Microsoft Technology Licensing, LLC

Chao Liu

Yi-Min Wang

Chao Zhou

Li-wei He

Hung-Chih Yang

Xiaoxin Yin

Jinliang Fan

Wenzhao Tan

Fan Guo

Chao Zhaou

A system is described herein that includes a preference deriver component that receives a predefined preference rule that indicates a hierarchy pertaining to entities belonging to a domain, wherein each of the entities has attributes and values for such attributes corresponding thereto, and wherein the preference deriver component outputs preferences between various subsets of entities based at least in part upon the preference rule. The system also includes a learning component that learns a computer-implemented ranker component that is configured to rank the entities belonging to the domain, wherein the learning component learns the computer-implemented ranker based at least in part upon the preferences between the various subsets of the entities output by the preference deriver component.

Learning a ranker to rank entities with automatically derived domain-specific preferences

The claimed subject matter provides a system or method for web crawling hidden files. An example method includes loading a web page with a browser agent, and executing any dynamic elements hosted on the web page using the browser agent to insert pre-determined values. A list of form controls may be retrieved from the web page using the browser agent, and the controls may be analyzed using a driver component. Form control values may be sent from the driver component to the browser agent, and an event may be submitted to the web page by the browser agent or scripted content may be run to trigger operations on the web page corresponding to the form control values. A URL may be generated for various form control values using a generalizer.

Interactive web crawler

Technologies pertaining to fact lookup queries are described herein. A relational database is automatically built by extracting attribute identities and attribute values from a one dimensional table, wherein the one dimensional table does not include an entity identity that corresponds to the attribute identity and the attribute value. The entity identity is inferred, and the attribute value is indexed in a relational database by the entity identity and the attribute identity. When a query is issued by a user that includes the entity identity and the attribute identity, the corresponding attribute value in the relational database is returned to the user.

Automated database generation for answering fact lookup queries

The claimed subject matter provides a system or method for web crawling hidden files. An exemplary method includes loading a web page with a browser agent, and executing any dynamic elements hosted on the web page using the browser agent to insert pre-determined values. A list of form controls may be retrieved from the web page using the browser agent, and the controls may be analyzed using a driver component. Form control values may be sent from the driver component to the browser agent, and an event may be submitted to the web page by the browser agent or scripted content may be run to trigger operations on the web page corresponding to the form control values. A URL may be generated for various form control values using a generalizer.

Architecture that scales up the non-negative matrix factorization (NMF) technique to a distributed NMF (denoted DNMF) to handle large matrices, for example, on a web scale that can include millions and billions of data points. To analyze web-scale data, DNMF is applied through parallelism on distributed computer clusters, for example, with thousands of machines. In order to maximize the parallelism and data locality, matrices are partitioned in the short dimension. The probabilistic DNMF can employ not only Gaussian and Poisson NMF techniques, but also exponential NMF for modeling web dyadic data (e.g., dwell time of a user on browsed web pages).

Distributed non-negative matrix factorization

Techniques are described for generating a statistical model from observed click chains. The model can be used to compute a probability that a document is relevant to a given search query. With the model, a probability of a user examining a given document in a given search result conditionally depends on: a probability that a preceding document in the given search result is examined by a user viewing the given search result; a probability that the preceding document is clicked on by a user viewing the given search result, which conditionally depends directly on the probability that the preceding document is examined and on a probability of relevance of the preceding document.

Growing community of inventors

Redmond, WA, United States of America

Chao Liu