Yahoo! Inc.

Nuance Communications, Inc.

Oath Inc.

Excalibur Ip, LLC

Adwait Ratnaparkhi

Benjamin Douglas

William Jarrold

Deepak Ramachandran

Abraham Hossain Bagherjeiran

Andrew Hatch

Suju Rajan

Dragomir Dimitrov Yankov

Peter Frederick Patel-Schneider

Peter Zei-Chan Yeh

Peter Yeh

Rajesh Girish Parekh

Scott J Gaffney

Rohit Chandra

Ming M Yang

Ben M Shahshahani

Parikshit Shah

Renjie Tang

Amruta Sadanand Joshi

Ralesh Parekh

Dun Liu

Sachidanand Alle

Wanlin Pang

Zengvan Zhang

In some natural language understanding (NLU) applications, results may not be tailored to the user's query. In an embodiment of the present invention, a method includes tagging elements of automated speech recognition (ASR) data based on an ontology stored in a memory. The method further includes indexing tagged elements to an entity of the ontology. The method further includes generating a logical form of the ASR data based on the tagged elements and the indexed entities. The method further includes mapping the logical form to a query to a respective corresponding database stored in the memory. The method further includes issuing the query to the respective corresponding databases. The method further includes presenting results of the query to the user via a display or a voice response system.

System and method for natural language driven search and discovery in large data sources

A system and method enable exploration of cold ads in an online information system. Cold ads are ads that are new to the system and do not yet have reliable click through rate (CTR) estimates or click probabilities. The method and system selectively boost the click probabilities of cold ads to increase their likelihood of participating in and winning an auction so that the cold ad get more impressions, but without adversely affecting revenue in a production environment when cold ads are introduced.

Exploration for search advertising

Presenting natural-language-understanding (NLU) results can include redundancies and awkward sentence structures. In an embodiment of the present invention, a method includes, responsive to receiving a result to a NLU query, loading a matching template of a plurality of templates stored in a memory. Each template has mask fields associated with at least one property. The method compares the properties of the mask fields of each of the templates to properties of the query and properties of the result, and selects the matching template. The method further completes the matching template by inserting fields of the result into corresponding mask fields of the matching template. The method may further suppress certain mask fields of the matching template to increase brevity and improve the naturalness of the response when appropriate based on the results of the NLU query. The method further presents the completed matching template to a user via a display.

Aspects discussed herein present a solution for utilizing large-scale knowledge graphs for inference at scale and generating explanations for the conclusions. In some embodiments, aspects discussed herein learn inference paths from a knowledge graph and determine a confidence score for each inference path. Aspects discussed herein may apply the inference paths to the knowledge graph to improve database lookup, keyword searches, inferences, etc. Aspects discussed herein may generate a natural language explanation for each conclusion or result from one or more inference paths that led to that conclusion or result. Aspects discussed herein may present the best conclusions or results to the user based on selection strategies. The presented results or conclusions may include generated natural language explanations rather than links to documents with word occurrences highlighted.

Utilizing large-scale knowledge graphs to support inference at scale and explanation generation

A method is provided for representing and updating the state of a dialog involving a series of queries and commands to an artificial intelligence system. Each statement within the dialogue may be modeled as a relational tree spanning nodes corresponding to named entities within the statement. A data structure may be used to store each of these trees and to modify them as the dialog progresses. A subsequent statement in the dialog may be parsed and its contents used to update an ongoing search initiated within that dialog. Statements may be used for the update process despite being fragmentary or not corresponding to any predetermined grammar. An algorithm is disclosed for updating the trees within the data structure after a new statement is parsed.

