This work explores computing distributionalsimilarity between sub-parses, I.e.,fragments of a parse tree, as an extensionto general lexical distributional similaritytechniques. In the same way that lexicaldistributional similarity is used to estimatelexical semantic similarity, we propose usingdistributional similarity between subparsesto estimate the semantic similarity ofphrases. Such a technique will allow us toidentify paraphrases where the componentwords are not semantically similar. Wedemonstrate the potential of the method byapplying it to a small number of examplesand showing that the paraphrases are moresimilar than the non-paraphrases.
展开▼