Measuring accuracy of semantic graphs with exogenous datasets

Abstract

Provided is a process including: obtaining a semantic similarity graph having nodes corresponding to documents in an analyzed corpus and edges indicating semantic similarity between pairs of the documents; for at least a plurality of nodes in the graph, evaluating accuracy of the edges based on neighboring nodes and an external corpus by performing operations including: identifying the neighboring nodes based on adjacency to the respective node in the graph; selecting documents from an external corpus based on references in the selected documents to entities mentioned in the documents of the neighboring nodes; and determining how semantically similar the respective node is to the selected documents.

Description

CROSS-REFERENCE TO RELATED APPLICATIONS[0001]The present patent is a continuation of U.S. patent application Ser. No. 15 / 016,693, filed 5 Feb. 2016, having the same title, the entire content of which is hereby incorporated by reference.BACKGROUND[0002]1. Field[0003]The present invention relates generally to computational linguistics and, more specifically, to measuring accuracy with exogenous datasets of graphs formed with unsupervised learning techniques.[0004]2. Description of the Related Art[0005]Often people wish to draw inferences based on information contained in, and distributed among, relatively large collections of documents, e.g., substantially more documents than they have time to read or the cognitive capacity to analyze. Certain types of inferences implicate relationships between those documents. For example, it may be useful to organize documents by the subject matter described in the documents, sentiments expressed in the documents, or topics addressed in the document...

AI Technical Summary

Problems solved by technology

Through the sophisticated use of computers, inferences that would otherwise be impractical are potentially attainable, even on relatively large collections of documents.

In many cases, the collections of documents are relatively large, for example, more than 100 documents, and in many cases more than 10,000 documents, making it difficult to gauge whether computer-implemented analyses are accurate.

For instance, an algorithm may work well for certain classes of topics or documents within the corpus, but other classes of topics or documents may yield low-quality results.

Further time and cognitive limitations make it difficult for a human being to effectively manually review each of the documents and compare that document to the algorithm's assessment, causing many analyses relying solely on human review to lead to false conclusions or misleading results, and making it difficult to compare the performance of algorithms.

Images