UBIU for Multilingual Coreference Resolution in OntoNotes

Proceedings of the Joint Conference on EMNLP and CoNLL: Shared Task, pages 88–94, Jeju Island, Korea, July 13, 2012. c⃝2012 Association for Computational Linguistics UBIU for Multilingual Coreference Resolution in OntoNotes Desislava Zhekova Sandra K¨ubler Joshua Bonner Marwa Ragheb Yu-Yin Hsu Indiana University Bloomington, IN, USA {dzhekova, skuebler, jebonner, mragheb, hsuy}@indiana.edu Abstract The current work presents the participa- tion of UBIU (Zhekova and K¨ubler, 2010) in the CoNLL-2012 Shared Task: Model- ing Multilingual Unrestricted Coreference in OntoNotes (Pradhan et al., 2012). Our system deals with all three languages: Arabic, Chi- nese and English. The system results show that UBIU works reliably across all three lan- guages, reaching an average score of 40.57 for Arabic, 46.12 for Chinese, and 48.70 for En- glish. For Arabic and Chinese, the system pro- duces high precision, while for English, preci- sion and recall are balanced, which leads to the highest results across languages. 1 Introduction Multilingual coreference resolution has been gain- ing considerable interest among researchers in re- cent years. Yet, only a very small number of sys- tems target coreference resolution (CR) for more than one language (Mitkov, 1999; Harabagiu and Maiorano, 2000; Luo and Zitouni, 2005). A first attempt at gaining insight into the comparability of systems on different languages was accomplished in the SemEval-2010 Task 1: Coreference Resolution in Multiple Languages (Recasens et al., 2010). Six systems participated in that task, UBIU (Zhekova and K¨ubler, 2010) among them. However, since sys- tems participated across the various languages rather irregularly, Recasens et al. (2010) reported that the data points were too few to allow for a proper com- parison between different approaches. Further sig- nificant issues concerned system portability across the various languages and the respective language tuning, the influence of the quantity and quality of diverse linguistic annotations as well as the perfor- mance and behavior of various evaluation metrics. The CoNLL-2011 Shared Task: Modeling Unre- stricted Coreference in OntoNotes (Pradhan et al., 2011) targeted unrestricted CR, which aims at iden- tifying nominal coreference but also event corefer- ence, within an English data set from the OntoNotes corpus. Not surprisingly, attempting to include such event mentions had a detrimental effect on over- all accuracy, and the best performing systems (e.g., (Lee et al., 2011)) did not attempt event anaphora. The current shared task extends the task definition to three different languages (Arabic, Chinese and En- glish), which can prove challenging for rule-based approaches such as the best performing system from 2011 (Lee et al., 2011). In the current paper, we present UBIU, a memory- based coreference resolution system, and its re- sults in the CoNLL-2012 Shared Task. We give an overview of UBIU in Section 2. In Section 3, we present the system results, after which Section 4 lays out some conclusive remarks. 2 UBIU UBIU (Zhekova and K¨ubler, 2010) is a corefer- ence resolution system designed specifically for a multilingual setting. As shown by Recasens et al. (2010), multilingual coreference resolution can be approached by various machine learning methods since …

Content truncated to 500 words