Abstract: Techniques are disclosed for organizing and distributing artifacts generated by processing pipelines for the training or application of machine learning models. An application may subscribe to a playlist of a stream of events and locally store a copy of the playlist. The subscriber may merge locally stored and/or selected events to generate a merged stream of events. The subscriber may then execute the merged event stream including the newly added instance of the event.

Abstract: Techniques are disclosed for revising training data used for training a machine learning model to exclude categories that are associated with an insufficient number of data items in the training data set. The system then merges any data items associated with a removed category into a parent category in a hierarchy of classifications. The revised training data set, which includes the recategorized data items and lacks the removed categories, is then used to train a machine learning model in a way that avoids recognizing the removed categories.

Abstract: The present embodiments relate to a cloud shell extension framework. A cloud infrastructure (CI) service can invoke a cloud shell from a console to the CI. The service may request that context data to be added for use in a terminal session. The cloud shell extension can include a tool or script that can be used to obtain context data or sample code for preparing an environment specific to the service prior to the user interaction with the environment. The cloud shell extension can allow for a service to register an extension in the cloud shell framework, and their extension can be invoked when a client initiates the cloud shell with the service. The extension can allow for the invoked service to forward context data, such as environment variables, to be set in the terminal session for the client.

Abstract: Example embodiments facilitate prioritizing the recycling of computing resources, e.g., server-side computing systems and accompanying resources (e.g., non-volatile memory, accompanying firmware, data, etc.) leased by customers in a cloud-based computing environment, whereby computing resources (e.g., non-volatile memory) to be forensically analyzed/inspected, sanitized ,and/or updated are prioritized for recycling based on estimates of when the computing resources are most likely to require recycling, e.g., via background sanitizing and updating. Computing resources that are likely to be recycled first are given priority over computing resources that are more likely to be recycled later. By prioritizing the recycling of computing resources according to embodiments discussed herein, other cloud-based computing resources that are used to implement computing resource recycling can be efficiently allocated and preserved.

Abstract: Embodiments described herein include receiving an application flow defining logic for triggering automated actions for a candidate post on one or more web-based communication channels. An artificial intelligence (AI) powered web crawler may crawl a set of websites for target content. Responsive to identifying target content, machine learning may be leveraged to identify an entity name for an entity that is a subject of the target content and associate the entity with one or more event categories. Based on the event categories, a segment may be generated at runtime for the application flow where segment membership was not known at the time the application flow was defined or before application runtime.

Abstract: Techniques for configuring autosave triggers in a computing environment based on environment and data conditions are disclosed. A system trains a machine learning model based on data attributes and environmental attributes to generate autosave value triggers for a computing environment. The autosave value triggers are triggered by different conditions. For example, one autosave trigger may be triggered when an error condition is detected. Another may be triggered when a certain number of operations are performed. The machine learning model generates autosave trigger values scores for one or more autosave triggers. The system may implement the autosave triggers in the computing environment based on the autosave trigger values.

Abstract: We disclose a system that executes an inferential model in VRAM that is embedded in a set of graphics-processing units (GPUs). The system obtains execution parameters for the inferential model specifying: a number of signals, a number of training vectors, a number of observations and a desired data precision. It also obtains one or more formulae for computing memory usage for the inferential model based on the execution parameters. Next, the system uses the one or more formulae and the execution parameters to compute an estimated memory footprint for the inferential model. The system uses the estimated memory footprint to determine a required number of GPUs to execute the inferential model, and generates code for executing the inferential model in parallel while efficiently using available memory in the required number of GPUs. Finally, the system uses the generated code to execute the inferential model in the set of GPUs.

Abstract: Techniques for generating confidence scores for machine learning predictions are disclosed. The confidence score for a predicted label corresponding to a target data point is based at least in part on how well the machine learning model predicts labels for other data points that are similar to the target data point. The system uses k data points, closest to the target data point, with known labels to compute the confidence score of a predicted label for the target data point. The accuracy of the predictions and the distance of each of the k data points from the target data point are used to compute a confidence score for a label predicted for the target data point.

Abstract: “Resource guarantee” refers to a unit of a resource that is guaranteed and therefore designated to a consumer. A multi-phased constraint programming (CP) approach is used to determine assignments of resource guarantees of a set of consumers to a set of hosts in a resource system. Phase I uses CP to segregate non-split consumers from split consumers. Phase II uses CP to assign each cotenant group of non-split consumers to a respective host. Phase III uses CP to assign resource guarantees of the split consumers across the hosts, wherein resource guarantees of a single split consumer may be splits across different hosts. Each phase involves execution of a CP solver based on a different CP data model. A CP data model declaratively expresses combinatorial properties of a problem in terms of constraints. CP is a form of declarative programming.

