Jointly Disambiguating and Clustering Concepts and Entities with Markov Logic

Proceedings of COLING 2012: Technical Papers, pages 815–832, COLING 2012, Mumbai, December 2012. Jointly Disambiguating and Clustering Concepts and Entities with Markov Logic Angela Fahrni1 Michael Strube1 (1) Heidelberg Institute for Theoretical Studies, Schloss-Wolfsbrunnenweg 35, 69118 Heidelberg, Germany Angela.Fahrni@h-its.org, Michael.Strube@h-its.org ABSTRACT We present a novel approach for jointly disambiguating and clustering known and unknown concepts and entities with Markov Logic. Concept and entity disambiguation is the task of identifying the correct concept or entity in a knowledge base for a single- or multi-word noun (mention) given its context. Concept and entity clustering is the task of clustering mentions so that all mentions in one cluster refer to the same concept or entity. The proposed model (1) is global, i.e. a group of mentions in a text is disambiguated in one single step combining various global and local features, and (2) performs disambiguation, unknown concept and entity detection and clustering jointly. The disambiguation is performed with respect to Wikipedia. The model is trained once on Wikipedia articles and then applied to and eval- uated on different data sets originating from news papers, audio transcripts and internet sources. KEYWORDS: Word Sense Disambiguation. 815 1 Introduction Recent advances in knowledge extraction from resources such as Wikipedia have allowed to create various large-scale knowledge bases and concept networks such as Yago (Suchanek et al., 2008), DBpedia (Bizer et al., 2009) or WikiNet (Nastase and Strube, 2012). To exploit the wealth of world knowledge in these resources for natural language processing tasks such as information extraction, text segmentation or summarization, words and phrases in a document first need to be linked to the relevant entries in the respective knowledge base, i.e. to be disambiguated. This problem has been tackled quite successfully by systems such as WikipediaMiner (Milne and Witten, 2008), which are mainly based on word sense disambiguation techniques (Agirre and Edmonds, 2006; Navigli, 2009) thus giving research on word sense disambiguation a new spin. Concept and entity disambiguation is the task of identifying the correct concept or entity in a knowledge base for a single- or multi-word noun (mention) given its context.1 In this paper we disambiguate with respect to the English Wikipedia and consider each article as a concept. One advantage of linking to Wikipedia is that the internal hyperlinks can be used as training data. Concept disambiguation models the relation between mentions and concepts (Figure 1a). For instance, the system needs to identify if the mention crocodile in the first text points to AMERICAN CROCODILES (ANIMAL), to CROCODILE (LOCOMOTIVE) or to the person RENÉ LACOSTE (TENNIS PLAYER) whose nick name is Crocodile. We define concept disambiguation as the task of disambiguating both common nouns such as crocodile or biologist and proper nouns such as FLORIDA or STATES. While the disambiguation of common nouns is usually called word sense disambiguation (WSD), the disambiguation of proper nouns is also known as entity linking. Although most of the work in concept disambiguation and WSD assumes that the knowledge base is complete, several studies show that …

Content truncated to 500 words