Research on document similarity hasshown that complex representations arenot more accurate than the simple bag-ofwords.Term clustering, e.g. using latentsemantic indexing, word co-occurrencesor synonym relations using a word ontologyhave been shown not very effective.In particular, when to extend the similarityfunction external prior knowledge isused, e.g. WordNet, the retrieval systemdecreases its performance. The critical issueshere are methods and conditions tointegrate such knowledge.In this paper we propose kernel functionsto add prior knowledge to learningalgorithms for document classification.Such kernels use a term similaritymeasure based on the WordNet hierarchy.The kernel trick is used to implement suchspace in a balanced and statistically coherentway. Cross-validation results showthe benefit of the approach for the SupportVector Machines when few training data isavailable.
展开▼
