Abstract
Relational similarity measures the correspondence of the semantic relations that exist between the two words in word pairs. Accurately measuring relational similarity is important for various natural language processing tasks such as, relational search, noun-modifier classification, and analogy detection. Despite this need, the features that accurately express the relational similarity between two word pairs remain largely unknown. So far, methods have been proposed based on linguistic intuitions such as the functional space proposed by Turney [1], which consists purely of verbs. In contrast, we propose a data-driven approach for discovering feature spaces for relational similarity measurement. Specifically, we use a linear-SVM classifier to select features using training instances, where two pairs of words are labeled as analogous or non-analogous. We evaluate the discovered feature space by measuring the relational similarity for relational classification task in which we aim to classify a given word-pair to a specific relation from a predefined set of relations. Linear classifier for ranking the best feature for relational space has been compared with different methods namely, Kullback Leibler divergence (KL), Pointwise Mutual Information (PMI). Experimental results show that our proposed classification method accurately discovers a discriminative features for measuring relational similarity. Furthermore, experiments show that the proposed method requires small number of relational features while still maintaining reasonable relational similarity accuracy.