Evaluating Computational Models of Similarity against a Human Rated Dataset.

The first genuine human-rated similarity set for the Romanian language aligned with an English similarity dataset is presented. Two computational models automatically learn the similarity scores for the word pairs in the similarity set. The first model learns the similarity scores from language corpora. The second one assigns a similarity score based on the taxonomic structure of a semantic network. We studied what model captures human similarity scores best, the language influence on the perception of similarity, and the impact of parts of speech on similarity. [ABSTRACT FROM AUTHOR]