SUCRE: A Modular System for Coreference Resolution

Proceedings of the 5th International Workshop on Semantic Evaluation, ACL 2010, pages 92–95, Uppsala, Sweden, 15-16 July 2010. c⃝2010 Association for Computational Linguistics SUCRE: A Modular System for Coreference Resolution Hamidreza Kobdani and Hinrich Sch¨utze Institute for Natural Language Processing University of Stuttgart, Germany kobdani@ims.uni-stuttgart.de Abstract This paper presents SUCRE, a new soft- ware tool for coreference resolution and its feature engineering. It is able to sep- arately do noun, pronoun and full coref- erence resolution. SUCRE introduces a new approach to the feature engineering of coreference resolution based on a rela- tional database model and a regular feature definition language. SUCRE successfully participated in SemEval-2010 Task 1 on Coreference Resolution in Multiple Lan- guages (Recasens et al., 2010) for gold and regular closed annotation tracks of six languages. It obtained the best results in several categories, including the regular closed annotation tracks of English and German. 1 Introduction In this paper, we introduce a new software tool for coreference resolution. Coreference resolution is the process of finding discourse entities (mark- ables) referring to the same real-world entity or concept. In other words, this process groups the markables of a document into equivalence classes (coreference entities) so that all markables in an entity are coreferent. There are various publicly available systems that perform coreference resolution, such as BART (Versley et al., 2008) and GUITAR (Stein- berger et al., 2007). A considerable engineering effort is needed for the full coreference resolution task, and a significant part of this effort concerns feature engineering. Thus, a system which is able to extract the features based on a feature defini- tion language can help the researcher reduce the implementation effort needed for feature extrac- tion. Most methods of coreference resolution, if providing a baseline, usually use a feature set sim- ilar to (Soon et al., 2001) or (Ng and Cardie, 2002) and do the feature extraction in the preprocessing stage. SUCRE has been developed to provide a more flexible method for feature engineering of coreference resolution. It has a novel approach to model an unstructured text corpus in a structured framework by using a relational database model and a regular feature definition language to define and extract the features. Relational databases are a well-known technology for structured data mod- eling and are supported by a wide array of soft- ware and tools. Converting a text corpus to/from its equivalent relational database model is straight- forward in our framework. A regular language for feature definition is a very flexible method to extract different features from text. In addition to features defined di- rectly in SUCRE, it accepts also externally ex- tracted/generated features. Its modular architec- ture makes it possible to use any externally avail- able classification method too. In addition to link features (features related to a markable pair), it is also possible to define other kinds of features: atomic word and markable features. This ap- proach to feature engineering is suitable not only for knowledge-rich but also for knowledge-poor datasets. It is also language …

Content truncated to 500 words