Abstract:
|
This paper describes Knowledge-Based and Data-Driven approaches we have followed for generic Textual Georeferencing of Informal Documents. Textual georeferencing consists in assigning a set of geographical coordinates to formal (news, reports,..) or informal (blogs, social networks, chats, tagsets,...) texts and documents. The system presented in this paper has been designed to deal with informal documents from social sites. The paper describes four Georeferencing approaches, experiments, and results at the MediaEval 2014 Placing Task (ME2014PT) evaluation, and posterior experiments. The task consisted of predicting the most probable geographical coordinates of Flickr images and videos using its visual, audio and metadata associated features. Our approaches used only Flickr users textual metadata annotations and tagsets. The four approaches used for this task were: 1) a Geographical Knowledge-Based (GeoKB) approach that uses Toponym Disambiguation heuristics, 2) the Hiemstra Language Model (HLM), TFIDF and BM25 Information Retrieval (IR) approaches with Re-Ranking, 3) a combination of the GeoKB and the IR models with Re-Ranking (GeoFusion). 4) a combination of the GeoFusion with a HLM model derived from the English Wikipedia georeferenced pages. The HLM approach with Re-Ranking showed the best performance in accuracy within a margin of distance errors ranging from 10m to 1km. The GeoFusion approaches achieved the best results in accuracies from 10km to 5,000km. Both approaches achieved state-of-the-art results at ME2014PT evaluation and posterior experiments, including the best results for distance accuracies of 1000km and 5,000km in the task where only the official training dataset can be used to predict the coordinates. |