Microsoft Technology Licensing, LLC

Jay Goyal

Neil Sharman

Vinay Deshpande

Utkarsh Jain

Yinzhe Yu

Vibhaakar Sharma

Daniel Yuan

Gaurav Sareen

Walter Sun

Jingfeng Li

Lin Guo

Hugh Evan Williams

Gopi Prashanth Gopal

Junaid Ahmed

Sanaz Ahari

Nitin Agrawal

Abhishek Kumar Singh

Milad Shokouhi

Derrick Leslie Connell

Pratibha Permandla

Andrew Peter Oakley

Christopher A Meyers

Nicholas Craswell

Oren Firestein

Thomas Arthur Ledbetter

Brahm Kiran Singh

Methods and systems for quickly serving documents are provided. Documents may be served to users, for example, in response to search query inputs. Documents may be individually communicated to a document server prior to batching the documents. By individually communicating documents to document servers, the document experiences sub-second latency before it is available to a user. The documents may also be modified individually such that real-time serving is not interrupted.

Receiving individual documents to serve

Methods, systems, and computer-storage media for improving the freshness, or the apparent freshness, of search results are described. In an embodiment, the first portion of search results presented on a search results page are based on responsiveness to the search query and a second portion of results describe only recently published documents that are responsive to the search query. In an embodiment, a more recent version of the document, which is not directly used to determine responsiveness, is used to build the caption for a search result. Another way to make search results appear fresh is to include a publication time within the search result caption. In one embodiment, the publication time is generated by calculating a point in time between when a document is first added to a search index and the previous time the search engine visited the site where the document was found.

Enhancing freshness of search results

Methods and systems for quickly serving documents are provided. Documents may be served to users, for example, in response to search query inputs. Documents may be individually communicated to a document server individually prior to batching the documents. In such a real-time serving system, serving components may fail. To ensure real-time serving despite the failure, spares are utilized to replace the failing serving components such that the spare can immediately begin receiving documents. The spare can also be synchronized with other serving components to obtain the memory of the failing serving component prior to the failure.

Recovery of a document serving environment

Systems and methods for applying user behavior data to improve search query result ranking are provided. Upon receiving an update file indicating that recent, significant user behavior data is available for a document associated with an inverted index, the update file is published periodically and frequently to an index server. After filtering out the relevant update information from the update file, the index server extracts identifiers of the documents having the associated user behavior data. The update file and the identifier of the documents are utilized to update an in-memory index containing representations of metadata indicative of the user behavior. The in-memory index is continuously updated and utilized to serve search query results in response to user search queries. Search query results from the in-memory index are ranked using the user behavior data prior to serving. Thus, results associated with recent, significant user-behavior metadata receive prominent placement on the search results page.

Using behavior data to quickly improve search ranking

Methods, systems, and media are provided for an optimized search engine index. The optimized index is formed by merging small lower level indexes of fresh documents together into a hierarchical cluster of multiple higher level indexes. The optimized index of fresh documents is formed via a single threaded process, while a fresh index serving platform concurrently serves fresh queries. The hierarchy of higher level indexes is formed by merging lower and/or higher level indexes with similar expiration times together. Therefore, as some indexes expire, the remaining un-expired indexes can be re-used and merged with new incoming indexes. The single threaded process provides fast serving of fresh documents, while also providing time to integrate the fresh indexes into a long term primary search engine index, prior to expiring.

Hierarchical merging for optimized index

Methods, systems, and media are provided for identifying and clustering queries that are rising in popularity. Resultant clustered queries can be compared to other stored queries using textual and temporal correlations. Fresh indices containing information and results from recently crawled content sources are searched to obtain the most recent query activity. Historical indices are also searched to obtain temporally correlated information and results that match the clustered query stream. A weighted average acceleration of a spike can be calculated to distinguish between a legitimate spike and a non-legitimate spike. Legitimate clusters are combined with other stored clusters and presented as grouped content results to a user output device.

Detecting spiking queries

Growing community of inventors

Bellevue, WA, United States of America

Jay Goyal