Abstract: A system and method for enabling a cloud computing environment. In accordance with an embodiment, the system can include a variety of hardware and/or software components and features, which can be used in delivering an infrastructure, platform, and/or applications to support public and private clouds.

Abstract: Disclosed is an approach to implement improved anomaly detection. Improved anomaly detection is provided using MSET-SPRT via Monte Carlo simulation that can address problems with conventional MSET-SPRT approaches and provide improved system performance and accuracy.

Abstract: A library uplift tool may automatically identify changes to elements of an between different versions of a software library interface. The public interfaces of the two versions may be compared to identify changes to datatypes, variable or function names, parameter names, parameter lists, and other elements of the library interface. Locations in application source code using the software library may then be identified and linked with the changes between the library versions. A report or user interface may then display specific locations in the application source code that are affected by the changes to the library versions. For each change, a suggested action may be presented in the interface, and an example of modified source code may be generated. The interface may be interactive such that modified source code may be automatically propagated to source code files for compatibility with the new library version.

Abstract: The disclosed embodiments relate to a system that compactly stores time-series sensor signals. During operation, the system receives original time-series signals comprising sequences of observations obtained from sensors in a monitored system. Next, the system formulizes the original time-series sensor signals to produce a set of equations, which can be used to generate synthetic time-series signals having the same correlation structure and the same stochastic properties as the original time-series signals. Finally, the system stores the formulized time-series sensor signals in place of the original time-series sensor signals.

Abstract: Techniques for improving a semantic parser of a dialog system, by breaking the semantic parser into a coarse semantic parser and a fine semantic parser, are described. A method described herein includes accessing an utterance received in a dialog system. The utterance is a text-based natural language expression. The method further includes applying a coarse semantic parser to the utterance to determine an intermediate logical form for the utterance. The intermediate logical form indicates one or more intents in the utterance. The method further includes applying a fine semantic parser to the intermediate logical form to determine a logical form for the utterance. The logical form is a syntactic expression of the utterance according to an established grammar, and the logical form includes one or more parameters of the one or more intents. The logical form can be used to conduct a dialog with a user of the dialog system.

Abstract: Embodiments provide systems and methods for navigating a dialogue flow using a trained intelligence bot. Upon initiation of a chat session between a user and a trained intelligence bot, one or more utterances can be received. The utterances can be processed using the trained intelligence bot to resolve an intent from among a plurality of predefined intents, where the intelligence bot is trained to resolve predefined intents based on training data associated with the predefined intents. A predefined dialogue flow associated with the resolved intent can be navigated using the intelligence bot, where the intelligence bot guides the user through the dialogue flow using context variables that are associated with the user or the chat session. The user can be provided enterprise data retrieved by the intelligence bot using a retrieval request generated based on one or more of the navigation of the dialogue flow or the context variables.

Abstract: According to one or more embodiments, lookup, insertion, and deletion operations are allowed to continue during actions required for collision remediation. When relocation operations are used to resolve a collision, information encoded in header portions of the hash table entries that store the key-value pairs indicates when the associated key-value pairs are undergoing relocation. This information facilitates continued access to the RKVS during the relocation process by allowing other processes that access the RKVS to handle relocations without failure. Furthermore, when hash table expansion is needed in order to resolve a collision, a second, larger, hash table is allocated, and lookup operations continue on both the old hash table and the new hash table. One or more embodiments further prevent insertion, lookup, and deletion failures in the RKVS using flags, encoded in header information in hash table entries, that reflect the state of the respective key-value pairs in the store.

Abstract: Some embodiments process data as defined by a polymorphic data model. A server may provide the polymorphic data model to a client application from a datastore. The functionality or features of the client application may be controlled as a function of the polymorphic data model. The client application may thus provide different functionalities when provided with different polymorphic data models. Updates to the polymorphic data model may be used to update the client application's functionality without any updates to the client application's executable code.

