Oracle International Corporation

Oracle America, Inc.

Sun Microsystems, Inc.

Stephen Joseph Green

Paul B Lamere

Jeffrey H Alexander

Michael Louis Wick

Jeffrey L Alexander

Jean-Baptiste Frederic George Tristan

Douglas Eck

Karl R Haberl

Guy Lewis Steele, Jr

Jean-Baptiste Tristan

Swetasudha Panda

Philip Ogren

Ariel Gedaliah Kobren

Manzil Zaheer

François Maillet

Bernard Horan

Debiasing pre-trained sentence encoders with probabilistic dropouts may be performed by various systems, services, or applications. A sentence may be received, where the words of the sentence may be provided as tokens to an encoder of a machine learning model. A token-wise correlation using semantic orientation may be determined to determine a bias score for the tokens in the input sentence. A probability of dropout that for tokens in the input sentence may be determined from the bias scores. The machine learning model may be trained or tuned based on the probabilities of dropout for the tokens in the input sentence.

Debiasing pre-trained sentence encoders with probabilistic dropouts

A scalable hierarchical coreference method that employs a homomorphic compression scheme that supports addition and partial subtraction to more efficiently represent the data and the evolving intermediate results of probabilistic inference. The method may encode the features underlying conditional random field models of coreference resolution so that cosine similarities can be efficiently computed. The method may be applied to compressing features and intermediate inference results for conditional random fields. The method may allow compressed representations to be added and subtracted in a way that preserves the cosine similarities.

Systems and methods for scalable hierarchical coreference

Embodiments perform string similarity analysis by receiving candidate strings of a collection to be searched and transforming each candidate string into one or more features. Embodiments generate a feature index that maps each of the features to one or more candidate strings which include the features, and transform the feature index into a low-memory index by byte encoding a sorted integer array into an encoded byte array. The transforming into a low-memory index further includes hashing each feature to an integer value and storing using a tightly-packed map a hashed value to feature identification ('ID') mapping, using a first tightly-packed list to map each of the feature IDs to one or more candidate strings which include the features, and using a second tightly-packed list to store the feature IDs.

Memory reduced string similarity analysis

Herein is described an unsupervised learning method to discover topics and reduce the dimensionality of documents by designing and simulating a stochastic cellular automaton. A key formula that appears in many inference methods for LDA is used as the local update rule of the cellular automaton. Approximate counters may be used to represent counter values being tracked by the inference algorithms. Also, sparsity may be used to reduce the amount of computation needed for sampling a topic for particular words in the corpus being analyzed.

Learning topics by simulation of a stochastic cellular automaton

Meaningful words or phrases may be extracted from the information and used as tags. Weights may be determined for the tags, and tag clouds may be generated for the items. The tag clouds may be stored to a data store. Information specifying a tag cloud may be received. Recommended items for which the tag clouds most closely match the specified tag cloud may be identified. Standard vector space distance calculations, for example the cosine distance between the tag clouds, may be used to determine cloud similarity. The results may be filtered to optimize relevance, novelty and familiarity in accordance with preferences of the user. The recommended items may be displayed to a user interface. Users may interact with the user interface to steer the recommendations towards more relevant content.

Method and apparatus for generating recommendations from descriptive information

A system performs search and retrieval. The system monitors one or more user interface ('UI') fields configured to receive text input in a UI. The system determines that the one or more UI fields are being used to enter a textual description, and performs a search on a knowledge base based on document similarity to identify documents that are similar to a portion of the textual description that has already been entered in the one or more UI fields. The system then provides one or more of the documents in a UI field of the UI, and repeats the monitoring, the determining, the performing, and the providing.

Knowledge base search and retrieval based on document similarity

Systems, methods, and other embodiments associated with generating suggested shortcuts are described. In one embodiment, a method includes identifying a user that accesses an intranet portal and tracking navigation paths taken by the user within the intranet portal. The navigation paths are stored in a user profile associated with the user, where the navigation paths indicate a sequence of visited locations and a destination location for the sequence. When the user visits a location within the intranet portal, a list of suggested destination locations are generated as navigation shortcuts to allow the user to directly access a desired destination.

Personalized navigation shortcuts

Method and apparatus for transparent, steerable recommendations. A steerable recommender uses tag clouds including descriptive tags and associated weights to generate recommendations. Users may dynamically interact with the recommender via a user interface to steer the recommendations. A tag cloud for an item is displayed, items for which associated tag clouds are most similar to the displayed tag cloud are identified, and the items are displayed as recommendations. The strength of similarity of the items to the displayed tag cloud may be displayed. The user may modify a tag cloud, for example by changing the weight of a tag or by adding or removing a tag, and the recommendations may be automatically updated to reflect the modification. A recommended item may be selected to display the tag cloud corresponding to the item. A user may select a user interface element to request information on why a particular item was recommended.

Method and apparatus for user-steerable recommendations

A method for accessing a file system including computing a first numerical similarity score for a first stored document and a second numerical similarity score for a second stored document by comparing a plurality of weighted active terms with a plurality of weighted indexed terms, determining a document order of the first stored document followed by the second stored document based on the first numerical similarity score exceeding the second numerical similarity score, generating a list of similar documents including the first stored document followed by the second stored document based on the document order, and displaying, in a file system interface and on the computer display, the list of similar documents while an active document is open in an active document interface.

Method and system for accessing a file system

A resource analyzer selects a resource (e.g., document) from a grouping of resources. The grouping of resources can be any type of social tagging system used for information retrieval. The selected resource has an assigned uncontrolled tag and an assigned controlled tag. The controlled tag is a term derived from a controlled vocabulary of terms. Having selected the resource for analyzing, the resource analyzer identifies a first set of resources in the grouping of resources having also been assigned a same value as the uncontrolled tag as the selected resource. Similarly, the resource analyzer identifies a second set of resources in the grouping of resources having also been assigned a same value as the controlled tag. With this information, the resource analyzer then produces a comparison result indicative of a similarity between the first set of resources and the second set of resources.

Relating similar terms for information retrieval

A method for searching and resource discovery in a distributed enterprise (DE), the method including: generating a first classifier for a first repository in the DE; generating a second classifier for a second repository in the DE, where the second classifier has a vector element identifying a location of the second repository; submitting a copy of the second classifier to a web server of the first repository; obtaining a resource at the first repository after submitting the copy of the second classifier; matching the resource to the copy of a second classifier; sending a copy of the resource to the second repository using the vector element; and storing the copy of the resource in the second repository.

Method and apparatus for searching and resource discovery in a distributed enterprise system

A method for accessing a file system that includes creating a document index for an active document, identifying a similarity score for each of a plurality of stored documents to generate a list of similar documents, wherein the similarity score is computed by comparing the document index for the active document with a search index for the plurality of stored documents, and displaying the list of similar documents in a file system interface.

A media playlist is generated from a set of two or more seed media selections as well as the desired number of selections in the playlist. The seed selections represent waypoints in 'media space'. A playlist is generated by finding a path of 'stepping stone' selections between the waypoints, with the constraints that no selection is presented more than once and the maximum distance between stepping stones is minimized. In one embodiment, an &#x201c;endless&#x201d; playlist by adding a final path from the last playlist selection back to a selection that is nearest the first selection, but has not been presented yet. In still another embodiment, a separate set of &#x201c;avoid&#x201d; selections is maintained. When generating the playlist, given a set of paths that meet the aforementioned path generation constraints, paths will be selected that maximize the distance from the set of &#x201c;avoid&#x201d; selections.

Growing community of inventors

Burlington, MA, United States of America

Stephen Joseph Green