Skip to content

LatentRelationalAnalysis

fozziethebeat edited this page Oct 26, 2011 · 5 revisions

Latent Relational Analysis

Introduction

Latent Relational Analysis is a semantic space algorithm that measures relational similarity. When two word pairs have a high degree of relational similarity, we say they are "analogous". LRA is an extension of the [Vector Space Model] (/fozziethebeat/S-Space/wiki/VectorSpaceModel) (VSM), which constructs a matrix based on predefined patterns.

LRA improves VSM by:

  • deriving patterns automatically from a corpus
  • using Singular Value Decomposition to smooth frequency data
  • generating synonyms that are used to explore word pairs with similar meanings

The algorithm requires a search engine with a very large corpus of text and a broad coverage thesaurus of synonyms.

LRA takes as input a set of word pairs, and constructs a matrix that can be used to find the relational similarity between any two word pairs.

For more information on Latent Relational Analysis and its implementation, see the following papers:

LRA Implementation

The implementation of Latent Relational Analysis is located in one file, LRA.java. This implementation uses the [Lucune Search Engine] (http://lucene.apache.org/java/docs/) for optimal indexing and filtering of word pairs using any given corpus. It also uses Wordnet through the [JAWS] (http://lyle.smu.edu/~tspell/jaws/index.html) interface in order to find alternate word pairs from given input pairs.