Duluth-WSI: SenseClusters Applied to the Sense Induction Task of SemEval-2

Proceedings of the 5th International Workshop on Semantic Evaluation, ACL 2010, pages 363–366, Uppsala, Sweden, 15-16 July 2010. c⃝2010 Association for Computational Linguistics Duluth-WSI: SenseClusters Applied to the Sense Induction Task of SemEval-2 Ted Pedersen Department of Computer Science University of Minnesota, Duluth Duluth, MN 55812 tpederse@d.umn.edu http://senseclusters.sourceforge.net Abstract The Duluth-WSI systems in SemEval-2 built word co–occurrence matrices from the task test data to create a second order co–occurrence representation of those test instances. The senses of words were in- duced by clustering these instances, where the number of clusters was automatically predicted. The Duluth-Mix system was a variation of WSI that used the combina- tion of training and test data to create the co-occurrence matrix. The Duluth-R sys- tem was a series of random baselines. 1 Introduction The Duluth systems in the sense induction task of SemEval-2 (Manandhar et al., 2010) were based on SenseClusters (v1.01), a freely available open source software package which relies on the premise that words with similar meanings will oc- cur in similar contexts (Purandare and Pedersen, 2004). The data for the sense induction task in- cluded 100 ambiguous words made up of 50 nouns and 50 verbs. There were a total of 8,915 test in- stances and 879,807 training instances provided. Note that neither the training nor the test data was sense tagged. The training data was made avail- able as a resource for participants, with the under- standing that system evaluation would be done on the test instances only. The organizers held back a gold standard annotation of the test data that was only used for evaluation. Five Duluth-WSI systems participated in this task, six Duluth-Mix systems, and five Duluth Random systems. The WSI and Mix systems al- most always represented the test instances using second order co–occurrences, where each word in a test instance is replaced by a vector that shows the words with which it co-occurs. The word vec- tors that make up a test instance are averaged to- gether to make up a new representation for that instance. All the test instances for a word are clus- tered, and the number of senses is automatically predicted by either the PK2 measure or Adapted Gap Statistic (Pedersen and Kulkarni, 2006). In the Duluth systems the co-occurrence matri- ces are either based on order-dependent bigrams or unordered pairs of words, both of which can be separated by up to some given number of interven- ing words. Bigrams are used to preserve distinc- tions between collocations such as cat house and house cat, whereas co–occurrences do not con- sider order and would treat these two as being equivalent. 2 Duluth-WSI systems The Duluth-WSI systems build co-occurrence ma- trices from the test data by identifying bigrams or co–occurrences that occur with up to eight inter- mediate words between them in instances of am- biguous nouns, and up to 23 intermediate words for the verbs. Any bigram (bi) or co–occurrence (co) that occurs more than 5 times with up to the allowed number of …

Content truncated to 500 words