Abstract: Conventionally, text summarization has been rule-based method and neural network based which required large dataset for training and the summary delivered had to be assessed by user in terms of relevancy. System and method are provided by present disclosure that generate causal insight summaries wherein event of importance is detected, and it is determined why event is relevant to a user. Text description is processed for named entities recognition, polarities of sentences identified, extraction of causal effects sentences (CES) and causal relationship identification in text segments which correspond to impacting events. Named entities are then role labeled. A score is computed for named entities, polarities of sentences, causal effects sentences, causal relationships, and the impacting events. A causal insight summary is generated with overall polarity being computed/determined.

Abstract: Methods, systems, and apparatus, including computer programs encoded on computer storage media, for recognizing speech using a spiking neural network acoustic model implemented on a neuromorphic processor are described. In one aspect, a method includes receiving, a trained acoustic model implemented as a spiking neural network (SNN) on a neuromorphic processor of a client device, a set of feature coefficients that represent acoustic energy of input audio received from a microphone communicably coupled to the client device. The acoustic model is trained to predict speech sounds based on input feature coefficients. The acoustic model generates output data indicating predicted speech sounds corresponding to the set of feature coefficients that represent the input audio received from the microphone. The neuromorphic processor updates one or more parameters of the acoustic model using one or more learning rules and the predicted speech sounds of the output data.

Abstract: The deep semantic feature based few-shot intent recognition method for air traffic control instructions, belonging to the technical field of air traffic control; the method of the present invention solves technical problems in the prior art relating to incomplete and unreliable security situation awareness and semantic representation due to poor intent recognition ability for instructions with a small sample size, as well as difficulty in improving the applicability of models while keeping a small amount of labeled language information; the present invention uses the one obtained air traffic control instruction set to construct a domain language model and adopts an unsupervised learning method to enhance the mining and representation of deep-level features of air traffic control air-ground communication, thereby further improving the reliability of safety situation awareness and semantic representation.

Abstract: At least one processor may obtain a document comprising text tokens. The at least one processor may determine, based on a pre-trained language model, word embeddings corresponding to the text tokens. The at least one processor may determine, based on the word embeddings, named entities corresponding to the text tokens; and one or more accuracy predictions corresponding to the named entities. The at least one processor may compare the one or more accuracy predictions with at least one threshold. The at least one processor may associate, based on the comparing, the named entities with one or more confidence levels. The at last one processor may deliver the named entities and the one or more confidence levels.

Abstract: Described are methods and systems are for generating dynamic conversational queries. For example, as opposed to being a simply reactive system, the methods and systems herein provide means for actively determining a user's intent and generating a dynamic query based on the determined user intent. Moreover, these methods and systems generate these queries in a conversational environment.

Abstract: The disclosed technology teaches a method for interactively generating a graphical summary of a meeting that provides a reflection of a group's conversation with textual and graphical elements in a tapestry. The method comprises a combination of transcript analysis using natural language processing or other machine learning models and automatic selection of graphic symbols detected as relevant to the transcript by statistical analyses. The method further includes the ability to interact with a generated tapestry for additional customization, wherein a user's customization selection is fed back into the machine learning models to fine tune the result.

Abstract: A method and a device for acquiring character information in a picture, a non-transitory storage medium, a page processing method, and a knowledge graph construction method are disclosed. The method for acquiring character information in a picture includes: acquiring a picture and extracting at least one piece of character information in the picture; and checking-and-correcting the at least one piece of character information based on a knowledge graph.

Abstract: Techniques and systems are described for performing semantic text searches. A semantic text-searching solution uses a machine learning system (such as a deep learning system) to determine associations between the semantic meanings of words. These associations are not limited by the spelling, syntax, grammar, or even definition of words. Instead, the associations can be based on the context in which characters, words, and/or phrases are used in relation to one another. In response to detecting a request to locate text within an electronic document associated with a keyword, the semantic text-searching solution can return strings within the document that have matching and/or related semantic meanings or contexts, in addition to exact matches (e.g., string matches) within the document. The semantic text-searching solution can then output an indication of the matching strings.

