Semantic Grounding of Tag Relatedness in Social Bookmarking Systems Ciro Cattuto, Dominik Benz, Andreas Hotho, Gerd Stumme ISWC 2008 Hyewon Lim January 11, 2011

Outline  Introduction  Folksonomy Definition and Data  Measures of Relatedness  Qualitative Insights  Semantic Grounding  Discussion and Perspectives 2

Introduction  Folksonomy –Underlying data structure of the social bookmarking systems –Consist of a set of users, a set of tags, a set of resources, & a set of tag assignments –Research communities interested in extracting machine-processable semantic structures from them  Focus on similarity and relatedness of tags –They carry the semantic information within a folksonomy –Provide the link to ontologies 3

Introduction  Define similarity and relatedness for a folksonomy –1. Map the tags to a thesaurus or lexicon like Roget’s thesaurus or WorkNet & measure the relatedness there by means of well-known metrics –2. Define measures of relatedness directly on the network structure of the folksonomy  Vocabulary of folksonomies includes many community-specific terms  Measures of tag relatedness in a folksonomy –Use statistical information about different types of co-occurrence –Adopt the distributional hypothesis 4

Folksonomy Definition and Data  Del.icio.us (Nov 2006) –Restricted dataset to the 10,000 most frequent tags & resources/users that have been associated with at least one of those tags –|U| = 476,378 users |T| = 10,000 tags |R| = 12,660,470 resources |Y| = 101,491,722 tag assignments 5

Measures of Relatedness  Five measures of tag relatedness –Co-occurrence count –Three distributional measures  Use the cosine similarity in the vector spaces spanned by U, T, & R –FolkRank 6

Measures of Relatedness Co-occurrence  Tag-tag co-occurrence graph –Weighted undirected graph  Weight of the edge –The number of posts that contain both t 1 and t 2 –Computation  Create a sorted list of all tag pairs which occur together in a post  Group this list by each tag and sort by count Ontologies are us: A unified model of social networks and semantics, Peter Mika, Journal of Web Semantics,

Measures of Relatedness Distributional Measures  Vector space R X –X ∈ {U, T, R} –Each tag t is represented by a vector v t ∈ R X  Tag Context Similarity (TagCont) –Computed in the vector space R T –Entries of the vector v t ∈ R T are defined by v tt’ := w(t, t’)  v tt = 0  two tags to be considered related when they occur in a similar context, and not when they occur together 8

Measures of Relatedness Distributional Measures  Resource Context Similarity (ResCont) –Computed in the vector space R R –The vector v t ∈ R R  Counting how often a tag t is used to annotate a certain resource r ∈ R v tr := card{ u ∈ U | (u, t, r) ∈ Y }  User Context Similarity (UserCont) –Computed in the vector space R U –The vector v t ∈ R U  Counting how often a tag t is used to annotate a certain resource r ∈ R v tu := card{ r ∈ R | (u, t, r) ∈ Y }  Cosine similarity cossim(t 1, t 2 ) = –Ranges from 0 to 1 9

Measures of Relatedness FolkRank  PageRank algorithm –A web page is important if there are many pages linking to it, and if those pages are important themselves  FolkRank –A graph-based measure –An adaptation of PageRank to folksonomies –A resource which is tagged with important tags by important users becomes important itself 10

Qualitative Insights  A few examples of the related tags 11

Qualitative Insights  A few examples of the related tags (cont’d) –In many cases TagCont & ResCont probide more synonyms than the others  web2.0 – its alternative spellings games – semantically similar tags tobuy – equivalent functional tags –java – python, perl and c++ (TagCont): considered as siblings  TagCont is measuring the frequency of co-occurrence with other tags in the global context of the folksonomy, whereas the co-occurrence measure and FolkRank measure the frequency of co-occurrence with other tags in the same posts –TagCont & ResCont seem to yield equivalent results, especially in terms of synonym identification 12

Qualitative Insights 13

Qualitative Insights  Whether the most closely related tags are shared across measures of relatedness –UserCont & TagCont does not exhibit a strong similarity  2.66 tags w/ ResCont: can be attributed to shared synonym tags –FolkRank --- co-occurrence (6.81)  Given a tag t, its related tags according to FolkRank are tags with a high frequency of co-occurrence with t 14

Qualitative Insights  Average rank of the related tags as a function of the rank of the original tag –Co-occ & FolkRank  Most of the related tags are high-frequency tags, independently of the original tag –Context measures  Tags obtained from context relatedness span a broader range of ranks 15

Semantic Grounding  Ground measures of tag relatedness by using WordNet –In WordNet words are grouped into synsets  Several factors limit the WordNet coverage of del.icio.us tags –WordNet only covers the English language  Del.icio.us contains tags from different languages, tags that are not words at all, and is an open-ended system –Structure of WordNet itself  Cannot compute adjective using is-a network for nouns and verbs 16

Semantic Grounding  An assessment of the measure of relatedness –can be carried out by measuring – in WordNet – the average semantic distance between a tag and the corresponding most closely related tag  1. Loop over the tags that are both in del.icio.us and WordNet  2. For each of those tags we use the chosen measure to find the corresponding most related tag  3. If the most related tag is also in WordNet, measure the semantic distance between the synsets 17

Semantic Grounding  The normalized distribution P(n) of shortest-path lengths n connecting a tag to its closest related tag in WordNet –TagCont & ResCont: synonyms or siblings of the original tag –Other measures: more general tags –Higher value at n=2  sibling relation? 18

Semantic Grounding  Average edge type composition 19

Semantic Grounding  Measure for every path the hierarchical displacement ∆l in WordNet – difference in hierarchical depth between the synsets 20

Discussion and Perspectives  Contributions –1. introduces a systematic methodology for characterizing measures of tag relatedness in a folksonomy –2. the question of emergent semantics  Globally meaningful tag relations can be harvested from an aggregated and uncontrolled folksonomy vocabulary  Specifically, the measures based on TagCont & ResCont are capable of identifying tags belonging to a common semantic concepts 21

Discussion and Perspectives  The relatedness measures is best for –Synonym discovery  TagCont & ResCont deliver not only spelling variants, but also terms that belong to the same WordNet synset –Concept hierarchy  FolkRank & co-occurrence seem to yield more general tags  provide valuable input for algorithms to extract taxonomic relationships between tags –Tag recommendations  FolkRank & co-occurrence  FolkRank delivered superior and more personalized results than co-occ –Query expansion  TagCont & ResCont could be used to discover synonyms  Expand original tag query using the tags obtained by TagCont & ResCont 22