Learning to Automatically Discover Meronyms

We present a system for automatically discovering meronyms (noun pairs in a part-whole relationship) from text corpora. More precisely, our system begins by parsing and extracting dependency paths similar to (but not the same as) those used by (Snow et al., 2004). For each noun pair we calculate an empirical distribution over dependency relations, which are then used as features of a Support Vector Machine classi er. Noun pairs are labeled as meronyms if there exists a path traversing only meronym and hypernym links between the nouns. Since the method of labeling training examples treats sentences as bags of words, our training examples are extremely noisy. However, we are able to nd a classi er that performs better than similar previous work.