms, and, more specifically, to an automated application interface governance platform to automate development, maintenance, and governance functions for application interfaces, such as harmonizing, validating, and/or replicating application program interfaces (“APIs”). For example, a method may include identifying a subset of application interfaces, synthesizing a data structure for each application interface, analyzing the data structure against other data structures to identify duplicative portions among multiple data structures, substituting a reference to a location into a portion of multiple application interfaces.

Abstract: Systems and processes for operating an intelligent automated assistant are provided. For example, a first message is received, wherein a semantic representation of the first message is obtained. In accordance with a determination that the semantic representation of the first message is within a predefined distance of a stored semantic representation, one or more stored reply message representations within the predefined distance of the semantic representation of the first message are identified. The one or more stored reply message representations may correspond to one or more messages previously sent, by a user of the electronic device, in response to a previously received second message. The second message may be semantically similar to the first message. One or more reply messages are provided based on the one or more identified reply message representations, wherein the one or more reply message suggestions are provided as suggested responses to the first message.

Abstract: According to one embodiment, a first phrase having one or more keywords is received to be added to a phrase blacklist, where the first phrase has been determined to be related to unwanted content. An analysis is performed on the keywords of the first phrase to identifying a set of one or more related phrases that are related to the first phrase. The first phrase and the set of related phrases are added in the phrase blacklist. The phrase blacklist contains a plurality of phrases that have been determined to be related to unwanted content. The phrase blacklist is utilized to prevent a content item from appearing in a search result, in response to a search query having a phrase matching at least one of the phrases in the phrase backlist.

Abstract: Provided is a system and method for adapting sentiment analysis to user profiles to reduce bias in customer or user generated content, specifically a system and method that discounts or adjusts bias in sentiment data based on the channel from which the content was received and/or the demographic of the user. The system includes a means to detect sentiment bias for any product, service, or company across multiple channels of customer data; a means to construct models to quantize bias by specific demographics and channels; and a means to adjust sentiment model output to reduce inflation by biased groups.

Abstract: A conversational analytics system may provide for a conversational interface to any relational database. A controlled natural language may be constructed in an automated manner from a given database (e.g., from schema and values associated with a relational database). For instance, a user natural language expression may be converted to an expression in the constructed controlled natural language and the controlled natural language expression may be converted into a sequence of one or more queries in a query language (e.g., queries in structured query language (SQL)). Such an intermediate controlled natural language may provide queries without ambiguity (e.g., as each expression or phrase in the controlled natural language may be mapped to one sequence of SQL queries). Accordingly, any natural language user utterance that ultimately follows the controlled natural language may be automatically converted into a sequence of one or more SQL queries sent to the database.

Abstract: Novel distractorless authorship verification technology optionally combines with novel algorithms to solve authorship attribution as to an open set of candidates—such as without limitation by analyzing the voting of “mixture of experts” and outputting the result to a user using the following: if z (z=pi?pj? pi+pj?(pi?pj)2/n) is larger than a first predetermined threshold then author j cannot be the correct author; or if z (z=pi?pj? pi+pj?(pi?pj)2/n) is smaller than a second predetermined threshold then author i cannot be the correct author; or if no author garners significantly more votes than all other contenders then none of the named authors is the author of a document in question—in a number of novel applications. Personality profiling and authorship attribution may also be used to verify user identity to a computer.

Abstract: A computerized method for reducing domain noise, creating and summarizing human-written sentences into clusters for efficient tagging in natural language processing comprising: receiving a typed, handwritten or printed text; implementing an optical character recognition (OCR) process on human written text to generate a digital version of the human written text; splitting the digital version of the typed, handwritten or printed text into an array of sentences, using a sentence splitter to generate a split sentence version; determining a domain of the human written text; based on the domain, implementing a domain noise reduction process on the split sentences version; hierarchically clustering the split sentences version after the domain noise reduction process; and summarizing the clustered sentences and reducing the amount of data to be tagged.

Abstract: A user engages in spoken voice or messaging platform-based natural language dialogues with an automated bot. The bot identifies actions and subjects of the actions from the dialogues and processes Application Programming Interfaces (API) to perform travel-related queries or process travel-related transactions with travel services. Real-time responses to queries or completed transaction notifications are provided back to the user by the bot during the dialogues.

