Microsoft Technology Licensing, LLC

Snap Inc.

Manish Malik

Qifa Ke

Doran Chakraborty

Jonathan R Tiao

Yu Shi

Xiaochuan Ni

Yiwei Ma

Jason Brewer

Andreas Bode

Rangan Majumder

Pushpraj Shukla

Raghavan Muthuregunathan

James Li

Lili Zhang

Mingtao Zhang

Chang Kuang Huang

Jinchao Ye

Hangqi Zhao

Yang Zhou

Yining Zhang

Ding Zhou

Jiarui Ren

Yinan Na

Dan Xie

Shuo Han

Techniques for creating an interest graph include obtaining content items from multiple content sources and applying tailored (e.g., source-specific) preprocessing to the content items based on their respective content source. Text is extracted and salient keywords and key phrases are identified using unsupervised machine learning models. The keywords and key phrases become nodes in an interest graph, each node comprising an embedding of a keyword or key phrase in a common embedding space, with edges representing semantic similarity based on embeddings or co-engagement patterns. The graph provides an expansive, granular, and dynamic taxonomy easily adaptable to emerging interests. The interest graph overcomes limitations of conventional taxonomies that lack depth, fail to capture niche interests, and cannot adapt to reflect evolving user preferences. The described techniques construct a rich interest graph from diverse content for improved content understanding.

Automatic techniques for constructing an evolving interest taxonomy from user-generated content

Systems and techniques for sequence to sequence conversational query understanding are described herein. A query may be received that includes multiple words. It may be identified that the query is to be reformulated based on an attention value for an attention word in the query. Relationships may be determined among words of the query and words in a previously submitted query and words in results from the previously submitted query. The query may be reformulated based on the relationships. The reformulated query may be employed to retrieve query results.

Sequence to sequence conversational query understanding

Representative embodiments disclose mechanisms to provide direct answers to a query submitted by a user. The mechanisms are tailored so that the answers presented have a high confidence of being correct. A plurality of document segments that are relevant to the query are selected. The selected segments are submitted to a trained machine reading comprehension model along with the query. The result is an extracted answer for one or more of the submitted segments. A subset of the extracted answers are clustered and an answer for each cluster having at least a threshold number of answers are selected as direct answers. The direct answers are presented in a format suitable to the number of selected direct answers.

Intelligent question answering using machine reading comprehension

Examples of the present disclosure describe systems and methods relating to generating a relevance score on a given natural language answer to a natural language query for ranking the answer among other answers for the query, while generating a summary passage and a likely query to the given passage. For instance, multi-layered, recurrent neural networks may be used to encode the query and the passage, along with a multi-layered neural network for information retrieval features, to generate a relevant score for the passage. A multi-layered, recurrent neural network with soft attention and sequence-to-sequence learning task may be used as a decoder to generate a summary passage. A common encoding neural network may be employed to encode the passage for the ranking and the summarizing, in order to present concise and accurate natural language answers to the query.

Method and system for ranking and summarizing natural language passages

Non-limiting examples of the present disclosure describe a unified ranking model that may be used by a plurality of entry points to return ranked results in response to received query data. The unified ranking model is provided as a service for a plurality of entry points. A query is received from an entry point of the plurality of entry points. Results data for the query data is retrieved. A unified ranking model is executed to rank the results data. Execution of the unified ranking model manipulates feature data of the unified ranking model based on user context signals associated with the received query data and acquired result retrieval signals corresponding with the retrieved results data. Execution of the unified ranking model generates ranked result data. Ranked results data is returned to the processing device corresponding with the entry point. Other examples are also described.

Single unified ranker

Aspects of the technology described herein increase the efficiency of a search session by determining whether fresh content is likely to be responsive to the user's query. Whether fresh content is likely to be responsive to a specific query is determined by retrieving social media posts that are responsive to the query. The social media posts are evaluated for virality, which is the tendency of a social media post to be circulated rapidly and widely from one Internet user to another. The virality of a social media post can be determined by comparing a number of times the social media post has been re-communicated by individual users. Queries that return viral social media posts may be classified as seeking fresh content.

Efficient retrieval of fresh internet content

Computer-readable media, computerized methods, and graphical user interfaces (GUI's) for interactively presenting features within the panes of a display area rendered in a GUI workspace are provided. Initially, a computing device memory is inspected to discover a profile that accommodates a user-visual pattern. When the profile is not discovered, a default page layout is retrieved and communicated to the display device for presentation within the display area. If the profile is discovered, commonly viewed screen locations, based upon a frequency of eye fixation upon each of the screen locations, are ordered. Contextually relevant features are selected for presentation within the display area. The selected features are ranked based on predetermined criteria and each associated with the commonly viewed screen locations by comparing the ranking with the ordering. A page layout is constructed by spatially placing the features in proximity to the screen locations associated therewith, and is rendered within the display area.

Dynamic feature presentation based on vision detection

Growing community of inventors

Cupertino, CA, United States of America

Manish Malik