Abstract: The invention relates to a method and a system for improving performance of text summarization and has an object of improving performance of a technique for generating a summary from a given paragraph. According to the invention to achieve the object, a method for improving performance of text summarization includes: an a step of generating an embedding vector by vectorizing a natural language-based context; a b step of generating a graph using the embedding vector and calculating a first likelihood of each of at least one node included in the graph; a c step of generating a second likelihood by assigning a weight to the first likelihood according to a result of comparing at least one node included in the graph with the context; and a d step of calculating a third likelihood for all candidate paths present in the graph based on the second likelihood, selecting a path having a highest third likelihood, and generating a summary based on the path.

Abstract: A method is described, for processing images of persons or objects to generate an identity preservative feature descriptor learnt for each person or object. The method includes obtaining an image of a person or object, extracting at least one spatial attribute of the person or object from the obtained image, and extracting at least one appearance feature of the person or object from the image by using a mapping function to translate image pixels into appearance attributes represented by at least one numerical feature. The method also includes combining the at least one spatial attribute and the at least one appearance feature to generate the unique feature descriptor representing the person or the object, to assign the unique feature descriptor to the image to enable feature descriptors representing the same person or object to be compared to feature descriptors representing different people or objects given a predefined mathematical pseudo-distance metric according to a least a distance from each other.

Abstract: An example apparatus for recognizing speech includes an audio receiver to receive a stream of audio. The apparatus also includes a key phrase detector to detect a key phrase in the stream of audio. The apparatus further includes a model adapter to dynamically adapt a model based on the detected key phrase. The apparatus also includes a query recognizer to detect a voice query following the key phrase in a stream of audio via the adapted model.

Abstract: The invention provides a system and method for analyzing similarity of natural language data. The system comprises a neural network subsystem adapted for reading graph format input data comprising a plurality of nodes having node values, and a similarity estimation subsystem utilizing the neural network subsystem and being trained for estimating similarity of a first and a second graphs, the similarity estimation subsystem being capable of producing at least one similarity value. In addition, there is provided a similarity explainability subsystem adapted to calculate importance values for a plurality of nodes or subgraphs of the second graph, which are used to create a reduced second graph indicate sub-blocks of the second block of natural language.

Abstract: Systems and methods for conducting communications between a user and an Artificial Intelligence (AI) character model are provided. An example method includes determining a context of a dialog between the AI character model and the user, the context being determined based on a data stream received from a client-side computing device associated with the user; receiving a message of the user in the dialog; and generating, based on the context and the message, an input to a language model configured to predict a response to the message; providing the input to the language model to obtain the response; and transmitting the response to the client-side computing device, where the client-side computing device presents the response to the user. The input to the language model includes the message expanded by a keyword associated with the context. The context includes an intent of the user and an emotional state of the user.

Abstract: Various embodiments for a chatbot improvement system are described herein. An embodiment operates by receiving input from a user via a chatbot interface. A first vertical corresponding to the user is identified. The input from the user is interpreted based on a first set of keywords corresponding to the first vertical. A first confidence score is calculated for the first vertical based on the interpretation of the input using the first set of keywords. It is determined whether the first confidence score exceeds a threshold. If the threshold is exceeded, a response to the input is generated based on the first set of keywords. If the threshold is not exceeded, the response to the input based on the second set of keywords. The generated response is provided for display via the chatbot interface.

Abstract: A method of training a neural network as a natural language processing, NLP, model, comprises: inputting annotated training data to first architecture portions of the neural network, the first architecture portions being executed respectively in a plurality of distributed client computing devices in communication with a server computing device, the training data being derived from text data private to the client computing device in which the first architecture portion is executed, the server computing device having no access to any of the private text data; deriving from the training data, using the first architecture portions, weight matrices of numeric weights which are decoupled from the private text data; concatenating the weight matrices, in a second architecture portion of the neural network executed in the server computing device, to obtain a single concatenated weight matrix; and training, on the second architecture portion, the NLP model using the concatenated weight matrix.

Abstract: A text style transfer system is described that generates different stylized versions of input text by rewriting the input text according to a target style. To do so, the text style transfer system employs a variational autoencoder to derive separate content and style representations for the input text, where the content representation specifies semantic information conveyed by the input text and the style representation specifies one or more style attributes expressed by the input text. The style representation using counterfactual reasoning to identify different transfer strengths for applying the target style to the input text. Each transfer strength represents a minimum change to the input text that achieves a different expression of the target style. The transfer strengths are then used to generate style representation variants, which are each concatenated with the content representation of the input text to generate the plurality of different stylized versions of the input text.

