UKP-WSI: UKP Lab Semeval-2013 Task 11 System Description

Second Joint Conference on Lexical and Computational Semantics (*SEM), Volume 2: Seventh International Workshop on Semantic Evaluation (SemEval 2013), pages 212–216, Atlanta, Georgia, June 14-15, 2013. c⃝2013 Association for Computational Linguistics UKP-WSI: UKP Lab Semeval-2013 Task 11 System Description Hans-Peter Zorn† and Iryna Gurevych†‡ †Ubiquitous Knowledge Processing Lab (UKP-TUDA) Department of Computer Science, Technische Universit¨at Darmstadt ‡Ubiquitous Knowledge Processing Lab (UKP-DIPF) German Institute for Educational Research and Educational Information www.ukp.tu-darmstadt.de Abstract In this paper, we describe the UKP Lab sys- tem participating in the Semeval-2013 task “Word Sense Induction and Disambiguation within an End-User Application”. Our ap- proach uses preprocessing, co-occurrence ex- traction, graph clustering, and a state-of-the- art word sense disambiguation system. We developed a configurable pipeline which can be used to integrate and evaluate other com- ponents for the various steps of the complex task. 1 Introduction The task “Evaluating Word Sense Induction and Word Sense Disambiguation in an End-User Ap- plication” of SemEval-2013 (Navigli and Vannella, 2013) aims at an extrinsic evaluation scheme for WSI to overcome the difficulties inherent to WSI evaluation. The task requires building a WSI sys- tem and combining it with a WSD step to assign the induced sentences to example instances. Word sense disambiguation (WSD) is the task of determining the correct meaning for an ambigu- ous word from its context. WSD algorithms usu- ally choose one sense out of a given set of possible senses for each word. A resource that enumerates possible senses for each word is called a sense in- ventory. Manually created inventories come usually in form of lexical semantic resources, such as Word- Net or more specifically created inventories such as OntoNotes (Hovy et al., 2006). Word sense induction (WSI) on the other hand aims to create such an inventory from a corpus in an unsupervised manner. For each word that should be disambiguated, a WSI algorithm creates a set of context clusters that will be used to define and de- scribe the senses. We build our system upon the open-source DKPro framework 1 and a corresponding WSD component (upcoming). Input for the task comes as two files. One contains the search queries, also referred as topics. Sense in- duction will be performed for each of those topics. The second file contains 6400 entries from the re- sult pages of a search engine. Each entry consists of the title, a snippet and the URL of the corresponding web page. 2 Related Work One of the early approaches to WSI (Sch¨utze, 1998) maps words into a vector space and represents word contexts as vector-sums and use cosine vec- tor similarity, clustering is performed by expectation maximization (EM) clustering. Dorow and Wid- dows (2003) use the BNC to build a co-occurrence graph for nouns, based on a co-occurrence fre- quency threshold. They perform Markov cluster- ing on this graph. Pantel and Lin (2002) proposes a clustering approach called clustering by commit- tee (CBC). This algorithm first selects the words with the highest similarity based on mutual infor- mation and …

Content truncated to 500 words