Abstract: In one or more embodiments, a cloud operation reservation system is provided through which cloud operations may be scheduled and managed. The cloud operation reservation system includes logic for defining a set of time windows that are available to perform one or more cloud operations on cloud targets and presenting the set of time windows to one or more tenants of a cloud service. Tenants may browse the presented set of time windows and submit reservation requests to update and/or perform other operations on cloud targets. In response to receiving, a request to reserve one or more slots in a particular time window, the cloud operation reservation system schedules one or more cloud operations to be performed on a cloud target that is available to at least one tenant.

Abstract: A method for identifying a category of offerings can include receiving, by a data management platform (DMP) executing on a computing platform, user data for a user of an end-user device. The method can also include mapping, by the DMP, identifiers (IDs) from received user data for the user to a plurality of different categories of offerings, wherein a particular category of offerings of the plurality of different categories of offerings corresponds to a leading commerce stream executed by the user. The method can also include matching, by the DMP, the category of the leading commerce stream to a second category of offerings for a following commerce stream based on commerce data characterizing a plurality of commerce streams executed in a window of time. The method can further include providing, by the DMP, the second category of offerings.

Abstract: An autonomous agent creates a first semantic tree from a question and second semantic tree from a candidate answer. The agent identifies, between the first semantic tree and the second semantic tree, common subtrees and calculates a semantic alignment score from a sum of sizes of each of the common subtrees. The agent forms a first syntactic tree for the question and a second syntactic tree for the candidate answer. The agent identifies a number of common syntactic nodes between the first syntactic tree and the second syntactic tree. The agent calculates a syntactic alignment score based on the number of common syntactic nodes. Responsive to determining that a sum of the semantic alignment score and the syntactic alignment score is greater than a threshold, the agent outputs the candidate answer to a device.

Abstract: Techniques are disclosed for providing adaptive recommendations for a data set. A data set can include one or more columns of data. The data set can be profiled in order to identify actions that can be applied to the data in order to enrich the data. The data set and actions that were applied to the data set can be stored. Actions that are applied to subsequent data sets can take into account the actions that were applied to prior data sets having similar profiles.

Abstract: Vertex/edge table rows are mapped to unique integer identifiers, to enable construction of in-memory representation of a graph from existing, unmodified RDBMS tables. The unique integer identifiers are based on an encoding of primary key values of the table rows. The unique integer identifiers are used as graph indexes of the in-memory representation.

Abstract: A distributed event processing system is disclosed that receives continuous data streams, registers a continuous query against the data streams, and continuously executes the query as new data appears in the streams. In certain embodiments, the distributed event processing system deploys and executes applications (e.g., event processing applications) by distributing the execution of the application on a cluster of machines within the system. In certain embodiments, the system provides users with the ability to specify logical rules in an application. The system processes the logical rules by generating a DAG of transformations representing the logical rules and converts the DAG of transformations into a logical rules Resilient Distributed Dataset (RDD) DAG of transformations. The system processes events in an event batch against the RDD DAG of transformations to generate a set of output results for the user.

Abstract: The present disclosure relates to techniques for using variant inconsistency attack (VIA) as a simple and effective adversarial attack method to create useful adversarial examples for adversarial training of machine-learning models. In one particular aspect, a method is provided that includes obtaining a set of input examples for attacking a machine-learning model (the set of examples do not have corresponding labels), modifying an example from the set of examples in a utility preserving manner to generate a pair of modified examples, attacking the machine-learning model with the pair of modified examples in order generate a pair of predictions for the pair of modified examples, comparing the pair of predictions to determine whether the pair of predictions are the same or different, and in response to the pair of predictions being different, adding the pair of modified examples to a set of adversarial examples.

Abstract: Resource allocation limitations include resource limits and resource guarantees. A consumer is vulnerable to interruption by other consumers if using more resources than guaranteed. Resources are designated and/or assigned to consumers based on resource limits and resource guarantees. A constraint programming (CP) solver determines resource limits and resource guarantees that minimize vulnerability and/or vulnerability cost based on resource usage data. A CP data model includes limit elements, guarantee elements, and vulnerability elements. The CP data model further includes guarantee-vulnerability constraints, which relies on exceedance distributions generated from resource usage data for the consumers. The CP data model declaratively expresses combinatorial properties of a problem in terms of constraints. CP is a form of declarative programming.