Abstract: A method for reporting a success of a dialog session including generating an outcome data indicative of a positive outcome of the dialog session with at least one of the plurality of natural language responses, receiving a natural language query, converting the natural language query into a structured database query based on the metadata, performing the structured database query on a database, generating a result in response to a response data received in response to the structured database query, generating a natural language responses in response to the result and the plurality of natural language response data, sending, by the chatbot, the natural language response, generating a result metadata in response to a comparison of the result to the outcome data, and generating an outcome log in response to the result metadata wherein the outcome log includes an indication of the positive outcome of the dialog session.

Abstract: Systems and methods including one or more processors and one or more non-transitory storage devices storing computing instructions configured to run on the one or more processors and cause the one or more processors to perform receiving one or more user utterances; re-training a pre-trained natural language processing (NLP) algorithm on the one or more user utterances; using the pre-trained NLP algorithm, as re-trained, as one or more layers in a neural network; combining at least one first output of at least one first output layer of the neural network with at least one second output of at least one second output layer of the neural network to create a final output of the neural network, wherein: the at least one first output layer of the neural network is different than the at least one second output layer of the neural network; and wherein the pre-trained NLP algorithm, as re-trained, is used to determine the final output of the neural network; and coordinating displaying a customized graphical user interf

Abstract: Devices and techniques are generally described for cohort determination in natural language processing. In various examples, a first natural language input to a natural language processing system may be determined. The first natural language input may be associated with a first account identifier. A first machine learning model may determine first data representing one or more words of the first natural language input. A second machine learning model may determine second data representing one or more acoustic characteristics of the first natural language input. Third data may be determined, the third data including a predicted performance for processing the first natural language input by the natural language processing system. The third data may be determined based on the first data representation and the second data representation.

Abstract: An apparatus and method for converting data is disclosed. The apparatus includes at least a processor and a memory communicatively connected to the at least a processor, wherein the memory contains instructions configuring the at least a processor to extract, using a data extraction module, user activity data from user device data, classify the user activity data into one or more user activity data groups, convert, using a data converting module, the user activity data to system data as a function of the one or more user activity data groups and generate, using a report generation module, a user activity report as a function of the system data.

Abstract: A system and method for developing a treatment plan using multi-stage machine learning. A method includes determining a treatment plan for a patient based on at least one mental health disorder of a patient, wherein the treatment plan includes a plurality of digital therapeutics exercise tasks, wherein each digital therapeutics exercise task is selected from among a category of digital therapeutics exercise tasks corresponding to a type of mental health disorder of the at least one mental health disorder of the patient; and administering treatment to the patient by prescribing the treatment plan to the patient and causing data for administering the treatment plan to a user device of the patient.

Abstract: A computer-implemented method, system and computer program product for generating rules to be utilized by a rule-based system. A parser analyzes a received string of words according to grammatical rules to generate a parse tree. The parse tree may then be reviewed by a user to determine if such a parse tree represents the appropriate syntactic structure of the string of words. If the user believes that the parse tree does not represent the appropriate syntactic structure of the string, then the user, who does not need to possess the expertise of natural language processing, may select the nodes and/or edges of the parse tree in a particular order that corresponds to the correct syntactic structure of the string of words. A grammatical rule is then created for parsing the string of words based on the user selection of nodes and/or edges of the parse tree in a particular order.

Abstract: Highly accurate document search, especially intellectual property-related document search, is achieved with a simple input method. A processing portion has a function of generating text analysis data from text data input to an input portion; a function of extracting a search word from words included in the text analysis data; and a function of generating first search data from the search word on the basis of weight dictionary data and thesaurus data. A memory portion stores second search data generated when the first search data is modified by a user. The processing portion updates the thesaurus data in accordance with the second search data.

Abstract: Systems and methods for 3D printer management can allow or reject printing of an object based on a model that is trained with machine learning. In one example, the model classifies the object according to object type. The object type can be compared against a list to determine whether to block the object from printing. The object can also be compared to previously printed objects. The object can be blocked from printing if a combination of the object and a previously printed object amounts to a threshold amount of a blacklisted item. The lists can be specific to users, such as an organizational group to which the user belongs. A print server can apply the model prior to forwarding the object to a 3D printer for printing. Both the models and the lists can evolve based on machine learning, such as based on which print decisions receive override from administrators.

