Microsoft Technology Licensing, LLC

Thomas Minka

John Michael Winn

John Guiver

Alexander Armin Spengler

Gang Wang

Zheng Chen

Dmitriy Meyerzon

Ralf Herbrich

Nikita Voronkov

Weizhu Chen

Michael James Taylor

Matteo Venanzi

Thore K H Graepel

Eldar Akchurin

Samuel Alexander Webster

Vladimir V Gvozdev

Stephen Robertson

Anthony Arnold Wieser

Alexandre De Baynast

Ivan Korostelev

Pavel Myshkov

Yordan Kirilov Zaykov

Zhikai Fan

Vitaly V Filimonov

Elena Pochernina

Dmitry Starostin

Marius Bunescu

Panagiotis Periorellis

Aleksandr Klimov

Sanil Rajput

In various examples there is a computer-implemented method of database construction. The method comprises storing a knowledge graph comprising nodes connected by edges, each node representing a topic. Accessing a topic type hierarchy comprising a plurality of types of topics, the topic type hierarchy having been computed from a corpus of text documents. One or more text documents are accessed and the method involves labelling a plurality of the nodes with one or more labels, each label denoting a topic type from the topic type hierarchy, by, using a deep language model; or for an individual one of the nodes representing a given topic, searching the accessed text documents for matches to at least one template, the template being a sequence of words and containing the given topic and a placeholder for a topic type; and storing the knowledge graph comprising the plurality of labelled nodes.

Knowledge base with type discovery

Examining time series sequences representing performance counters from executing programs can provide significant clues about potential malfunctions, busy periods in terms of traffic on networks, intensive processing cycles and so on. An unsupervised anomaly detector can detect anomalies for any time series. A combination of known techniques from statistics, signal processing and machine learning can be used to identify outliers on unsupervised data, and to capture anomalies like edge detection, spike detection, and pattern error anomalies. Boolean and probabilistic results concerning whether an anomaly was detected can be provided.

Unsupervised anomaly detection for arbitrary time series

Probabilistic parsing is described for calculating information about the structure of text and other ordered sequences of items to enable downstream systems such as machine translation systems, information retrieval systems, document classification systems and others to use the structure information. In various embodiments, a parsing inference component comprises inference algorithm(s) compiled from a probabilistic program which defines a stochastic process for generating text or other ordered sequences of items. In examples, the parsing inference component receives one or more observations or examples of text that are compatible with the stochastic process defined by the probabilistic program. The parsing inference component may apply the inference algorithms to the text to update one or more probability distributions over strings or other values relevant to the parse. In some examples, the parsing inference component uses the inference results to complete partial examples to assist a user with information retrieval tasks.

Probabilistic parsing

A 'General Click Model' (GCM) is constructed using a Bayesian network that is inherently capable of modeling 'tail queries' by building the model on multiple attribute values that are shared across queries. More specifically, the GCM learns and predicts user click behavior towards URLs displayed on a query results page returned by a search engine. Unlike conventional click modeling approaches that learn models based on individual queries, the GCM learns click models from multiple attributes, with the influence of different attribute values being measured by Bayesian inference. This provides an advantage in learning that enables the GCM to achieve improved generalization and results, especially for tail queries, than conventional click models. In addition, most conventional click models consider only position and the identity of URLs when learning the model. In contrast, the GCM considers more session-specific attributes in making a final prediction for anticipated or expected user click behaviors.

Click modeling for URL placements in query response pages

There is a desire to provide a way to determine relative skills of players of games such as computer games, chess, tennis and any other suitable type of game. Our earlier Bayesian Scoring system is implemented in Xbox Live (trade mark) and is currently commercially available under the trade name TrueSkill (trade mark). Here we build on our earlier work and use a new method of computation to enable processing times to be significantly reduced. Message passing techniques are adapted to enable computation of updated skill beliefs to be obtained quickly even in the case of multiple teams of multiple players.

Determining relative skills of players

A compiler for probabilistic programs is described. The inputs to the compiler are a definition of a model and a set of inference queries. The model definition is written as a probabilistic program which describes a system of interest. The compiler transforms statements in the probabilistic program to generate source code which performs the specified queries on the model. The source code may subsequently be compiled into a compiled algorithm and executed using data about the system. The execution of the compiled algorithm can be repeated with different data or parameter settings without requiring any recompiling of the algorithm.

Compiler for probabilistic programs

Methods to enable optimization of discontinuous rank metrics are described. The search scores associated with a number of search objects are written as score distributions and these are converted into rank distributions for each object in an iterative process. Each object is selected in turn and the score distribution of the selected object is compared to the score distributions of each other object in turn to generate a probability that the selected object is ranked in a particular position. For example, with three documents the rank distribution may give a 20% probability that a document is ranked first, a 60% probability that the document is ranked second and a 20% probability that the document is ranked third. In some embodiments, the rank distributions may then be used in the optimization of discontinuous rank metrics.

Optimization of discontinuous rank metrics

Growing community of inventors

Cambridge, United Kingdom

Thomas Minka