The participation of the University of Hildesheim focused on the monolingual German and English tasks of GeoCLEF 2007. Based on the results of GeoCLEF 2005 and GeoCLEF 2006, the weighting and expansion of geographic named entities (NE) and Blind Relevance Feedback were combined. This year an improved model for German Named Entity Recognition was evaluated.
1 Introduction
2 Geographic Retrieval System
The system we augmented for this experimentation with (geographic) NEs in GIR is based on a retrieval system
applied to ad-hoc retrieval in previous CLEF campaigns [
Named Entity Recognition was carried out with the open source machine learning tool LingPipe3, which identifies named entities and classifies them into the categories Person, Organization, Location and Miscellaneous according to a trained statistical model.
HiMoDeNe2 HiMoDeNe2Na HiMoDeNe3 HiMoEnBase HiMoEnNe HiMoEnNaNe HiMoEnNe2
HiMoDeNe1 HiMoDeNe2 HiMoDeNe2Na HiMoDeNe3 HiMoEnBase HiMoEnNe HiMoEnNaNe HiMoEnNe2 German German German German English English English English
German German German German English English English English x x x x After experimentation with the GeoCLEF data of 2006 we submitted runs differing in parameters and query processing steps.
Run descriptions and results measured as Mean Average Precision (MAP) are shown in Table 1 for submitted monolingual runs and in Table 2 for the corresponding results with the training topics of 2006. With the training topics of 2006 best results were made expanding the query with 40 geographic terms from the best 30 documents giving each a relative weight of 0.2 compared to the rest of the query (for German) and using 20 terms from top5 documents with a relative weight of 0.5 for English (Table 2). While in the case of the English topics this hold true for the submitted runs, for German topics the base run without NER performed best (Table 1).
The worse results for the English topics indicate more difficult topics (concerning our retrieval system) for 2007. With the German results remaining on almost the same level, the optimised NER-model for German seems to improve retrieval quality.
Summing up, we could not find a substantial positive impact of additional geographic information, but the effect of investment in optimizing the Geo-NE model seems to be positive. 5
Conclusion and Outlook Optimised Geo-NE models seem to have positive effect on retrieval quality for monolingual tasks. For future experiments, we intend to integrate geographic ontologies to expand entities with neighbouring places, villages and regions. Furthermore we will integrate Wikipedia as translation tool for Geo-NEs to participate in multilinual tasks of GeoCLEF in the future.
0.2-30-40 (nm) 0.15-30-60 (nm) 1.0-10-4 (nm)
0.5-5-20 0.5-5-20 2-10-3
0.2-30-40 (om) 0.2-30-40 (nm) 0.15-30-60 (nm) 1.0-10-4 (nm)
0.5-5-20 0.5-5-20 2-10-3