Abstract: A distributed computing system is described that leverages a nearline storage layer to minimize the downtime required for bootstrapping a new computing cluster in the distributed computing system. The system executes a computing cluster comprising a set of computing nodes and determines a set of one or more data segments to be written to a nearline storage system. The system writes the data segments to the nearline storage system. In certain examples, the system receives a request to create a second computing cluster and responsive to the request, bootstraps the second computing cluster using the set of data segments stored on the nearline storage system. The system additionally leverages the nearline storage layer to accelerate query processing by the computing nodes of a computing cluster.

Abstract: Update-anywhere replication of queuing operations on a replicated message queue is performed. A dequeue ready time (“ready time”) is associated by each participating persistent storage server with a queue message to be dequeued. Unless a queue message is already locked by a distributed dequeue transaction, a participating leader PSS initiates a distributed dequeue transaction for the queue message once the ready time for the queue message is reached, subject to certain conditions. An initiator PSS is in effect designated for a queue message; the initiator PSS associates a desired ready time for the queue message. The designated PSS is referred to herein as the primary leader PSS and the ready time the primary PSS associates with queue message is referred as the primary ready time. The other participating leader PSSs are backup leader PSSs which serve as backups for dequeuing a queue message. Each backup leader PSS associates a later “backup ready time” with the queue message.

Abstract: Described herein are systems and methods for implementing a task assignment protocol. In one or more embodiments, a task management system receives task data and resource data. Responsive to the receipt, the task management system receives input for selection of a resource. A candidate subset of tasks that match the properties of the resource is then identified. Upon receipt of selection of the candidate task, a task icon updates. The task icon may update to reflect the resource identifier associated the assigned resource. Additionally, or alternatively, a resource icon is updated to reflect the availability of the resource. The protocol repeats until one or more resources are allocated to the tasks. A resource chart additionally displays to aid in evaluation of resource availability.

Abstract: Indications of a plurality of events whose occurrence is detected in a particular execution of a program are obtained. One or more partitions of a trace object corresponding to the execution are constructed, including a first partition corresponding to a first subset of the events. The first partition comprises a header portion which includes a compressed representation of one or more event chains, and a data portion comprising a compressed events record indicating an occurrence, during the execution, of a particular sequence of events indicated by an event chain. The trace object is stored.

Abstract: A method for supporting multiple locations in a preferred-locality attribute NF discovery and selection includes receiving, from producer NFs, NF registration messages including NF profiles with locality attributes and registering the producer NFs by storing the NF profiles including the locality attributes in an NF registration database local to an NRF. The method further includes receiving, from a requesting NF, an NF discovery request having query parameters for discovering producer NFs, the query parameters including a preferred-locality attribute identifying a plurality of different preferred producer NF locations. The method further includes retrieving NF profiles from the NF registration database having locality attributes that match any of the preferred producer NF locations. The method further includes generating an NF discovery response including the NF profiles retrieved from the NF registration database. The method further includes forwarding the NF discovery response to the requesting NF.

Abstract: A method for detecting malicious code may include generating, from deserialization examples, a finite automaton including states. The states may include labeled states corresponding to the deserialization examples. A state may correspond to a path from a start state to the state. The method may further include while traversing the states, generating a state mapping including, for the state, a tracked subset of the states, determining that the path corresponds to a path type, inferring, using the path type and the state mapping, a regular expression for the state, and determining, for a new deserialization example and using the regular expression, a polarity indicating whether it is safe to deserialize the new deserialization example.

Abstract: Techniques are disclosed for utilizing control packets to manage flows by a smart network interface card (smartNIC). In one example, an accelerator of the smartNIC determines that a cache entry of a cache that is managed by the accelerator is a candidate for removal. The cache entry stores flow state of a particular flow. The accelerator generates a control packet that includes flow information of the particular flow that is formatted utilizing a particular header format, the flow information operable for generating a hash that indexes to the cache entry. The accelerator includes an instruction within the control packet that requests a programming data plane of the smartNIC to provide instructions for removing the cache entry from the cache. Upon receiving the control packet, the programming data plane generates and transmits a second instruction to the accelerator for removing the cache entry from the cache.

Abstract: Embodiments provide systems and methods for learning extract, transform, and load mappings by example. A plurality of features can be extracted from a source schema and a target schema. Example ETL mappings can be provided to a machine learning algorithm that comprise definitions for extracting data from source tables and loading the extracted data into target tables. Using the machine learning algorithm and based on the source schema, target schema, and extracted features, one or more ETL rules can be predicted that define logic for extracting data from the source schema and loading the extracted data into the target schema. Additional ETL mappings can be generated based on the predicted ETL rules, the additional ETL mappings providing additional definitions for extracting data from the source schema and loading the extracted data into the target schema.

