Abstract: The disclosed technology teaches a deep end-to-end speech recognition model, including using multi-objective learning criteria to train a deep end-to-end speech recognition model on training data comprising speech samples temporally labeled with ground truth transcriptions.

Abstract: Techniques for providing feed-based case management on an online social network. A user interface that includes a publisher and an information feed is provided. Information associated with a record is received and the record is updated in a database system. A feed item associated with the database system update is generated and provided in the information feed. The feed item includes one or more actionable selections providing a reference to the publisher. Selection of one of the one or more actionable selections causes the publisher to be operable to receive further information associated with the record. The record is capable of being updated based on the further information. In some implementations, the record is a case and the user interface serves as a single interface for all interactions required in solving the case.

Abstract: We introduce an attentional neural machine translation model for the task of machine translation that accomplishes the longstanding goal of natural language processing to take advantage of the hierarchical structure of language without a priori annotation. The model comprises a recurrent neural network grammar (RNNG) encoder with a novel attentional RNNG decoder and applies policy gradient reinforcement learning to induce unsupervised tree structures on both the source sequence and target sequence. When trained on character-level datasets with no explicit segmentation or parse annotation, the model learns a plausible segmentation and shallow parse, obtaining performance close to an attentional baseline.

Abstract: Disclosed are methods, apparatus, systems, and computer-readable storage media for recommending an event to a user. In some implementations, one or more servers receive information identifying a plurality of events. The one or more servers store data of the plurality of events in a first one or more data tables having an action field, an intent field, and a user field, and analyze the data of the first one or more data tables to generate one or more pairs, each pair including information identifying a set of events and a target event. The one or more servers may calculate a similarity score for each of the one or more pairs and store the respective similarity score in a second one or more data tables having a set field, a target event field, and a similarity score field.

Abstract: The technology disclosed presents a novel spatial attention model that uses current hidden state information of a decoder long short-term memory (LSTM) to guide attention and to extract spatial image features for use in image captioning. The technology disclosed also presents a novel adaptive attention model for image captioning that mixes visual information from a convolutional neural network (CNN) and linguistic information from an LSTM. At each timestep, the adaptive attention model automatically decides how heavily to rely on the image, as opposed to the linguistic model, to emit the next caption word. The technology disclosed further adds a new auxiliary sentinel gate to an LSTM architecture and produces a sentinel LSTM (Sn-LSTM). The sentinel gate produces a visual sentinel at each timestep, which is an additional representation, derived from the LSTM's memory, of long and short term visual and linguistic information.

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: Methods and systems are provided for implementing session synchronization to synchronize applications at different user systems (e.g., mobile device, laptop, desktop, tablet) of a particular user. As will be explained in greater detail below, this platform level solution that can maintain session information received any number of different user systems of the particular user and share this collective session information with each of the other user systems as session-synchronization information that can be used, at any user system of that user, to ensure that when that user switches from using one user system to another user system that that applications at the “new” user system being switched to are synchronized with any other applications that were most recently in use at all of the other user systems of that user.

Abstract: Disclosed are some implementations of systems, apparatus, methods and computer program products for publisher and share action integration in a user interface (UI) for automated messaging. In some implementations, information is displayed in a first component and in a second component in a UI on a display. The first component is a publisher configured for composing and sharing messages on different communication channels using different share actions. The second component includes links to data items. When one of the data items is selected, a set of share actions for sharing the selected item using the publisher can be identified and displayed. When one of the displayed share actions is selected, it can be determined that the selected share action is applicable to one of the communication channels. A reference to the selected item can be provided in a message being composed in the publisher for sharing on the one communication channel.

Abstract: The technology disclosed presents a novel spatial attention model that uses current hidden state information of a decoder long short-term memory (LSTM) to guide attention and to extract spatial image features for use in image captioning. The technology disclosed also presents a novel adaptive attention model for image captioning that mixes visual information from a convolutional neural network (CNN) and linguistic information from an LSTM. At each timestep, the adaptive attention model automatically decides how heavily to rely on the image, as opposed to the linguistic model, to emit the next caption word. The technology disclosed further adds a new auxiliary sentinel gate to an LSTM architecture and produces a sentinel LSTM (Sn-LSTM). The sentinel gate produces a visual sentinel at each timestep, which is an additional representation, derived from the LSTM's memory, of long and short term visual and linguistic information.