Abstract: Systems and methods for generating response records for individual users based on user responses to prompts. Exemplary implementations may: provide prompts to individual users, receive user input from individual ones of the users indicating responses to the one or more provided prompts; determine, based on the received user input from individual ones of the users, response information for individual ones of the responses to the prompts; aggregate the individual responses and corresponding response information to generate response records for individual ones of the users; store the response records for individual ones in the users; and/or other exemplary implementations.

Abstract: Embodiments of this disclosure disclose a sentence processing method and device. The method may include performing word segmentation operation on a source sentence to be encoded to obtain m words. The method may further include obtaining an ith word in the m words using an ith encoding processing node in the n encoding processing nodes, and obtaining an (i?1)th word vector from an (i?1)th encoding processing node. The method may further include performing linear operation and non-linear operation on the ith word and the (i?1)th word vector using the first unit of the ith encoding processing node to obtain an ith operation result, and outputting the ith operation result to the at least one second unit for processing to obtain an ith word vector. The method may further include generating, in response to obtaining m word vectors, a sentence vector according to the m word vectors.

Abstract: A computer-implemented method, a computer program product, and a computer system for dependency tree-based data augmentation for sentence well-formedness judgement. A computer applies a dependency parser to generate a dependency tree for a sentence. The computer removes one or more nodes in the dependency tree, according to a removal ratio for a predetermined rating score. The computer generates, from the dependency tree, a partial tree for the sentence. The computer outputs a rated sentence based on the partial tree. The rated sentence is used as training data.

Abstract: There is a need for more accurate and more efficient natural language solutions with greater semantic intelligence. This need can be addressed, for example, by natural language processing techniques that utilize predictive entity scoring. In one example, a method includes determining an overall prevalence score for the input entity data object with respect to a scored document corpus and a target section; determining a qualified prevalence score for the input entity data object with respect to a high-scoring subset of the scored document corpus; processing the input entity data object using an entity scoring machine learning model to generate the predicted entity score, wherein the entity scoring machine learning model may characterized by a plurality of multiplicative hyper-parameters and one or more additive hyper-parameters; and performing one or more prediction-based actions based at least in part on the predicted entity score.

Abstract: A method of using a computing device using natural language software to respond to a Boolean question is presented. The method includes receiving, by a computing device, a question and at least one passage. The computing device classifies the question as a Boolean type question. The computing device generates evidence from the at least one passage to the Boolean question. The computing device generates a response to the Boolean question using the generated evidence. The computing device provides the generated evidence that supports the response to the Boolean question.

Abstract: A system that manages a dialog between a user and an AI system using a hierarchically structured tree of topics. An illustrative application is collection of a personal history. A topic manager repeatedly selects a topic and requests the AI system to generate a question based on summary information stored in the associated topic node; this avoids sending the entire dialog to the AI system each time, which would be impossible for long conversations due to AI context window limitations. The AI system analyzes each response from the user to identify new topics to add to the tree, and to update the summary associated with each topic node. Vector embeddings of topic names may be used to place new topic nodes in the most relevant position in the tree. Topics may be selected using a biased tree walk that preferentially selects nodes that have been less explored in previous questions.

Abstract: Embodiments operate a machine learning (“ML”) notebook in a cloud infrastructure executing a plurality of ML notebooks. Embodiments receive a plurality of previously executed ML notebook feature engineering commands from the plurality of ML notebooks. Embodiments store the plurality of previously executed ML notebook feature engineering commands, including a relationship between the feature engineering commands. Embodiments mine the stored commands to generate feature engineering sets of feature engineering commands, the feature engineering sets comprising feature engineering commands that are frequently used together and an order of use of the feature engineering commands. Embodiments then receive a context of a current feature engineering command and data used in the context and recommend a next feature engineering command to be executed after the current feature engineering command.

Abstract: Systems and methods for content framing monitoring and intervention are disclosed. In one example, a method includes receiving, at a computing device, first text conveying first information, performing natural language processing of the first text, determining a first framing of the first information based on the natural language processing of the first text, determining second text conveying the first information with a second framing, different from the first framing, and outputting the second text on an electronic display.

Abstract: A system may include a memory and a processor in communication with the memory. The processor may be configured to perform operations that include analyzing interactions by a user within a network and generating a user profile for the user. The operations by the processor may further include identifying an attempt by the user to share a post via the network and prompting the user with a personalized alert to evaluate the post, wherein the personalized alert is generated based on the interactions, the user profile, and the properties of the post.