Abstract: Techniques are described herein for training and evaluating machine learning (ML) models for document processing computing applications based on in-domain and out-of-domain characteristics. In some embodiments, an ML system is configured to form feature vectors by mapping unknown tokens to known tokens within a domain based, at least in part, on out-of-domain characteristics. In other embodiments, the ML system is configured to map the unknown tokens to an aggregate vector representation based on the out-of-domain characteristics. The ML system may use the feature vectors to train ML models and/or estimate unknown labels for the new documents.

Abstract: System and method for aggressive credit waiting in a high performance computing environment. In accordance with an embodiment, systems and methods can provide for an indexed matrix of credit wait policies between ports within a single switch. In addition, systems and methods can provide for an array of credit wait polices at an egress port from a switch, the array being indexed by virtual lane.

Abstract: A computer comprising multiple processors and non-uniform memory implements multiple threads that perform a lock operation using a shared lock structure that includes a pointer to a tail of a first-in-first-out (FIFO) queue of threads waiting to acquire the lock. To acquire the lock, a thread allocates and appends a data structure to the FIFO queue. The lock is released by selecting and notifying a waiting thread to which control is transferred, with the thread selected executing on the same processor socket as the thread controlling the lock. A secondary queue of threads is managed for threads deferred during the selection process and maintained within the data structures of the waiting threads such that no memory is required within the lock structure. If no threads executing on the same processor socket are waiting for the lock, entries in the secondary queue are transferred to the FIFO queue preserving FIFO order.

Abstract: Techniques are disclosed for tracing memory components in asset management systems. A computing device may receive an indication that a new device has been connected to a network. The computing device receives a first set of memory specifications from the new device and a second set of memory specifications from a SoV database. The computing device then generates a memory-asset data structure that stores a third set of memory specifications, each memory specification of the third set of memory specifications being a memory specification that is in both the first set of memory specifications and the second set of memory specifications. The computing device assigns, memory specifications of the third set of memory specifications, a data privacy level that is based on a sensitivity of data stored in the component of the new device. The computing device may then transmit the memory-asset data structure.

Abstract: Techniques for modifying queries in a set of nested queries are disclosed. A graphical user interface displays a query detail region alongside a nested query display region. The graphical user interface includes functionality to provide for modification of queries in the nested set of queries. Based on a selection by a user, a query modification tool promotes a query attribute from a child query to one or more parent queries. Based on another selection by a user associated with one query in the set of nested queries, the system deletes an attribute from each query in the set of nested queries. Responsive to a selection to create multiple conditions for a query rule, the system modifies the functionality of the user interface to enable entry of multiple condition characteristics. Based on a further selection, the system creates the multiple conditions for the query rule.

Abstract: The present embodiments relate to implementing change data on no-master NoSQL data stores. An optimized node can be identified from a plurality of NoSQL data storage nodes and a specialized node can be connected (e.g., collocated) to the optimized node. The specialized node can maintain change data capture (CDC) data provided by client nodes in a hash map that can be used as a point of truth for coordinating CDC data across the plurality of NoSQL data storage nodes. The plurality of NoSQL data storage nodes can identify and coordinate all read/write data obtained from multiple client devices in a geographically separated large-scale (e.g., planet scale) system to identify change data in a distributed data store. The specialized data can provide read data to devices in the large-scale system to reconcile inconsistencies in change data across nodes in the large-scale system.

Abstract: Implementations generally relate to a method of automatically generating applications for multiple platforms based on one source of application code. Implementations include receiving a signal from a client device to initiate an application generation process, for example, through a graphical user interface (GUI). In an implementation, the application generation process analyzes the application code to differentiate between core application code, which is shared between many platforms, and device-specific code, and then separates the core application code from the device-specific code. The core application code may then be deployed for use in a plurality of applications such as a web application, progressive web application, or, as taken together with device-specific code, as a native mobile application. Applications may be deployed with features that are automatically determined based on their deployment method and the feature set of each platform.

Abstract: Techniques are disclosed for processing flows by a smart network interface card (smartNIC) based on modifying a packet. In one example, a smartNIC accelerator receives a packet from a first port of the smartNIC, the first port being connected to a splitter device that splits a first data path into a second data path and a third data path, and the packet arriving at the first port via the second data path. The accelerator modifies the packet to indicate that the packet arrived at the first port via the second data path. The accelerator inserts the modified packet into a queue that is associated with both the second data path and the third data path. A programming data plane of the smartNIC receives and then processes the modified packet based on determining that the packet arrived at the first port via the second data path.

