BART: A Multilingual Anaphora Resolution System

Proceedings of the 5th International Workshop on Semantic Evaluation, ACL 2010, pages 104–107, Uppsala, Sweden, 15-16 July 2010. c⃝2010 Association for Computational Linguistics BART: A Multilingual Anaphora Resolution System Samuel Broscheit∗, Massimo Poesio‡, Simone Paolo Ponzetto∗, Kepa Joseba Rodriguez‡, Lorenza Romano†, Olga Uryupina‡, Yannick Versley⋄, Roberto Zanoli† ∗Seminar f¨ur Computerlinguistik, University of Heidelberg ‡CiMeC, University of Trento †Fondazione Bruno Kessler ⋄SFB 833, University of T¨ubingen broscheit@cl.uni-heidelberg.de, massimo.poesio@unitn.it, ponzetto@cl.uni-heidelberg.de, kepa.rodriguez@unitn.it, romano@fbk.eu, uryupina@gmail.com, versley@sfs.uni-tuebingen.de, zanoli@fbk.eu Abstract BART (Versley et al., 2008) is a highly mod- ular toolkit for coreference resolution that supports state-of-the-art statistical approaches and enables efficient feature engineering. For the SemEval task 1 on Coreference Resolu- tion, BART runs have been submitted for Ger- man, English, and Italian. BART relies on a maximum entropy-based classifier for pairs of mentions. A novel entity- mention approach based on Semantic Trees is at the moment only supported for English. 1 Introduction This paper presents a multilingual coreference reso- lution system based on BART (Versley et al., 2008). BART is a modular toolkit for coreference resolution that supports state-of-the-art statistical approaches to the task and enables efficient feature engineer- ing. BART has originally been created and tested for English, but its flexible modular architecture en- sures its portability to other languages and domains. In SemEval-2010 task 1 on Coreference Resolution, BART has shown reliable performance for English, German and Italian. In our SemEval experiments, we mainly focus on extending BART to cover multiple languages. Given a corpus in a new language, one can re-train BART to obtain baseline results. Such a language-agnostic system, however, is only used as a starting point: substantial improvements can be achieved by incor- porating language-specific information with the help of the Language Plugin. This design provides ef- fective separation between linguistic and machine learning aspects of the problem. 2 BART Architecture The BART toolkit has five main components: pre- processing pipeline, mention factory, feature extrac- tion module, decoder and encoder. In addition, an independent LanguagePlugin module handles all the language specific information and is accessible from any component. The architecture is shown on Figure 1. Each module can be accessed independently and thus adjusted to leverage the system’s performance on a particular language or domain. The preprocessing pipeline converts an input doc- ument into a set of lingustic layers, represented as separate XML files. The mention factory uses these layers to extract mentions and assign their basic properties (number, gender etc). The fea- ture extraction module describes pairs of mentions {Mi, Mj}, i < j as a set of features. The decoder generates training examples through a process of sample selection and learns a pairwise classifier. Finally, the encoder generates testing ex- amples through a (possibly distinct) process of sam- ple selection, runs the classifier and partitions the mentions into coreference chains. 3 Language-specific issues Below we briefly describe our language-specific ex- tensions to BART. These issues are addressed in more details in our recent papers (Broscheit et al., 2010; Poesio et al., 2010). 3.1 Mention Detection Robust mention …

Content truncated to 500 words