Duluth at SemEval 2016 Task 14: Extending Gloss Overlaps to Enrich Semantic Taxonomies

Proceedings of SemEval-2016, pages 1328–1331, San Diego, California, June 16-17, 2016. c⃝2016 Association for Computational Linguistics Duluth at SemEval–2016 Task 14 : Extending Gloss Overlaps to Enrich Semantic Taxonomies Ted Pedersen Department of Computer Science University of Minnesota Duluth, MN, 55812 USA tpederse@d.umn.edu Abstract This paper describes the Duluth systems that participated in Task 14 of SemEval 2016, Se- mantic Taxonomy Enrichment. There were three related systems in the formal evalua- tion which are discussed here, along with nu- merous post–evaluation runs. All of these systems identified synonyms between Word- Net and other dictionaries by measuring the gloss overlaps between them. These systems perform better than the random baseline and one post–evaluation variation was within a re- spectable margin of the median result attained by all participating systems. 1 Introduction The goal of Task 14 in SemEval–2016 was to enrich a semantic taxonomy with new word senses. In par- ticular, this task sought to augment WordNet1 with senses that are present in another dictionary. Task 14 drew glosses of words not found in WordNet from a variety of sources that will collectively be referred to as OtherDict in this paper. The method the Duluth systems took was based on scoring overlaps between WordNet and Other- Dict glosses. A OtherDict sense was assigned to the WordNet sense with the highest overlapping score. Overlaps are defined to be exact matches between words and phrases in the glosses. Each overlap is assigned a score that is the square of the number of words in the overlap, and then all the overlaps be- tween a pair of glosses is summed to provide the final score. 1http://wordnet.princeton.edu The premise of relying on overlaps is that senses that are defined using many of the same words are certainly related, and indeed if they are defined using the same words they are likely synonyms. This is a well established and reliable intuition that goes back at least to (Lesk, 1986). Closely related words that may not be synonyms (such as hyponyms or hyper- nyms) will use many of the same words in their defi- nitions, but then have specific differentia that distin- guishes among them. Task 14 asked participants to distinguish between merging a OtherDict sense into an existing Word- Net synset (i.e, a synonym) or attaching it as a hy- ponym (i.e., a more specific example) of a WordNet sense. However, our systems only merge OtherDict senses into WordNet synsets. It seems clear though that this merge versus attach problem can be tack- led by setting some kind of threshold for the amount of overlap, where more significant degrees of over- lap should result in a synonym merge whereas less overlap could indicate a hyponym attach. As yet we have not been successful in determining a reliable method for finding such a threshold. As a result we simply assumed every OtherDict sense would attach to its closest (most overlapping) WordNet sense. Each Duluth system carried out the same pre- processing on both the WordNet and OtherDict …

Content truncated to 500 words