Illinois-Coref: The UI System in the CoNLL-2012 Shared Task

Proceedings of the Joint Conference on EMNLP and CoNLL: Shared Task, pages 113–117, Jeju Island, Korea, July 13, 2012. c⃝2012 Association for Computational Linguistics Illinois-Coref: The UI System in the CoNLL-2012 Shared Task Kai-Wei Chang Rajhans Samdani Alla Rozovskaya Mark Sammons Dan Roth University of Illinois at Urbana-Champaign {kchang10|rsamdan2|rozovska|mssammon|danr}@illinois.edu Abstract The CoNLL-2012 shared task is an extension of the last year’s coreference task. We partici- pated in the closed track of the shared tasks in both years. In this paper, we present the im- provements of Illinois-Coref system from last year. We focus on improving mention detec- tion and pronoun coreference resolution, and present a new learning protocol. These new strategies boost the performance of the system by 5% MUC F1, 0.8% BCUB F1, and 1.7% CEAF F1 on the OntoNotes-5.0 development set. 1 Introduction Coreference resolution has been a popular topic of study in recent years. In the task, a system requires to identify denotative phrases (“mentions”) and to cluster the mentions into equivalence classes, so that the mentions in the same class refer to the same en- tity in the real world. Coreference resolution is a central task in the Natural Language Processing research. Both the CoNLL-2011 (Pradhan et al., 2011) and CoNLL- 2012 (Pradhan et al., 2012) shared tasks focus on resolving coreference on the OntoNotes corpus. We also participated in the CoNLL-2011 shared task. Our system (Chang et al., 2011) ranked first in two out of four scoring metrics (BCUB and BLANC), and ranked third in the average score. This year, we further improve the system in several respects. In Sec. 2, we describe the Illinois-Coref system for the CoNLL-2011 shared task, which we take as the baseline. Then, we discuss the improvements on mention detection (Sec. 3.1), pronoun resolu- tion (Sec. 3.2), and learning algorithm (Sec. 3.3). Section 4 shows experimental results and Section 5 offers a brief discussion. 2 Baseline System We use the Illinois-Coref system from CoNLL-2011 as the basis for our current system and refer to it as the baseline. We give a brief outline here, but fo- cus on the innovations that we developed; a detailed description of the last year’s system can be found in (Chang et al., 2011). The Illinois-Coref system uses a machine learn- ing approach to coreference, with an inference pro- cedure that supports straightforward inclusion of do- main knowledge via constraints. The system first uses heuristics based on Named Entity recognition, syntactic parsing, and shallow parsing to identify candidate mentions. A pair- wise scorer w generates compatibility scores wuv for pairs of candidate mentions u and v using ex- tracted features φ(u, v) and linguistic constraints c. wuv = w · φ(u, v) + c(u, v) + t, (1) where t is a threshold parameter (to be tuned). An inference procedure then determines the optimal set of links to retain, incorporating constraints that may override the classifier prediction for a given mention pair. A post-processing step removes mentions in singleton clusters. Last year, we found that …

Content truncated to 500 words