Xerox Corporation

Nidhi Singh

Nicolas Monet

Jean-Marc Coursimault

Herve Poirier

Frédéric Roulland

Herve Poirer

Methods and systems for dynamically training a web crawler. The web crawler maintains one or more categories each comprising a set of words. The method includes selecting at least one hyperlink in response to a query received from a user. The method further includes determining a hyperlink score for the at least one hyperlink based on a category score associated with each of one or more categories. The category score associated with each of the one or more categories is updated based at least in part on the hyperlink score. The updated category score is compared with the hyperlink score to select a category from the one or more categories. The set of words associated with the category is updated based on content of a web page pointed by the at least one hyperlink.

Systems and methods of web crawling

A method, non-transitory computer readable medium, and apparatus for predicting a location of a short text are disclosed. For example, the method trains a model for each parent node of a grid of a plurality of parent nodes of a geographical area using a plurality of previously received short texts having a known location, receives the short text, applies a hierarchical modeling algorithm to the short text to calculate a weighted probability for one or more child nodes of one or more of the plurality of parent nodes and predicts the location of the short text within the grid of the geographical area based upon a child node of the one or more child nodes having a highest weighted probability.

Method and apparatus for location prediction using short text

Online learning of an ensemble of classifiers or regressors is performed to predict a non-stationary time-varying parameter over a time series of episodes. For an episode, an ensemble action is selected from a set of ensemble actions based on ensemble action quality values (Q values) at an error state for the episode. The selected ensemble action is executed to update the ensemble. A cost of executing the selected ensemble action is computed or retrieved, and a reward is computed indicating how well the updated ensemble predicts the non-stationary time-varying parameter over the episode. The Q value for the selected ensemble action at the error state for the episode is updated based on both the reward and the cost of executing the selected ensemble action. The cost may be based on the ensemble action alone, or on both the ensemble action and the error state for the episode.

Cost-aware non-stationary online learning

A web page identified by a URL stored in a downloads queue is downloaded, and hyperlinks in the downloaded web page are identified. Each hyperlink is screened by parsing the hyperlink (optionally only the URL of the hyperlink) to identify features comprising character strings, computing for each feature values for one or more meta-features indicative of the hyperlinked web page being in a target language, aggregating the meta-feature values to generate a score for the hyperlink, and adding the URL of the hyperlink to the downloads queue conditional upon the score satisfying a screening criterion. The downloading, identifying, and screening are iteratively repeated to perform web crawling, and an index of web pages in the target language is constructed based on analysis of content of the downloaded web pages. The meta-features may include a transliterated target word meta-feature, a language code meta-feature, a country code meta-feature, or so forth.

Growing community of inventors

Meylan, France

Nidhi Singh