Multi-concept latent semantic analysis queries

Inventors

Jakubik, Paul A.

Assignees

Brainspace Corp

Abstract

A method includes accessing text, identifying a plurality of terms from the text, determining a plurality of term vectors associated with the identified plurality of terms, and clustering the determined plurality of term vectors into a plurality of clusters, the plurality of clusters comprising a first and a second cluster, the first and second clusters each comprising two or more of the determined term vectors. The method further includes creating a first pseudo-document according to the first cluster, creating a second pseudo-document according to the second cluster, identifying a first set of terms associated with the first cluster using latent semantic analysis (LSA) of the first pseudo-document, identifying a second set of terms associated with the second cluster using LSA of the second pseudo-document, and combining the first and second sets of terms into a list of output terms.

Core Innovation

The invention provides multi-concept latent semantic analysis (LSA) query methods that preserve multiple distinct concepts from query text. The method accesses text, identifies a plurality of terms, determines a plurality of term vectors associated with the identified terms, and calculates a weight of each term vector. The term vectors are clustered into a plurality of clusters, where each cluster is related to a distinct concept of the text.

For each concept-related cluster, the method identifies a set of terms associated with the cluster using LSA, and it determines cluster weights based at least on the weights of the term vectors of each cluster. The method determines output-term proportions from the cluster weights by calculating percentages of a list of output terms that should come from each cluster using ratios of the cluster weights to a sum of the cluster weights. It then selects one or more terms from each LSA-derived cluster term set according to the determined percentages, combines the selected terms, and stores the list of output terms having the distinct concepts of the text.

The invention further addresses term importance by using log-entropy statistics to distinguish important term vectors and unimportant term vectors, and it prevents clustering of important term vectors with unimportant term vectors. Additional refinements include log-entropy mixing when creating the list of output terms, optional cleaning of cluster-derived results toward a query pseudo-document using vector normalization and interpolation, and maintaining comprehensive LSA term spaces using libraries of LSA spaces to mitigate vocabulary limitations.

Claims Coverage

The document provides three independent claims: a computer system, a computer-implemented method, and a non-transitory computer-readable medium. Across the independent claims, the inventive features focus on weighted term-vector clustering into distinct concept clusters, LSA-based identification of per-cluster term sets, and proportion-based selection and combination of output terms to represent multiple distinct concepts.

Across the independent claims, the coverage centers on clustering weighted term vectors into distinct concept clusters, extracting concept-specific term sets using LSA, and selecting and combining output terms according to percentages derived from cluster weights so the output list has multiple distinct concepts of the text.

Stated Advantages

Documented Applications