Unsupervised All-words Word Sense Disambiguation with Grammatical Dependencies

We present experiments that analyze the necessity of using a highly interconnected word/sense graph for unsupervised allwords word sense disambiguation. We show that allowing only grammatically related words to influence each other’s senses leads to disambiguation results on a par with the best graph-based systems, while greatly reducing the computation load. We also compare two methods for computing selectional preferences between the senses of every two grammatically related words: one using a Lesk-based measure on WordNet, the other using dependency relations from the British National Corpus. The best configuration uses the syntactically-constrained graph, selectional preferences computed from the corpus and a PageRank tie-breaking algorithm. We especially note good performance when disambiguating verbs with grammatically constrained links.