Abstract: Disclosed is an improved approach to implement a mechanism to provide customer control over access to cloud infrastructure by the cloud provider's operator employees. This mechanism allow customer controlled access to any cloud infrastructure that belongs to or is otherwise allocated to the customer.

Abstract: Techniques for generating human-readable explanations (also referred to herein as “reasons”) for navigational recommendations are disclosed. Composing a human-readable explanation includes individually selecting words or phrases that are then analyzed, combined, rearranged, modified, or removed to generate the human-readable explanation for a navigational recommendation. A decoder trains a machine learning model to generate the human-readable reasons for the navigational recommendations based on (1) historical recommendation vectors, and (2) historical human-readable reasons associated with the recommendation vectors. The system generates a dictionary of human-readable reasons for recommendations, with each entry of the dictionary including: (1) a recommendation identifier (ID) associated with a recommended navigational target, (2) a reason identifier (ID) associated with a particular reason for the recommendation, and (3) a human-readable reason associated with the reason ID.

Abstract: The present disclosure relates to systems and methods that use an artificial intelligence (AI) model to generate outputs that can be evaluated to predict which logged entry items match entry request record line items of an entry request record. Additionally, the present disclosure relates to systems and methods for intelligently detecting anomalies within data sets.

Abstract: Techniques for managing real-time business event analysis and monitoring are provided. In some examples, a logical request may be received at a business intelligence server configured with a continuous query service. Additionally, the continuous query service may be configured to translate the logical request into a physical request. Further, business intelligence information may be provided based at least in part on querying against an event data stream with a continuous query configured based at least in part on the physical request translated by the continuous query service.

Abstract: A new and improved onboarding process for configuring and activating an email message receiving system (EMRS) to accept and process email messages on behalf of a customer is disclosed. The process comprises generating a customer-account-specific Domain Name Server (DNS) name that identifies a customer of the EMRS and a mail server for accepting and processing email messages for a domain associated with the customer. The process includes providing the customer-account-specific DNS name to the customer, receiving a request to process email messages for the domain associated with the customer and identifying using the customer-account-specific DNS name, the mail server for receiving the email messages for the domain. The process further includes verifying, using the customer-account-specific DNS name, that the customer has permission for receiving the email messages for the domain. The process includes processing the email messages and delivering the messages to a destination system.

Abstract: Techniques are disclosed for utilizing control packets to manage flows by a smart network interface card (smartNIC). In one example, an accelerator of the smartNIC determines that a cache entry of a cache that is managed by the accelerator is a candidate for removal. The cache entry stores flow state of a particular flow. The accelerator generates a control packet that includes flow information of the particular flow that is formatted utilizing a particular header format, the flow information operable for generating a hash that indexes to the cache entry. The accelerator includes an instruction within the control packet that requests a programming data plane of the smartNIC to provide instructions for removing the cache entry from the cache. Upon receiving the control packet, the programming data plane generates and transmits a second instruction to the accelerator for removing the cache entry from the cache.

Abstract: Disclosed is an improved approach to implement a cloud access control mechanism. The mechanism provides customer control over access to cloud infrastructure by the cloud provider's operator employees, where the approach provides an override mechanisms for allowing operator access without additional customer approval based upon configured policies/rules.

Abstract: A hierarchy of containers is received at a user interface of a design time tool. The hierarchy of containers includes a container nested inside of one or more other containers. Each of the containers includes a different part of a web client application. The design time tool associates state with each of the containers in the hierarchy of containers. Lifecycles for each of the containers are received at the user interface. The design time tool creates computer executable instructions in the web client application that enable state for each of the containers to be created and destroyed according to the hierarchy based on the respective lifecycles of each of the containers. The design time tool provides each of the containers access to its own state and state of containers that include it based on respective lifecycle events associated with each of the containers.

Abstract: A method of evaluating a language model using negative data may include accessing a first language model that is trained using a first training corpus, and accessing a second language model. The second language model may be configured to generate outputs that are less grammatical than outputs generated by the first language model. The method may also include training the second language model using a second training corpus, and generating output text from the second language model. The method may further include testing the first language model using the output text from the second language model.