GYDER: Maxent Metonymy Resolution

Proceedings of the 4th International Workshop on Semantic Evaluations (SemEval-2007), pages 161–164, Prague, June 2007. c⃝2007 Association for Computational Linguistics GYDER: maxent metonymy resolution Rich´ard Farkas University of Szeged Department of Informatics H-6720 Szeged, ´Arp´ad t´er 2. rfarkas@inf.u-szeged.hu Eszter Simon Budapest U. of Technology Dept. of Cognitive Science H-1111 Budapest, Stoczek u 2. esimon@cogsci.bme.hu Gy¨orgy Szarvas University of Szeged Department of Informatics H-6720 Szeged, ´Arp´ad t´er 2. szarvas@inf.u-szeged.hu D´aniel Varga Budapest U. of Technology MOKK Media Research H-1111 Budapest, Stoczek u 2. daniel@mokk.bme.hu Abstract Though the GYDER system has achieved the highest accuracy scores for the metonymy resolution shared task SemEval-2007 in all six subtasks, we don’tat consider the results (72.80% accuracy for org, 84.36% for loc) particularly impres- sive, and argue that metonymy resolution needs more features. 1 Introduction In linguistics metonymy means using one term, or one specific sense of a term, to refer to another, related term or sense. For example, in ‘the pen is mightier than the sword’ pen refers to writing, the force of ideas, while sword refers to military force. Named Entity Recognition (NER) is of key importance in numerous natural language pro- cessing applications ranging from information ex- traction to machine translation. Metonymic usage of named entities is frequent in natural language. On the basic NER categories person, place, organisation state-of-the-art systems generally perform in the mid to the high nineties. These sys- tems typically do not distinguish between literal or metonymic usage of entity names, even though this would be helpful for most applications. Resolving metonymic usage of proper names would therefore directly benefit NER and indirectly all NLP tasks (such as anaphor resolution) that require NER. Markert and Nissim (2002) outlined a corpus- based approach to proper name metonymy as a se- mantic classification problem that forms the basis of the 2007 SemEval metonymy resolution task. Instances like ‘He was shocked by Vietnam’ or ‘Schengen boosted tourism’ were assigned to broad categories like place-for-event, sometimes ignoring narrower distinctions, such as the fact that it wasn’t the signing of the treaty at Schengen but rather its actual implementation (which didn’t take place at Schengen) that boosted tourism. But the corpus makes clear that even with these (sometimes coarse) class distinctions, several metonymy types seem to appear extremely rarely in actual texts. The shared task focused on two broad named en- tity classes as metonymic sources, location and org, each having several target classes. For more details on the data sets, see the task description pa- per Markert and Nissim (2007). Several categories (e.g. place-for-event, organisation-for-index) did not contain a sufficient number of examples for machine learn- ing, and we decided early on to accept the fact that these categories will not be learned and to concen- trate on those classes where learning seemed feasi- ble. The shared task itself consisted of 3 subtasks of different granularity for both organisation and lo- cation names. The fine-grained evaluation aimed at distinguishing between all categories, while the medium-grained evaluation grouped different types of metonymic usage …

Content truncated to 500 words