Multilingual Lexicon Bootstrapping - Improving a Lexicon Induction System Using a Parallel Corpus

International Joint Conference on Natural Language Processing, pages 844–848, Nagoya, Japan, 14-18 October 2013. Multilingual Lexicon Bootstrapping - Improving a Lexicon Induction System Using a Parallel Corpus Patrick Ziering1 Lonneke van der Plas1 Hinrich Sch¨utze2 1Institute for NLP, University of Stuttgart, Germany 2CIS, University of Munich, Germany {Patrick.Ziering, Lonneke.vanderPlas}@ims.uni-stuttgart.de Abstract We address the task of improving the qual- ity of lexicon bootstrapping, i.e., of ex- panding a semantic lexicon on a given cor- pus. A main problem of iterative boot- strapping techniques is the fact that lex- icon quality degrades gradually as more and more false terms are added. We propose to exploit linguistic variation be- tween languages to reduce this problem of semantic drift with a knowledge-lean and language-independent ensemble method. Our results on English and German show that lexicon bootstrapping benefits signifi- cantly from the multilingual symbiosis. 1 Introduction High-quality semantic lexicons are an essen- tial resource for many natural language process- ing (NLP) tasks like information extraction or anaphora resolution. Methods for automatically bootstrapping semantic lexicons given a seed list often struggle with lexicon accuracy decrease in higher iterations depending on corpus size (Igo and Riloff, 2009). One reason for this is semantic drift, which occurs when erroneous terms and/or contexts are introduced into and then dominate the iterative process (Curran et al., 2007). For instance, the ambiguity found in female names such as Iris and Rose may cause the induced terms to drift into flower names (McIntosh and Curran, 2009). Examples from the patent domain, that we are focusing on in this work, are PROCESSES that may drift into the semantic class of OBJECTS when terms such as energy storage and spring coupling are induced. Previous work has used the cross-lingual corre- spondence between variations in linguistic struc- ture and variations in ambiguity as a form of natu- rally occurring supervision in unsupervised learn- ing for a number of tasks (Dagan et al., 1991; Snyder and Barzilay, 2010). On the lexical level, cross-lingual variations proved to remedy prob- lems related to polysemy for synonym acquisition (Van der Plas and Tiedemann, 2010) and word sense disambiguation (Lefever and Hoste, 2010). We hypothesize that cross-lingual divergences will be preeminently suitable to remedy problems related to semantic drift in iterative bootstrapping, where lexical and structural ambiguity give rise to erroneous terms and/or contexts. Languages are not isomorphic: ambiguous terms and con- texts are frequently language-specific. In our ex- ample above, the English term energy storage is ambiguous, however, in German, each reading has its own translation. Energy storage is translated with Energiespeicher in the OBJECT reading and Energiespeicherung in the PROCESS reading. Our multilingual ensemble lexicon bootstrap- ping system is inspired by Basilisk (Thelen and Riloff, 2002). Previous work has addressed se- mantic drift in Basilisk by conflict resolution be- tween several classes (Thelen and Riloff, 2002), by using web queries (Igo and Riloff, 2009) and by combining Basilisk in an ensemble with an SVM tagger and a coreference resolution system (Qadir and Riloff, 2012). These approaches are mono- …

Content truncated to 500 words