Highly Multilingual Coreference Resolution Exploiting a Mature Entity Repository

Proceedings of Recent Advances in Natural Language Processing, pages 254–260, Hissar, Bulgaria, 12-14 September 2011. Highly Multilingual Coreference Resolution Exploiting a Mature Entity Repository Josef Steinberger, Jenya Belyaeva, Jonathan Crawley, Leonida Della-Rocca, Mohamed Ebrahim, Maud Ehrmann, Mijail Kabadjov, Ralf Steinberger and Erik van der Goot EC Joint Research Centre 21027, Ispra (VA), Italy name.surname@jrc.ec.europa.eu Abstract In this paper we present an approach to large-scale coreference resolution for an ample set of human languages, with a par- ticular emphasis on time performance and precision. One of the distinctive features of our approach is the use of a mature multilingual named entity repository (per- sons and organizations) gradually com- piled over the past few years. Our exper- iments show promising results – an over- all precision of 94% tested on seven dif- ferent languages. We also present an ex- trinsic evaluation on seven languages in the context of summarization where we gauge the contribution of the coreference resolver towards the end summarization performance. 1 Introduction Recent work on coreference resolution has been largely dominated by machine learning ap- proaches and predominantly for the English lan- guage (Ng and Cardie, 2002; Ponzetto and Strube, 2006; Luo, 2007). This is in great part due to the availability of annotated corpora such as MUC-6/7 (Hirschman, 1998), ACE-2/3/4/5 (NIST, 2004), GNOME (Poesio et al., 2004) and large-scale crowdsourcing efforts like Phrase Detectives.1 One of the big advantages of machine learn- ing approaches is that they are reasonably easy to reproduce given that the set of input fea- tures are documented well, since there are many good open-source platforms for machine learning (e.g., WEKA2) and machine-learning-based coref- erence (e.g., BART3 (Versley et al., 2008)). However, intrinsic evaluations can pose prob- lems. As pointed out by (Stoyanov et al., 2009) 1http://www.phrasedetectives.org. 2http://www.cs.waikato.ac.nz/ml/weka/. 3http://www.bart-coref.org/. there is too much variation in reported results across data sets to be able to draw robust conclu- sions on the state-of-the-art in the area for which they proposed a method for reporting results on a data set that makes it easier to predict performance on other data sets (by breaking down results into names, types of pronouns, nominals etc.). Also, intrinsic evaluations can be highly sensitive to pre- processing (Mitkov, 2002). There is agreement in the community on the level of resolution difficulty on major types of coreferential expressions. For instance, proper names are considered to be the easiest to resolve, followed by pronouns, in turn followed by com- mon nouns. One of the main reasons why com- mon noun coreference is challenging is because they often share little or no surface linguistic fea- tures with their antecedents and require world or encyclopedic knowledge for their resolution (see (Kabadjov, 2007) for a study for English). For instance, Ponzetto and Strube (2006) proposed to use WordNet and Wikipedia to address the problem of bringing in world and/or encyclope- dic knowledge into their system for coreference resolution in English reporting improvements for common noun resolution. In this work we address two important remain- ing gaps in …

Content truncated to 500 words