Analogical Reasoning in Machines

Identifying pairs of words that are analogous is easy for humans. For example, most of us can easily understand that the words cat:meow are analogous to dog:bark. Our ability to identify the relational similarity between simple pairs of words is thought to underly many cognitive and linguistic processes. A new paper by Peter D. Turney titled Measuring Semantic Similarity by Latent Relational Analysis (PDF format), describes an algorithm called, you guessed it, Latent Relational Analysis (LRA) that can give machines the ability to measure the relational similarity of words. The algorithm uses a vector space model (VSM) in which words are represented by vectors determined by the frequencies of patterns. The similarity of words is calculated from the cosine of the angle between two word's vectors. The new algorithm was able to achieve human-level performance on analogy questions from a college-level multiple-choice test. For more information see the author's website.