Experiments with geographic knowledge for information extraction

Experiments with geographic knowledge for information extraction Dimitar Manov, Atanas Kiryakov, Borislav Popov Ontotext Lab, Sirma AI Ltd 38A Christo Botev Blvd, Sofia 1000, Bulgaria {mitac,naso,borislav}@sirma.bg Kalina Bontcheva, Diana Maynard, Hamish Cunningham University of Sheffield Regent Court, 211 Portobello St., Sheffield S1 4DP, UK {kalina,diana,hamish}@dcs.shef.ac.uk Abstract Here we present work on using spatial knowl- edge in conjunction with information extrac- tion (IE). Considerable volume of location data was imported in a knowledge base (KB) with entities of general importance used for seman- tic annotation, indexing, and retrieval of text. The Semantic Web knowledge representation standards are used, namely RDF(S). An exten- sive upper-level ontology with more than two hundred classes is designed. With respect to the locations, the goal was to include the most im- portant categories considering public and tasks not specially related to geography or related ar- eas. The locations data is derived from num- ber of publicly available resources and com- bined to assure best performance for domain- independent named-entity recognition in text. An evaluation and comparison to high perfor- mance IE application is given. 1 Introduction Information Extraction (IE) research has focused mainly on the recognition of course-grained entities like Loca- tion, Organization, Person, etc. (Sundheim, 1998). The application of Information Extraction to new areas like the Semantic Web and knowledge management has posed new challenges, from which the most relevant here is the need for finer-grained recognition of entities, such as lo- cations. In this paper we present some experiments with build- ing a reusable knowledge base of locations which is used as a component into an IE system, instead of a location gazetteer. This work is part of the Knowledge and Infor- mation Management (KIM) platform and still undergoing development and refinement. With respect to coverage, the goal was to include the most important location categories for a wide range of ap- plications and tasks, not specially related to geography or related areas. The locations data is derived from a num- ber of publicly available resources and combined to as- sure best performance for named-entity recognition. An evaluation and comparison to high performance IE sys- tem using very small location gazetteers is given. One important aspect of our work is that we choose to create a knowledge base of locations, structured accord- ing to an ontology and having relations between them, in- stead of having somewhat flat structures of gazetteer lists found in other IE systems. While a knowledge base can be plugged into an IE system instead of a flat gazetteer, it also has several unique advantages: • the extra information, especially the transitive sub- RegionOf relation can be used for disambiguation and reasoning • the location entities in the text can be recognised at the right level of granularity for the target applica- tion (i.e., as Location or as Country, City, etc). • the ontology and knowledge base can be modified by the user and any changes are reflected immedi- ately in the output of the IE system. The paper is structured …

Content truncated to 500 words