lexical relations between synsets, 1. e, one can browse for hyponyms and hyper nyms of a given word, other similar words or even antonyms The reason why we did not use Wordnet as our source of keywords and relationships at the first place comes from our plans to leverage our hierarchy for user modeling purposes. Tags acquired from social tagging systems are closer to the user than Wordnet, they are more"webbish" Even more, some relationships hich emerge from the social tagging systems could never come out of Wordnet for example, a relationship between words ie, png and bugs, pointing to th well-known problem of Internet Explorer's broken PNG support ). Relationshi cquired from social tagging systems are like ordinary people used them, not like linguistics have decided them to be However, we believe, that Wordnet can still significantly contribute to the quality of the tag-based models built upon the generated tag hierarchy. Informa- tion on semantic relationships between words can be used not only to identify particular subtrees, which should be merged (in case of synonymy) or divided (because of ambiguity), but also to add new words(along with their relation- ships)to the hierarchy, which would raise the probability that we will be able to map user's interest to our keywords 3 Experiment In order to determine the feasibility of the proposed approach to deriving rela- tionships between tags from folksonomies for the purposes of tag-based user mod eling and to determine the optimal setting of the algorithm, we performed several experiments with two different folksonomies. Our main concern was whether the algorithm creates cohesive groups of tags(subtrees)without significant flaws of 3.1 Data Wecollectedapartofdeliciousbookmarkingsite(http://delicious.com dataset by periodically polling their Rss feeds. First, we used a recent activity RSS feed to obtain a list of 128 448 unique user login names, next we used this information in user-scoped RSS feeds to obtain all tags and tagged pages for a given login name the second dataset, we took the anonymized folksonomy of users-tags- publications from the Citeulike(htTp: //citeulike. org), which is a system for tagging and searching for scholarly papers. Summarization of data we w able to acquire so far is listed in the Table 1 first thing we were interested in was whether these two folksonomies are used in a similar manner. The graph on Fig. 1 depicts a distribution of tags pages in delicious in a logarithmic scale. We can see that the distribution of tags hypernym-the generic term used to designate a whole class of specific instances. Y is a hypernym of X if X is a(kind of)ylexical relations between synsets, i.e., one can browse for hyponyms and hyper￾nyms1 of a given word, other similar words or even antonyms. The reason why we did not use Wordnet as our source of keywords and relationships at the first place comes from our plans to leverage our hierarchy for user modeling purposes. Tags acquired from social tagging systems are closer to the user than Wordnet, they are more “webbish”. Even more, some relationships which emerge from the social tagging systems could never come out of Wordnet (for example, a relationship between words ie, png and bugs, pointing to the well-known problem of Internet Explorer’s broken PNG support). Relationships acquired from social tagging systems are like ordinary people used them, not like linguistics have decided them to be. However, we believe, that Wordnet can still significantly contribute to the quality of the tag-based models built upon the generated tag hierarchy. Informa￾tion on semantic relationships between words can be used not only to identify particular subtrees, which should be merged (in case of synonymy) or divided (because of ambiguity), but also to add new words (along with their relation￾ships) to the hierarchy, which would raise the probability that we will be able to map user’s interest to our keywords. 3 Experiment In order to determine the feasibility of the proposed approach to deriving rela￾tionships between tags from folksonomies for the purposes of tag-based user mod￾eling and to determine the optimal setting of the algorithm, we performed several experiments with two different folksonomies. Our main concern was whether the algorithm creates cohesive groups of tags (subtrees) without significant flaws of the context. 3.1 Data We collected a part of delicious bookmarking site (http://delicious.com) dataset by periodically polling their RSS feeds. First, we used a ’recent activity’ RSS feed to obtain a list of 128 448 unique user login names, next we used this information in user-scoped RSS feeds to obtain all tags and tagged pages for a given login name. As the second dataset, we took the anonymized folksonomy of users-tags￾publications from the CiteULike (http://citeulike.org), which is a system for tagging and searching for scholarly papers. Summarization of data we were able to acquire so far is listed in the Table 1. First thing we were interested in was whether these two folksonomies are used in a similar manner. The graph on Fig. 1 depicts a distribution of tags on pages in delicious in a logarithmic scale. We can see that the distribution of tags 1 hypernym – the generic term used to designate a whole class of specific instances. Y is a hypernym of X if X is a (kind of) Y