Abstract: A system for generating a contextualized text using a character-based convolutional neural network architecture (CNN) is disclosed. The system includes a memory, a character-based CNN engine having executable modules and a processor configured to execute the character-based CNN engine and the executable modules for generating contextualized text using a character-based convolutional neural network architecture (CNN).

Abstract: Mechanisms for implementing a text encoder and text encoder operations are provided. A contrastive machine learning training operation trains an encoder of a machine learning computer model, to learn a sense and similarity preserving embedding that operates to encode input natural language text data to generate encoded natural language text data based on a sense attribute of one or more terms in the input natural language text data. The contrastive machine learning training operation learns to separate positive samples in training data from negative samples in the training data. The trained encoder processes a term specified in an input natural language text to generate an encoded natural language text based on the learned embedding and inputs, to a downstream computing system, the encoded natural language text to cause the downstream computing system to perform a computer natural language processing operation based on the embedding.

Abstract: Various embodiments of the present invention provide methods, apparatus, systems, computing devices, computing entities, and/or the like for perform predictive data analysis operations using natural language input data. For example, certain embodiments of the present invention utilize systems, methods, and computer program products that perform predictive data analysis operations by using sentence embedding machine learning models that are trained in coordination with similarity-based machine learning models.

Abstract: Disclosed implementations include systems, methods, and apparatus that process multiple, disparate streams of data, determine correlations and relationships between the data and provide natural language responses that provide insights for events or activities that have occurred and foresights for events or activities that are forecasted to occur. The disclosed implementations include a model that understands data statistics and provides both insights and foresights that are backed with statistical support that can be presented to and understood by operators. Still further, the disclosed implementations are capable of operating at edge locations that may be frequently or permanently disconnected from conventional or cloud based systems.

Abstract: A system may be configured to detect custom acoustic events, where the system generates an acoustic event profile for the custom acoustic event based on a natural language description provided by a user and using an audio sample of the described acoustic event. For example, the user may describe the custom acoustic event as “dog bark.” The system may ask the user questions to refine the description (e.g., dog breed, dog gender, age, etc.). Using an audio sample of the refined description, the system may then determine that audio captured in the user's environment is a potential sample of the custom acoustic event. Such captured audio may be presented to the user for confirmation, and then may be used to detect future occurrences of the custom acoustic event in the user's environment.

Abstract: In a method for generating a normalized sequential model using a processor, a sequential energy-based model computed by a parameterized neural network is provided. The sequential energy-based model defines an unnormalized probability distribution over a target sequence for a context source. The normalized sequential model is generated by projecting the sequential energy-based model onto a target autoregressive model that approximates a normalized distribution associated with the sequential energy-based model.

Abstract: Implementations set forth herein relate to employing dynamic regulations for governing responsiveness of multiple automated assistant devices, and specifically the responsiveness an automated assistant to a given spoken utterance that has been acknowledged by two or more of the assistant devices. The dynamic regulations can be context-dependent and adapted over time in order that the automated assistant can accommodate assistant interaction preferences that may vary from user to user. For instance, a spoken utterance such as “stop,” may be intended to affect different assistant actions based on a context in which the user provided the spoken utterance. The context can refer to a location of the user relative to other rooms in a home, a time of day, a user providing the spoken utterance, an arrangement of the assistant devices within a home, and/or a state of each device in the home.

Abstract: The present disclosure relates to a vehicle quality problem management system and a data processing method thereof. The system includes a management server including a server communication device that performs wireless communication with a mobile device, and a server processing device connected to the server communication device, and the server processing device receives a voice signal containing a current quality problem from the mobile device, converts the current quality problem in the voice signal to text using speech to text (STT), and registers the current quality problem converted into the text in a database (DB).

Abstract: An embodiment for creating a dialog based on anticipated questions for database driven conversations is provided. The embodiment may include receiving content from a database. The embodiment may also include identifying one or more schemas, one or more entities, and relational data in the content. The embodiment may further include identifying a semantic type and a number of distinct entries for each entity. The embodiment may also include presenting choices for one or more query targets and one or more filtering conditions to a user. The embodiment may further include prompting the user for one or more annotations and one or more clarifying questions for each chosen query target and filtering condition. The embodiment may also include generating a plurality of modular phrases and combining the plurality of modular phrases into one or more sentences and paraphrases of the one or more sentences.

