Self-Annotation for fine-grained geospatial relation extraction

Proceedings of the 23rd International Conference on Computational Linguistics (Coling 2010), pages 80–88, Beijing, August 2010 Self-Annotation for Fine-Grained Geospatial Relation Extraction Andre Blessing Hinrich Sch¨utze Institute for Natural Language Processing Universit¨at Stuttgart ner@ifnlp.org Abstract A great deal of information on the Web is represented in both textual and structured form. The structured form is machine- readable and can be used to augment the textual data. We call this augmentation – the annotation of texts with relations that are included in the structured data – self-annotation. In this paper, we intro- duce self-annotation as a new supervised learning approach for developing and im- plementing a system that extracts fine- grained relations between entities. The main benefit of self-annotation is that it does not require manual labeling. The in- put of the learned model is a represen- tation of the free text, its output struc- tured relations. Thus, the model, once learned, can be applied to any arbitrary free text. We describe the challenges for the self-annotation process and give re- sults for a sample relation extraction sys- tem. To deal with the challenge of fine- grained relations, we implement and eval- uate both shallow and deep linguistic anal- ysis, focusing on German. 1 Introduction In the last years, information extraction has be- come more important in domains like context- aware systems (e.g. Nexus (D¨urr et al., 2004)) that need a rich knowledge base to make the right de- cisions in different user contexts. Geospatial data are one of the key features in such systems and need to be represented on different levels of de- tail. Data providers do not cover all these lev- els completely. To overcome this problem, fine- grained information extraction (IE) methods can be used to acquire the missing knowledge. We define fine-grained IE as methods that recognize entities at a finer grain than standard categories like person, location, and organization. Further- more, the quality of the data in context-aware sys- tems plays an important role and updates by an in- formation extraction component can increase the overall user acceptance. For both issues an information extraction sys- tem is required that can handle fine-grained rela- tions, e.g., “X is a suburb of Y” or “the river X is a tributary of Y” – as opposed to simple con- tainment. The World Wide Web offers a wealth of information about geospatial data and can be used as source for the extraction task. The extraction component can be seen as a kind of sensor that we call text senor (Blessing et al., 2006). In this paper, we address the problem of de- veloping a flexible system for the acquisition of relations between entities that meets the above desiderata. We concentrate on geospatial entities on a fine-grained level although the approach is in principle applicable to any domain. We use a supervised machine learning approach, includ- ing several features on different linguistic lev- els, to build our system. Such a system highly depends on the quality and amount of labeled …

Content truncated to 500 words