Random-Walk Models of Term Semantics: An Application to Opinion-Related Properties

It has recently been proposed that term senses can be automatically ranked by how strongly they possess a given opinion-related property, by applying PageRank, the well known random-walk algorithm lying at the basis of the Google search engine, to a graph in which nodes are represented by eXtended WordNet synsets and links are represented by the binary relation si I sk (“the gloss of synset si contains a term belonging to synset sk”). In other words, these properties are seen as “flowing” through this graph, from the definiendum (i.e., the synset being defined) to the definiens (i.e., a synset which occurs in the gloss of the definiendum), with PageRank controlling the “hydraulics” of this flow. In this paper we contend that two other random-walk algorithms may be equally adequate to this task, and provide an intuitive justification to support this claim. The first is a random-walk algorithm different from PageRank which we apply to the “inverse” graph, i.e., with properties flowing from the definiens to the definiendum. The second algorithm is a bidirectional randomwalk algorithm, which assumes that properties may flow from the definiens to the definiendum and viceversa. We report results which significantly improve on the ones obtained by simple PageRank.