Unsupervised Coreference Resolution by Utilizing the Most Informative Relations

Proceedings of COLING 2014, the 25th International Conference on Computational Linguistics: Technical Papers, pages 644–655, Dublin, Ireland, August 23-29 2014. Unsupervised Coreference Resolution by Utilizing the Most Informative Relations Nafise Sadat Moosavi and Michael Strube Heidelberg Institute for Theoretical Studies gGmbH Schloss-Wolfsbrunnenweg 35 69118 Heidelberg, Germany {nafise.moosavi|michael.strube}@h-its.org Abstract In this paper we present a novel method for unsupervised coreference resolution. We introduce a precision-oriented inference method that scores a candidate entity of a mention based on the most informative mention pair relation between the given mention entity pair. We introduce an infor- mativeness score for determining the most precise relation of a mention entity pair regarding the coreference decisions. The informativeness score is learned robustly during few iterations of the expectation maximization algorithm. The proposed unsupervised system outperforms existing unsupervised methods on all benchmark data sets. 1 Introduction Due to the advent of the internet, the world wide web, social media, the electronic distribution of infor- mation and new means of communication, the amount of text available in many different languages is rising. Natural language processing (NLP) is in charge of automatic processing this growing data. NLP research has mainly focused on English and very few other languages. Therefore there is a rich set of annotated corpora for linguistic analysis tasks for these languages. However, there are no such corpora for thousands of other languages. Since unsupervised methods do not require annotated data for learning a model, employing unsupervised methods has become a popular and important area of research in NLP. In this paper, we propose a new precision oriented method for unsupervised coreference resolution. Our method evaluates the candidate entities of mentions based on the most precise relation of each mention and its candidate entity. Though we develop and evaluate our method for the English language, we intend to apply it to low resource languages in the future. Common coreference resolution approaches rely on a combination of different features for each de- cision (for an overview over such approaches, see Ng (2010)). However, a few approaches break down this combination having precision in mind (Baldwin, 1997; Zhou and Su, 2004; Haghighi and Klein, 2009; Lee et al., 2013). The idea of starting with high precision knowledge is used in various NLP tasks including parsing (Borghesi and Favareto, 1982), word alignment (Brown et al., 1993), and named en- tity classification (Collins and Singer, 1999) with different names like “islands of reliability”, “stepping stones”, and “cautiousness”. Lee et al. (2013) is a successful recent work that implements this idea as “sieve architecture”. Lee et al. (2013) first decide on the basis of more precise features, and then they extend these decisions by using less precise features in later sieves. In this system less precise knowledge is used for extending the decisions made by high precision knowledge. Our proposed inference method goes in the same direction but in a different way. The probability of each coreference decision is computed based on a single relation of a mention-entity. This single relation is the …

Content truncated to 500 words