Domain-sensitive Temporal Tagging for Event-centric Information Retrieval

by Jannik Strötgen

Institution: Universität Heidelberg
Department: The Faculty of Mathematics and Computer Science
Degree: PhD
Year: 2015
Record ID: 1103995
Full text PDF: http://www.ub.uni-heidelberg.de/archiv/18357


Temporal and geographic information is of major importance in virtually all contexts. Thus, it also occurs frequently in many types of text documents in the form of temporal and geographic expressions. Often, those are used to refer to something that was, is, or will be happening at some specific time and some specific place – in other words, temporal and geographic expressions are often used to refer to events. However, so far, event-related information needs are not well served by standard information retrieval approaches, which motivates the topic of this thesis: event-centric information retrieval. An important characteristic of temporal and geographic expressions – and thus of two components of events – is that they can be normalized so that their meaning is unambiguous and can be placed on a timeline or pinpointed on a map. In many research areas in which natural language processing is involved, e.g., in information retrieval, document summarization, and question answering, applications can highly benefit from having access to normalized information instead of only the words as they occur in documents. In this thesis, we present several frameworks for searching and exploring document collections with respect to occurring temporal, geographic, and event information. While we rely on an existing tool for extracting and normalizing geographic expressions, we study the task of temporal tagging, i.e., the extraction and normalization of temporal expressions. A crucial issue is that so far most research on temporal tagging dealt with English news-style documents. However, temporal expressions have to be handled in different ways depending on the domain of the documents from which they are extracted. Since we do not want to limit our research to one domain and one language, we develop the multilingual, cross-domain temporal tagger HeidelTime. It is the only publicly available temporal tagger for several languages and easy to extend to further languages. In addition, it achieves state-of-the-art evaluation results for all addressed domains and languages, and lays the foundations for all further contributions developed in this thesis. To achieve our goal of exploiting temporal and geographic expressions for event-centric information retrieval from a variety of text documents, we introduce the concept of spatio-temporal events and several concepts to "compute" with temporal, geographic, and event information. These concepts are used to develop a spatio-temporal ranking approach, which does not only consider textual, temporal, and geographic query parts but also two different types of proximity information. Furthermore, we adapt the spatio-temporal search idea by presenting a framework to directly search for events. Additionally, several map-based exploration frameworks are introduced that allow a new way of exploring event information latently contained in huge document collections. Finally, an event-centric document similarity model is developed that calculates document similarity on multilingual corpora…