Ebay, Incorporated

Paypal, Inc.

Google Inc.

Raghav Gupta

Jean-Michel Leon

Neelakantan Sundaresan

Alvaro Bolivar

Aditya Arora

Benjamin David Foster

Eric Noel Billingsley

Seema Shah

Srikanth Subramaniam

Vijay Raman

Suhail Ansari

Adam Nash

Xavier Stonestreet

Ken Sun

Roopnath Grandhi

Alec Reitter

Randall Scott Shoup

Alan Lewis

John Tapley

Eric J Farraro

Daniel Graham Douglas

Daniel Tsun Kao

James Everett Lloyd

Sichun Xu

Gregory Lincoln Stone

Todd Cameron Pringle

Yang Song

Richard Dean Henderson

Shuhui Chen

Shen Huang

Xiaoyuan Wu

Dengyong Zhou

Abhinav Kumar Rastogi

Sanqiang Zhao

Barbara Chang

Barb Chang

Denis Golovnya

Jeffrey Olson

Dilek Zeynep Hakkani-Tur

Louis Marcel Gino Monier

Preston Smalley

Chi-Hsien Chiu

Anselm Paul Baird-Smith

Pranav Khaitan

Vijay Kumar

Raffi Tutundjian

Xiaoxue Zang

Srinivas Kumar Sunkara

Thomas Nelson

Nicholas David Posner

Ramana M Idury

Richard Douglas Henderson

Provided is a knowledge distillation technique for training a student language model that, relative to a larger teacher language model, has a significantly smaller vocabulary, lower embedding dimensions, and/or hidden state dimensions. Specifically, aspects of the present disclosure are directed to a dual-training mechanism that trains the teacher and student language models simultaneously to obtain optimal word embeddings for the student vocabulary. In some implementations, this approach can be combined with learning shared projection matrices that transfer layer-wise knowledge from the teacher language model to the student language model. Example experimental results have also demonstrated higher compression efficiency and accuracy when compared with other state-of-the-art compression techniques, including the ability to compress the BERTmodel by more than 60×, with only a minor drop in downstream task metrics, resulting in a language model with a footprint of under 7 MB.

Extreme language model compression with optimal sub-words and shared projections

Methods, systems, and apparatus, including computer programs encoded on computer storage media for dialogue systems. A transcription of a user utterance is obtained. The transcription of the utterance is tokenized to identify multiple tokens for the utterance. Token-level utterance encodings corresponding to different tokens of the transcription are generated. A system action encoding from data indicating system actions previously performed by the dialogue system are generated. A dialogue context vector based on the utterance encoding and the system action encoding are generated. The token-level utterance encodings, the system action encoding, and the dialogue context vector are processed using a slot tagger to produce token-level output vectors. A limited set of candidate token classifications for the tokens of the user utterance are determined based on the token-level utterance encodings. A response for output is provided in response to the user utterance.

Language understanding and dialogue state tracking in dialogue systems

Provided is a knowledge distillation technique for training a student language model that, relative to a larger teacher language model, has a significantly smaller vocabulary, lower embedding dimensions, and/or hidden state dimensions. Specifically, aspects of the present disclosure are directed to a dual-training mechanism that trains the teacher and student language models simultaneously to obtain optimal word embeddings for the student vocabulary. In some implementations, this approach can be combined with learning shared projection matrices that transfer layer-wise knowledge from the teacher language model to the student language model. Example experimental results have also demonstrated higher compression efficiency and accuracy when compared with other state-of-the-art compression techniques, including the ability to compress the BERTmodel by more than 60&#xd7;, with only a minor drop in downstream task metrics, resulting in a language model with a footprint of under 7 MB.

A system to provide image processing services responsive to requests including image data includes a system layer that forwards a request to an image application processing interface. Image processing provides an image comparison, barcode recognition, and optical character recognition. The image processing compares the image data to products in a database in order to identify a matching product. The system layer receives the matching information and forwards to a user.

Method and apparatus for image recognition services

In one embodiment, a method is illustrated as including defining a set of perspective objects capable of being placed onto a modified web page, monitoring parameters of a web page, the parameters including a number of times a current object is executed on the web page, using an Artificial Intelligence (AI) algorithm to determine a perspective object with a preferred Return On Investment (ROI), and selecting the perspective object to be placed onto the modified web page.

Economic optimization for product search relevancy

In an example embodiment, a method of identifying an item depicted in an image is provided. In this method, the image depicting the item is accessed; in addition, other images and their item identifiers are also accessed. A match of the image with one of the other images is identified. With a match, the image is then associated with an item identifier of the matched image.

Identification of items depicted in images

Generally, the present disclosure is directed to systems and methods for performing task-oriented response generation that can provide advantages for artificial intelligence systems or other computing systems that include natural language processing for interpreting user input. Example implementations can process natural language descriptions of various services that can be accessed by the system. In response to a natural language input, systems can identify relevant values for executing one of the service(s), based in part on comparing embedded representations of the natural language input and the natural language description using a machine learned model.

Schema-guided response generation

A system to process navigation information includes a receiving module to receive a data request from a network entity. The data request includes a network entity identifier that uniquely identifies the network entity. The system further includes an application to generate a network event based on the data request, the network event including the navigation information. Finally, the system includes a publisher module to communicate the network event to a first server that is distinguished from a first plurality of servers based on the network entity identifier.

Method and system to process navigation information

In an example embodiment, a request that includes a first image is received. A second image and a description are accessed from an item listing. An item identifier that corresponds to the second image is parsed from the description. A first edge in the first image and a second edge in a second image are detected. A match between the first image and the second image is determined based on the detection. The first image is associated with the item identifier. Item information corresponding to the item identifier is accessed from web pages. The item information is then transmitted.

A computer-implemented system and method for keyword extraction are disclosed. The system in an example embodiment includes a keyword extraction component to extract relevant keywords from content of a web page, to identify items relevant to the extracted keywords, and to rank the relevant items.

System and method for keyword extraction

Systems and methods to communicate information are described. The system receives a query and retrieves a set of data items from a database based on the query. The system generates a distribution based on a count of data items from the set of data items that were retrieved for each domain in a set of domains and a count of requests to view data items from another set of data items for each domain in the set of domains. The system identifies whether the distribution exhibits a type of a distribution by comparing the distribution with at least one of a peak distribution and a hills distribution. Finally, the system generates interface information based on the type of distribution.

Methods and systems to communicate information

A system to process navigation information includes at least one hardware processor and an information storage and retrieval platform executed by the at least one hardware processor. The information storage and retrieval platform is configured to receive a data request from a client, the data request including a client identifier that identifies the client, generate a network event in response to the data request received from the client, and generate, by the at least one hardware processor, a hash value from the client identifier. The information storage and retrieval platform is also configured to identify an address based on the hash value, determine a server by a lookup of the server from among a plurality of servers, the lookup being based on the identified address, and communicate the network event to the server.

Growing community of inventors

Sunnyvale, CA, United States of America

Raghav Gupta