Titel: Multilingual Coreference Resolution with BART

BART, the Baltimore Anaphora Resolution Toolkit is a framework for
coreference resolution that was conceived as a re-engineered version
of the EML-R system by Ponzetto & Strube (2006) on the ELERFED project
at the JHU Summer Workshop 2007. After an introduction to the general
architecture of BART, I will use multiple case studies to present
recent developments highlighting the advantages of a framework-based
approach to coreference resolution.

1. Kernel-based Expletive Detection Tree kernels offer a
straightforward way to incorporate both structural and lexical
information in classification tasks. I will present an approach for
expletive detection that achieves near-state of the art results and
discuss the different ways to integrate the classifier in a
coreference system.

2. Parameter Optimization for a Multi-Way Ranking Resolver Ranking
classifiers are a very natural way of presenting the coreference
resolution problem, as they can be used to integrate both antecedent
choice and discourse-new detection in one single decision. However,
evaluation metrics such as Vilain et al.'s MUC measure have a stronger
bias towards over-merging that puts more simplistic decoding
strategies such as the one by Soon et al. (2001) at an advantage.
Hence, it is common, as in Luo et al. (2004), to include a parameter
that biases the coreference decisions towards a different
precision-recall tradeoffs.  It is advantageous to use different
ranking classifiers for different kinds of (potentially) anaphoric
expressions such as pronouns, nominals, and names, which however
complicates the task of finding good values for the bias parameter for
each classifier.

3. The LanguagePlugin Architecture for Multilingual Coreference
Resolution In joint work with Olga Uryupina and Massimo Poesio
(Trento) as well as Samuel Broscheid and Simone Ponzetto (Heidelberg),
BART has been adapted to do coreference resolution for Italian and
German, with state of the art results in both cases. Besides
presenting the interface that allows to easily add new languages, I
will also raise the question how and if multilingual coreference
resolution can allow us to make general statements on the utility of
certain features to a system.