Microsoft Technology Licensing, LLC

Guihong Cao

Zicheng Huang

Jun Liu

Mridu Baldevraj Narang

Jiayin Ge

Jianfeng Gao

Ahmed Hassan Mohamed

Chin-Yew Lin

Xiaodong Fan

Jian-Yun Nie

Fetiye Karabay

Eugene Joonsup So

Julian Manuel Varanda

Alyssa Nicole Ricken

Jianshu Ji

Jinxi Xu

Sharath Rao

Chao Gao

Yu Jia

Yufei Zhong

Innocent Mutabazi Ndimubanzi

Astha Suneja

Systems, methods, and computer-executable instructions for synthesizing facts from multiple sources. A user search string associated with a topic is received. Queries associated with the user search string are generated. Website content is queried, using the plurality of queries associated with the user search string, to generate results. The results are ranked based on a first feature. The results are filtered based on the first feature to generate intermediate results. The intermediate results are ranked based on a second feature. The second feature is most costly to calculate than the first feature. The filtered intermedia results are filtered based on the second feature to generate final results. Distinct facts from the final results are determined. The distinct facts include content from different URLs. The distinct facts associated with the topic are returned.

Synthesized facts from multiple sources

The present disclosure relates to systems and methods for automatically identifying additional content for a webpage. The systems and methods may create a side pane placed next to a webpage to supplement the webpage in a browser. The side pane may provide additional information or additional content to help users consume the webpage. The additional content may provide webpage insights and help the users further explore the webpage.

Automatic identification of additional content for webpages

The technology described herein makes improved use of limited screen space on a search results page by determining whether to present a question-and-answer experience and/or an entity details experience. This determination effects the amount of information presented and the format in which it is presented. In general, the question-and-answer experience provides less information and is more targeted to a question and query terms other than the entity. In contrast, the entity details experience provides more information about the entity that is not tailored to the query beyond the entity being included in the query. In one aspect, the determination of whether to show a question-and-answer experience and/or an entity details experience is based, at least in part, on an entity-details intent classification score ('intent classification score') generated by a machine classification system. The classification score may be processed in combination with other criteria to make a final determination.

Query entity-experience classification

A method of disambiguating user queries in a multi-turn dialogue include a set of user utterances includes using a predefined language model to recognize an ambiguous entity in an unresolved user utterance from the multi-turn dialogue. The method further includes outputting a clarifying question about the ambiguous entity, and receiving a clarifying user utterance. The method further includes identifying a disambiguating entity in the clarifying user utterance. The method further includes rewriting the unresolved user utterance as a rewritten utterance that replaces the ambiguous entity with the disambiguating entity, and outputting the rewritten utterance to one or more query answering machines.

Clarifying questions for rewriting ambiguous user utterance

A method of disambiguating user queries in a multi-turn dialogue including a set of user utterances. The method includes using a predefined language model to recognize an ambiguous entity in an unresolved user utterance from the multi-turn dialogue, and using the predefined language model to recognize entity constraints of the ambiguous entity. The method further includes, in a computer-accessible conversation history of the multi-turn dialogue, searching a set of previously-resolved entities for a candidate entity having entity properties with a highest confidence correspondence to the entity constraints of the ambiguous entity. The unresolved user utterance is rewritten as a rewritten utterance that replaces the ambiguous entity with the candidate entity. The rewritten utterance is output to one or more query answering machines.

Architecture for resolving ambiguous user utterance

A platform for generating query-answer pairs is described herein, wherein the platform is well-suited for generating query-answer pairs for a variety of topics. The platform comprises a topic-specific classifier that is configured to identify queries from a query log that are fact-seeking queries that belong to a topic. The platform further comprises a domain identifier component that identifies domains that are relevant to the topic, wherein the domains comprise pages. The platform further includes a passage extraction system that comprises a hierarchy of rankers, wherein the passage extraction system is configured to identify passages in the pages that belong to the domain that comprise answers to the fact-seeking queries that belong to the topic.

Computer-implemented platform for generating query-answer pairs

Methods, computer systems, and computer-storage media are provided for generating natural language outputs. Sets of triples may be used to map voice queries and answers to sentence structures that may be used as an output answer to the voice query. Sentence structures are only appropriate with certain sets of triples. One or more constraints may be associated with the set of triples to make sure sentence structures are only applied in correct situations. In order to be a valid sentence structure, each constraint associated therewith must be satisfied. If each constraint is satisfied, the sentence structure is valid and may be used as the format for an output answer. If each constraint is not satisfied, additional sentence structures associated with the set of triples may be evaluated until a valid sentence structure is identified. If no sentence structure is valid, no output is generated.

Generating natural language outputs

A probability distribution for a reference summary of a document is determined. The probability distribution for the reference summary is then used to generate a score for a machine-generated summary of the document.

Growing community of inventors

Sammamish, WA, United States of America

Guihong Cao