2/17/041Using WordNet toImprove User Modelling ina Web DocumentRecommender SystemCS 620 Class PresentationPresented by Haifeng HeBernardo Magnini and Carlo Strapparava2Problem A recommender system for a Web site ofmultilingual news Learns user’s interests from the requestedpages Build a model of the user Exploit the model to anticipate whichdocuments in the web site could be interestingfor the user3Previous Work SiteIF, a personal agent for a multilingualnews web site Word-based (word frequency and co-occurrence) Not accurate enough Misinterpret word sense4Main Idea Content-based document representation Build the user model as a semantic networkwhose nodes represent sense (not just words) Retrieve new documents with high semanticrelevance with respect to the use model More accurate and, independent from the language of thedocuments browsed(?!). The problems Require a repository for word senses(WordNet) Word sense disambiguation (WSD)5Word Domain Disambiguation Sense clustering with domain labels (Magniniand Strapparava, 2000) Each word has a domain label (MEDICINE, SPORT,etc) Reduce the WordNet polysemy Covers only noun synsets now6Example7Domain Disambiguation Two steps Given a word, for each domain label of theword, give a score, which is determined by thefrequency of the label among the senses The domain label with the highest score isselected .83 accuracy (Magnini and Strapparava, 2000)89Evaluation and Conclusions Compare the output of two systems againstthe judgments of a human advisor Word-based and synset based H the set of human proposals, S the set of thesystem proposals Precision = ; Recall = Precision increase 34%. Recall increase 15%.SSH || ∩SSH ||
View Full Document