Abstract: A CDN traffic is optimized by a client-side system that maps the servers in the CDN system. Content requests from client devices for domain names are forwarded to servers in the CDN system that may be selected from the map to prevent a cache miss in the a server for a particular request for content.

Abstract: A document retrieval system tracks user selections of documents from query search results and uses the selections as proxies for manual user labeling of document relevance. The system trains a model representing the significance of different document features when calculating true document relevance for users. To factor in positional biases inherent in user selections in search results, the system learns positional bias values for different search result positions, such that the positional bias values are accounted for when computing document feature features that are used to compute true document relevance.

Abstract: Techniques and architectures to provide trace functionality. Trace record data is received from a plurality of client threads executed by one or more processors. The trace record data is stored in a plurality of chunks maintained in an in-use list. The in-use list has a chunk for individual use by the corresponding client threads. Chunks in the in-use list are moved to a completed queue when a chunk in the in-use list is substantially full. A chunk from a free list is placed in the in-use list to replace removed chunks. The chunks from the completed queue are stored in at least one memory device.

Abstract: A system for introducing behaviorally tested live content (18) into an electronic mail message comprising at least one dynamic live content area (1006) sent through an e-mail service provider system (16) comprises memory and a click manager (28). The memory 5 includes behaviorally tested live content comprising image data. The click manager (28) receives an indication of the opening of the message by a recipient, whereupon the click manager (28) accesses the memory to retrieve behaviorally tested live content (18) and send the retrieved behaviorally tested live content (18) for rendering in the dynamic live content area (1006) of the electronic mail message opened by the one of the plurality of 0 recipients.

Abstract: Some embodiments of the present invention include a method for merging records associated with objects in an object database and include generating, by a database system, a graph database corresponding to an object database, wherein objects in the object database are represented as nodes of graphs in the graph database, and wherein relationships among the objects in the object database are represented as edges of the graphs in the graph database; receiving, by the database system, information about a first object and information about a dimension based on a first relationship associated with the first object; and performing, by the database system, a merge operation using the information about the first object and the information about the dimension on the graph database instead of on the object database.

Abstract: Systems, methods, and computer-readable media for providing a Policy Enforcement as a Service (PEaaS) are described. A processor may, in response to identification of a suspension of user(s) for one of one or more services, generate a suspension value indicative of the suspension and transmit the suspension value to a corresponding one of third party platforms of the service(s), respectively. The suspension value usable by the corresponding third party platform to determine whether to deny request(s) from the user for the distributed service, or not. A service of the service(s) comprises a distributed service provided by a plurality of hosts. In response to the service corresponding to the suspension including the distributed service, the suspension value may be transmitted to the first host, and the suspension value may be propagated to the one or more second hosts, respectively. Other embodiments may be described and/or claimed.

Abstract: A system receives an association of first item with first system user, generates first hash value by applying first hash function associated with first system user to first item identifier associated with first item, and sets a bit corresponding to first hash value in array. The system receives an association of second item with second system user, generates second hash value by applying second hash function associated with second user to second item identifier associated with second item, and sets a bit corresponding to second hash value in array. The system receives a request to determine whether third item is associated with first system user, generates third hash value by applying first hash function to third item identifier associated with third item, and outputs message that third item is not associated with first user if a bit corresponding to third hash value is not set in array.

Abstract: Data related to a product is received from a first subscriber and stored into a first section of the on-demand database reserved for the first subscriber. The first subscriber then communicates a request to share the product data with one or more second subscribers. Once the first subscriber receives an acceptance from the one or more second subscribers, the product data is stored in a second section of the on-demand database reserved for the second subscriber. Changes made to the data associated with the product are synchronized between the two subscribers.

Abstract: The invention provides, in some aspects, digital data processing methods of generating digital content pieces (e.g., email messages or portions thereof) that are customized in accord with individual recipient behaviors. Such methods include the step of generating and digitally transmitting to a digital data devices of a recipient a digital content piece that (i) has a call to action to which the recipient can respond and (ii) that has a plurality of features selected so as to maximize a probability, P(b1, b2, . . . , bM, x1, x2, . . . , xM), that the recipient will respond to that call to action, where that probability is defined by the relation P(b1,b2, . . . , bM,x1,x2, . . . , xM)=exp(?j=1, . . . , Mbjxj)/(1+exp(?j=1, . . . , Mbjxj)) where x1, x2, . . . , xM are values for each of a plurality, M, of features characterizing the digital content piece and/or the recipient, b1, b2, . . . , bM are respective coefficients for each of the values x1, x2, . . . , xM.