Extracting instance attributes from text

摘要

Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for extracting instance attributes from text are described. In one aspect, a method exploits weakly-supervised and unsupervised instance relatedness data, available in the form of labeled classes of instances and distributionally similar instances. The method organizes the data into a graph containing instances, class labels, and attributes. The method propagates attributes among related instances, through random walks over the graph.

主权项

1. A method executed by a data processing apparatus, comprising:
identifying, in a graph stored in a memory and representing information instances and attributes describing the information instances, first attributes describing a first information instance and second attributes describing a plurality of second information instances related to the first information instance, wherein the graph includes attribute nodes each representing a respective attribute, instance nodes each representing a respective information instance, and edges each connecting an attribute node and an instance node, each edge representing that a non-zero likelihood exists that the attribute of the connected attribute is related to the information instance of the connected instance node, and wherein the graph further includes paths connecting pairs of instance nodes, wherein the existence of a path indicates that a semantic relationship exists between the information instances of the connected instance nodes; performing a first random walk in the graph from the first information instance to a first attribute of the second attributes through the second information instances; calculating a first relatedness value for the first random walk from the first information instance to the first attribute of the second attributes; determining that the first relatedness value for the first random walk from the first information instance to a first attribute of the second attributes satisfies a threshold relatedness value; and creating an edge between the first attribute of the second attributes and the first information instance in the graph.