Bootstrapping coreference resolution using word associations

Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics, pages 783–792, Portland, Oregon, June 19-24, 2011. c⃝2011 Association for Computational Linguistics Bootstrapping Coreference Resolution Using Word Associations Hamidreza Kobdani, Hinrich Sch¨utze, Michael Schiehlen and Hans Kamp Institute for Natural Language Processing University of Stuttgart kobdani@ims.uni-stuttgart.de Abstract In this paper, we present an unsupervised framework that bootstraps a complete corefer- ence resolution (CoRe) system from word as- sociations mined from a large unlabeled cor- pus. We show that word associations are use- ful for CoRe – e.g., the strong association be- tween Obama and President is an indicator of likely coreference. Association information has so far not been used in CoRe because it is sparse and difficult to learn from small labeled corpora. Since unlabeled text is readily avail- able, our unsupervised approach addresses the sparseness problem. In a self-training frame- work, we train a decision tree on a corpus that is automatically labeled using word associa- tions. We show that this unsupervised system has better CoRe performance than other learn- ing approaches that do not use manually la- beled data. 1 Introduction Coreference resolution (CoRe) is the process of find- ing markables (noun phrases) referring to the same real world entity or concept. Until recently, most ap- proaches tried to solve the problem by binary classi- fication, where the probability of a pair of markables being coreferent is estimated from labeled data. Al- ternatively, a model that determines whether a mark- able is coreferent with a preceding cluster can be used. For both pair-based and cluster-based models, a well established feature model plays an important role. Typical systems use a rich feature space based on lexical, syntactic and semantic knowledge. Most commonly used features are described by Soon et al. (2001). Most existing systems are supervised systems, trained on human-labeled benchmark data sets for English. These systems use linguistic features based on number, gender, person etc. It is a challenge to adapt these systems to new domains, genres and lan- guages because a significant human labeling effort is usually necessary to get good performance. To address this challenge, we pursue an unsuper- vised self-training approach. We train a classifier on a corpus that is automatically labeled using asso- ciation information. Self-training approaches usu- ally include the use of some manually labeled data. In contrast, our self-trained system is not trained on any manually labeled data and is therefore a com- pletely unsupervised system. Although training on automatically labeled data can be viewed as a form of supervision, we reserve the term supervised sys- tem for systems that are trained on manually labeled data. The key novelty of our approach is that we boot- strap a competitive CoRe system from association information that is mined from an unlabeled cor- pus in a completely unsupervised fashion. While this method is shallow, it provides valuable informa- tion for CoRe because it considers the actual iden- tity of the words in question. Consider the pair of markables (Obama, President). It is …

Content truncated to 500 words