Abstract: A tunable monitoring system for detecting computer threats to third-parties is provided. The system may include a receiver operable to receive entity characteristics relating to a first entity and identification data identifying a third-party. The system may include a publicly-available data crawler operable to mine and retrieve publicly-available data. The data may relate to both the third-party and one or more of a plurality of entities. The plurality of entities may include entities that share a predetermined threshold number of entity characteristics with the first entity. The system may include a sentiment analysis module operable to score the retrieved data using a sentiment analysis scoring model. A potential relationship between the first entity and third-party may be dissolved when the score is above a predetermined negativity threshold. The potential relationship between the first entity and the third-party may be initiated when the score is below the predetermined negativity threshold.

Abstract: This disclosure describes techniques for identifying a voice-enabled device from a group of voice-enabled devices to respond to a speech utterance of a user. A speech-processing system may receive an audio signal representing the speech utterance captured in an environment of a voice-enabled device, and identify another voice-enabled device located in the environment. The system may analyze the audio signal using a different natural-language-understanding model for each of the voice-enabled devices to identify an intent for each of the voice-enabled devices to respond to the speech utterance. The system may determine confidence scores that the intents are responsive to the speech utterance, and select the intent with the highest confidence score. The system may use the selected intent to generate a command for the corresponding voice-enabled device to respond to the user.

Abstract: Embodiments described in this disclosure illustrate the use of self-/semi supervised approaches for label-efficient DST in task-oriented dialogue systems. Conversational behavior is modeled by next response generation and turn utterance generation tasks. Prediction consistency is strengthened by augmenting data with stochastic word dropout and label guessing. Experimental results show that by exploiting self-supervision the joint goal accuracy can be boosted with limited labeled data.

Abstract: An intelligent text cleaning method includes: acquiring a text set, and preprocessing the text set to obtain a word vector text set; subjecting the word vector text set to a full-text matrix numeralization to generate a principal word vector matrix and a text word vector matrix; inputting the principal word vector matrix to a BiLSTM model to generate an intermediate text vector; inputting the text word vector matrix to a convolution neural network model to generate a target text vector; and concatenating the intermediate text vector and the target text vector to obtain combined text vectors, inputting the combined text vectors to a pre-constructed semantic recognition classifier model, outputting an aggregated text vector, subjecting the aggregated text vector to reverse recovery using a word2vec reverse algorithm, and outputting a standard text. The present application realizes accurate text cleaning.

Abstract: Embodiments are directed to organizing conversation information. Two or more machine learning (ML) models and a plurality of sentences provided from a conversation may be employed to generate insight scores for each sentence such that each insight score correlates to a probability that its sentence includes one or more of an action or a question. In response to one or more sentences having insight scores that exceed a threshold value an information score and a definiteness score may be determined for the one or more sentences. And one or more insights associated with the conversation may be generated based on the one or more sentences. A report may be generated that associates the one or more insights with one or more portions of the conversation that include the one or more sentences that are associated with the insights.

Abstract: Systems, methods, and storage media for extracting patent document templates from a patent corpus are disclosed. Exemplary implementations may: obtain a patent corpus; receive one or more parameters; determine one or more subsets of the patent corpus by filtering the patent corpus based on the one or more parameters; identify one or more document clusters within individual ones of the one or more subsets of the patent corpus; obtain a patent document template corresponding to the first document cluster; and/or perform other operations.

Abstract: A system for generating textual instructions for manufacturers from hybrid textual and image data includes a manufacturing instruction generator that may generate a language processing module from a first training set including at least a training annotated file describing at least a first product to manufacture, the at least an annotated file containing one or more textual data, and at least an instruction set containing one or more manufacturing instructions to manufacture the at least a first product. Manufacturing instruction generator may use the language processing to generate textual instructions for manufacturers from at least an annotated file and may initiate manufacture using the generated manufacturing instructions.

Abstract: A method for execution by a computing device includes obtaining a phrase that includes string of words and generating a valid sequence of words utilizing the phrase. The method further includes identifying a set of identigens for each word of the valid sequence of words to produce sets of identigens. The method further includes identifying, for each identigen of the sets of identigens, a word type associated with phrase structure grammar rules to produce sets of identigen-type associations. The method further includes interpreting, utilizing the phrase structure grammar rules, the sets of identigen-type associations to produce an entigen group. The entigen group represents a most likely interpretation of the phrase.

