Joint Discourse-aware Concept Disambiguation and Clustering

This thesis addresses the tasks of concept disambiguation and clustering. Concept disambiguation is the task of linking common nouns and proper names in a text – henceforth called mentions – to their corresponding concepts in a predefined inventory. Concept clustering is the task of clustering mentions, so that all mentions in one cluster denote the same concept. In this thesis, we investigate concept disambiguation and clustering from a discourse perspective and propose a discourse-aware approach for joint concept disambiguation and clustering in the framework of Markov logic. The contributions of this thesis are fourfold: Joint Concept Disambiguation and Clustering. In previous approaches, concept disambiguation and concept clustering have been considered as two separate tasks (Schütze, 1998; Ji & Grishman, 2011). We analyze the relationship between concept disambiguation and concept clustering and argue that these two tasks can mutually support each other. We propose the – to our knowledge – first joint approach for concept disambiguation and clustering. Discourse-Aware Concept Disambiguation. One of the determining factors for concept disambiguation and clustering is the context definition. Most previous approaches use the same context definition for all mentions (Milne & Witten, 2008b; Kulkarni et al., 2009; Ratinov et al., 2011, inter alia). We approach the question which context is relevant to disambiguate a mention from a discourse perspective and state that different mentions require different notions of contexts. We state that the context that is relevant to disambiguate a mention depends on its embedding into discourse. However, how a mention is embedded into discourse depends on its denoted concept. Hence, the identification of the denoted concept and the relevant concept mutually depend on each other. We propose a binwise approach with three different context definitions and model the selection of the context definition and the disambiguation jointly. Modeling Interdependencies with Markov Logic. To model the interdependencies between concept disambiguation and concept clustering as well as the interdependencies between the context definition and the disambiguation, we use Markov logic (Domingos & Lowd, 2009). Markov logic combines first order logic with probabilities and allows us to concisely formalize these interdependencies. We investigate how we can balance between linguistic appropriateness and time efficiency and propose a hybrid approach that combines joint inference with aggregation techniques. Concept Disambiguation and Clustering beyond English: Multi- and Cross-linguality. Given the vast amount of texts written in different languages, the capability to extend an approach to cope with other languages than English is essential. We thus analyze how our approach copes with other languages than English and show that our approach largely scales across languages, even without retraining. Our approach is evaluated on multiple data sets originating from different sources (e.g. news, web) and across multiple languages. As an inventory, we use Wikipedia. We compare our approach to other approaches and show that it achieves state-of-the-art results. Furthermore, we show that joint concept disambiguating and clustering as well as joint context selection and disambiguation leads to significant improvements ceteris paribus