Abstract: The in-home care of seniors is augmented using Internet of Things (IOT) technologies. In-home sensors monitor a senior and their caregiver. Physical conditions and psychological conditions may be monitored. In some implementations, a machine learning system has a classifier trained to detect a specified condition, such as depression. The system may perform various transformations of raw sensor data into a format indicative of a particular condition. In one implementation, a psychological or medical condition has symptoms in which each symptom has one or more measurable events. Mappings between symptoms, events, sensor data, and sensor transformation functions may be supported.

Abstract: Systems and methods are provided for generating a content item banner for display based on a session type and a content item type. A user profile is accessed at a computing device, and a session type associated with consumption of at least a content item of a plurality of content items is determined based on the user profile. A content item type associated with the plurality of content items is determined, and a content item banner for a subsequent content item of the plurality of content items is identified based on the session type and the content item type. The content item banner is generated for display.

Abstract: A document processing system is configured to identify, for each accessed electronic document in a first set of multiple electronic documents, a set of identified multi-word phrases determined to be in ordered text information in the accessed electronic document, each multi-word phrase of the set of identified multi-word phrases including adjacent words in the ordered text information; and determine, for each accessed electronic document in the first set of multiple electronic documents, a selected document type from the first set of document types based at least on an analysis of the set of identified multi-word phrases with respect to multi-word-phrase characteristics identified by a first definition and associated with each document type in a first set of document types associated with a first document-set type.

Abstract: Systems, methods and non-transitory computer readable media for attributing generated audio contents to training examples are provided. A first audio content generated using a generative model may be received. The generative model may be a result of training a machine learning model using training examples. Each training example may be associated with a respective audio content. Properties of the first audio content may be determined. For each training example of the training examples, the respective audio content may be analyzed to determine properties of the respective audio content. The properties of the first audio content and the properties of the audio contents associated with the training examples may be used to attribute the first audio content to a subgroup of the training examples. A respective data-record associated with a source associated with the training examples of the subgroup may be updated based on the attribution.

Abstract: Techniques are described herein for producing machine-generated findings given a set of user experience test results. In some embodiments, the system generates the findings using an artificial intelligence and machine learning engine. The findings may highlight areas that are predicted to provide the most insight into optimizing a product's design. A finding may be generated based on all or a subset of the test result elements, including qualitative and/or quantitative data contained therein. A finding may summarize a subset of the UX test results that are interrelated. A finding may link a summary to one or more references extracted from the set of test results to show support for the machine-generated insights in the underlying raw test data. Machine-generated findings reports may provide near instantaneous guidance for optimizing product designs while removing extraneous information from a vast quantity of raw test result data.

Abstract: Methods, apparatus, systems, and articles of manufacture are disclosed to determine mutex entropy for malware classification. An example apparatus includes interface circuitry to access a mutex associated with a software application, the mutex to include a mutex identifier string, normalizer circuitry to normalize the mutex identifier string, character probability circuitry to determine character probabilities of characters within the normalized mutex identifier string, the character probabilities based on a historical mutex character distribution, entropy calculator circuitry to calculate an entropy value for the mutex based on the character probabilities, classifier circuitry to classify the mutex as clean or malicious based on the entropy value, and protector circuitry to mitigate malicious attacks based on the classification.

Abstract: In various embodiments, a bootstrapping training subsystem performs sampling operation(s) on a training database that includes subjective scores to generate resampled dataset. For each resampled dataset, the bootstrapping training subsystem performs machine learning operation(s) to generate a different bootstrap perceptual quality model. The bootstrapping training subsystem then uses the bootstrap perceptual quality models to quantify the accuracy of a perceptual quality score generated by a baseline perceptual quality model for a portion of encoded video content. Advantageously, relative to prior art solutions in which the accuracy of a perceptual quality score is unknown, the bootstrap perceptual quality models enable developers and software applications to draw more valid conclusions and/or more reliably optimize encoding operations based on the perceptual quality score.

Abstract: Systems and methods for multilingual keyword generation for low-resource language passages are disclosed. For example, high-resource language passages describing items in a catalog or other such passages may include numerous keywords for classifying and searching for the item while a low-resource language description of the item may not include the same rich keyword environment due to lower usage and traffic in the low-resource language. The systems and methods herein provide for leveraging the high-resource language through a multilingual natural language processing algorithm to identify similar passages across language barriers and identify keywords in the high-resource language for input into a decoder to generate keywords in the low-resource language.