State tracking over machine-learned relational trees in a dialog system

A method and apparatus for finding predictive cross-category search queries for behavioral targeting in a networked online display advertising system. The methods include aggregating a training model dataset, the training model dataset comprising a history of clicks corresponding to historical advertisements. The training model dataset also contains plurality of targeting categories related to the history of clicks. Various techniques are disclosed for selecting a plurality of features from the training model dataset and calculating a click probability for a subject advertisement to be clicked by a user from a page, the calculating operations using features of the page that is to be presented to the user. Embodiments include mapping a particular query to one of the targeting categories and then presenting the subject advertisement selected on the basis of the value of the click probability. Normalization scales down the value of the click probabilities to filter out false positive categories.

Finding predictive cross-category search queries for behavioral targeting

A method for adaptive display of internet advertisements to look-alike users using a desired user profile dataset as a seed to machine learning modules. Upon availability of a desired user profile, that user profile is mapped other look-alike users (from a larger database of users). The method proceeds to normalize the desired user profile object, proceeds to normalize known user profile objects, then seeding a machine-learning training model with the normalized desired user profile object. A scoring engine uses the normalized user profiles for matching based on extracted features (i.e. extracted from the normalized user profile objects). Once look-alike users have been identified, the internet display system may serve advertisements to the look-alike users, and analyze look-alike users' behaviors for storing the predicted similar user profile objects into the desired user profile object dataset, thus adapting to changing user behavior.

Adaptive targeting for finding look-alike users

Disclosed are apparatus and methods for facilitating contextual selection of advertisements for displaying online via a computer network. In general, user features in the form of text are provided in conjunction with web page content for contextual advertisement matching. In one embodiment, a request for an advertisement to be displayed in a current web page that has been requested by a current user is received. The current user is associated with one or more current user characteristics from a plurality of different user characteristics, and the current web page has an associated content. A mapping model and the one or more current user characteristics are used to obtain a plurality of user-relevant terms for each of the one or more current user characteristics. A combination of the content of the current web page and obtained user-relevant terms are provided for selecting an advertisement for displaying with the current web page based on such combination.

Contextual advertising with user features

Methods, systems, and computer programs for categorizing the sensitivity of web pages are presented. In one method, a space of sensitive pages is identified based on the sensitivity categorization of a first plurality of web pages and a second plurality of web pages. The first plurality of web pages is obtained by performing search queries using known sensitive words, and the second plurality of web pages includes randomly selected web pages. Additionally, the method identifies a third plurality of web pages that includes web pages on or near the boundary between the space of sensitive pages and the space of non-sensitive pages. The space of sensitive pages is then redefined based on the sensitivity categorization of the first, second, and third pluralities of web pages. Once the space of sensitive pages is defined, the method is used to determine that a given web page is sensitive when the given web page is in the space of sensitive pages. Web pages are included in a marketing operation when the web pages are not sensitive.

Sensitivity categorization of web pages

An online advertising selects online advertisements for display on a network location taking into account a probability that a candidate online advertisement will receive a click on a particular website. The system may determine a network location identity of the network location and transform a set of advertisements into a set of ranked advertisements. The system may determine an advertisement rank of a first advertisement among the set of ranked advertisements. The system then may generate a click probability value. The click probability value may reflect a click probability of the first advertisement by dividing an exponent of a weighted sum of the network location identity and the advertisement rank by one plus the exponent of the weighted sum of the network location identity and the advertisement rank.

Clickable terms for contextual advertising

A system for training classifiers in multiple categories through an active learning system, including a computer having a memory and a processor, the processor programmed to: train an initial set of m binary one-versus-all classifiers, one for each category in a taxonomy, on a labeled dataset of examples stored in a database coupled with the computer; uniformly sample up to a predetermined large number of examples from a second, larger dataset of unlabeled examples stored in a database coupled with the computer; order the sampled unlabeled examples in order of informativeness for each classifier; determine a minimum subset of the unlabeled examples that are most informative for a maximum number of the classifiers to form an active set for learning; and use editorially-labeled versions of the examples of the active set to re-train the classifiers, thereby improving the accuracy of at least some of the classifiers.

Growing community of inventors

San Jose, CA, United States of America

Adwait Ratnaparkhi