网络搜索-填补信息断层.ppt

office microsoft office office tv show ms office download the office office shoes openoffice footware office uk office season 6 /The_office/ /title/tt0386676/ www.office.co.uk /en-us/products/ Con: sparse Precise information from Wisdom of crowds Pro: higher precision software Shoe supplier TV show software Con: irrelevant/spam/advertisement/ambiguity Great Context Describing the query Pro: higher recall A. Identify Search Intents (Data) leverage two types of auxiliary data Search result snippets Clickthrough A. Identify Search Intents (Algorithm) Regularized Topic Model top search result snippets virtual documents words in snippets words potential search intents topics PLSI model log-likelihood Topic Model Regularization two queries share many same clicked URLs convey similar search intent co-click matrix Search result snippets Clickthrough powerful constraint: B.Intent-Aware Similarity Measure (Pair-wise) Similarity independently measured by pair-wise metrics I. Extract intent-aware representations word vector representation original: intent-aware: expected search intent distribution for each word occurrence wl given query qi word vector representation under k-th search intent II. Apply Pair-wise similarity measures similarity under k-th search intent B.Intent-Aware Similarity Measure (Graph-based) similarity calculated over the query graph I. Extract intent-aware representations query similarity graph original: adjacency matrix Jaccard coefficient spectral embedding query representation under k-th search intent II. Apply Graph-based similarity measures similarity under k-th search intent intent-aware: query similarity graph under k-th search intent the probability that an edge will be generated between query qi with search intent sk and query qj with search intent sk’ Result Expected inter-intra ratio Part II: Ranking Ranking is a Central Problem! Ranking Web Search Information Filtering Recommendation Aggregation e.g. Diversity Output a perfect ranking from