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: Systems and methods of processing a query from a user. A method includes receiving, by a server computer, an initial question from a client computer. The initial question includes a plurality of words and the server computer can identify a set of words in the plurality of words. Then the server computer can determine a list of clarifying questions based on a subset of the set of words. The server computer presents the list of clarifying questions to the client computer and receiving clarifying answers to the clarifying questions. The server computer determines an answer to the initial question and presents the answer to the client computer.

Abstract: Provided are a computer program product, system, and method for pre-processing a table in a document for natural language processing. A table in a document is parsed to extract column headers, row headers, and data cells, which are processed to determine an initial set of a main element comprising an entity whose value is to be extracted, a conditional element that refines the entity, and a value element comprising a value for the entity. A user selection is received of at least one of the column headers, row headers, and data cells for at least one of the main element, conditional element, and the value element in the initial set to produce a modified set of the main element, conditional element, and value element. The modified set is provided to a natural language processing engine to perform natural language processing of the document including the table, using the modified set.

Abstract: Text is received from a first user. The text is associated with an electronic communication tool for communication to a second user. Candidate answers are generated based on the text using a question answering system. At least one of the candidate answers is provided to the first user.

Abstract: A system for reinforcement learning based controlled natural language generation is disclosed. The system includes a token generator subsystem to generate an initial output phrase including a sequence of output tokens. The system includes trained models associated with corresponding predefined tasks. Each trained model includes an attention layer to compute attention-based weights for each output token. The trained models include a scoring layer to generate a phrase sequence level score for the output phrase. The trained models include a reward generation layer to generate dense rewards for each output token based on the attention-based weights and the phrase sequence level score. The trained models include a feedback score generation layer to generate a feedback score based on the dense rewards and reward weights assigned to the dense rewards of the corresponding trained models. The feedback score generation layer provides the feedback score iteratively to the token generator subsystem.

Abstract: A method is disclosed. The method includes receiving a text phrase from a user and parsing the text phrase using a natural language parser to generate distinct words. A vector of values relating to the distinct words is generated and compared to vectors in a graph database which relate to learned communities of language. Based on the comparing, a most similar community in the learned communities of language is determined and an action set action set associated with the most similar community is queried. A response to the text phrase from the user is then generated based on the action set and provided to the user.

Abstract: Systems and methods for conversational user experiences and conversational database analysis disclosed herein improve the efficiency and accessibility of low-latency database analytics.

Abstract: Briefly stated, the invention is directed to retrieving a semantically matched knowledge structure. A question and answer pair is received, wherein the answer is received from a query of a search engine. A question is constraint-matched with the answer based on maximizing a plurality of constraints, wherein at least one of the plurality of the constraints is a similarity score between question and answer, wherein the constraint matching generates a matched sequence. For one or more answer sequences, a subsequence is found that are not parsed as answer slots. Query results are obtained from another search engine based on a combination of the answer or question, and the non-answer subsequence. And a KB based is refined on the query results and the constraint matching and based on a neural network training, for a further subsequent semantic matching, wherein the KB includes a dense semantic vector indication of concepts.

Abstract: The technology disclosed provides a so-called “pointer sentinel mixture architecture” for neural network sequence models that has the ability to either reproduce a token from a recent context or produce a token from a predefined vocabulary. In one implementation, a pointer sentinel-LSTM architecture achieves state of the art language modeling performance of 70.9 perplexity on the Penn Treebank dataset, while using far fewer parameters than a standard softmax LSTM.

Abstract: Due to the high language use variability in real-life, manual construction of semantic resources to cover all synonyms is prohibitively expensive and may result in limited coverage. Described herein are systems and methods that automate the process of synonymy resource development, including both formal entities and noisy descriptions from end-users. Embodiments of a multi-task model with hierarchical task relationship are presented that learn more representative entity/term embeddings and apply them to synonym prediction. In model embodiments, a skip-gram word embedding model is extended by introducing an auxiliary task “neighboring word/term semantic type prediction” and hierarchically organize them based on the task complexity. In one or more embodiments, existing term-term synonymous knowledge is integrated into the word embedding learning framework.