Knowledge-based and data-driven approaches for geographical information access

Zusammenfassung

Geographical Information Access (GeoIA) can be defined as a way of retrieving information from textual collections that includes the automatic analysis and interpretation of the geographical constraints and terms present in queries and documents. This PhD thesis presents, describes and evaluates several heterogeneous approaches for the following three GeoIA tasks: Geographical Information Retrieval (GIR), Geographical Question Answering (GeoQA), and Textual Georeferencing (TG). The GIR task deals with user queries that search over documents (e.g. ¿vineyards in California¿) and the GeoQA task treats questions that retrieve answers (e.g. ¿What is the capital of France?¿). On the other hand, TG is the task of associate one or more georeferences (such as polygons or coordinates in a geodetic reference system) to electronic documents. Current state-of-the-art AI algorithms are not yet fully understanding the semantic meaning and the geographical constraints and terms present in queries and document collections. This thesis attempts to improve the effectiveness results of GeoIA tasks by: 1) improving the detection, understanding, and use of a part of the geographical and the thematic content of queries and documents with Toponym Recognition, Toponym Disambiguation and Natural Language Processing (NLP) techniques, and 2) combining Geographical Knowledge-Based Heuristics based on common sense with Data-Driven IR algorithms. The main contributions of this thesis to the state-of-the-art of GeoIA tasks are: 1) The presentation of 10 novel approaches for GeoIA tasks: 3 approaches for GIR, 3 for GeoQA, and 4 for Textual Georeferencing (TG). 2) The evaluation of these novel approaches in these contexts: within official evaluation benchmarks, after evaluation benchmarks with the test collections, and with other specific datasets. Most of these algorithms have been evaluated in international evaluations and some of them achieved top-ranked state-of-the-art results, including top-performing results in GIR (GeoCLEF 2007) and TG (MediaEval 2014) benchmarks. 3) The experiments reported in this PhD thesis show that the approaches can combine effectively Geographical Knowledge and NLP with Data-Driven techniques to improve the efectiveness measures of the three Geographical Information Access tasks investigated. 4) TALPGeoIR: a novel GIR approach that combines Geographical Knowledge ReRanking (GeoKR), NLP and Relevance Feedback (RF) that achieved state-of-the-art results in official GeoCLEF benchmarks (Ferrés and Rodríguez, 2008; Mandl et al., 2008) and posterior experiments (Ferrés and Rodríguez, 2015a). This approach has been evaluated with the full GeoCLEF corpus (100 topics) and showed that GeoKR, NLP, and RF techniques evaluated separately or in combination improve the results in MAP and R-Precision effectiveness measures of the state-of-the-art IR algorithms TF-IDF, BM25 and InL2 and show statistical significance in most of the experiments. 5) GeoTALP-QA: a scope-based GeoQA approach for Spanish and English and its evaluation with a set of questions of the Spanish geography (Ferrés and Rodríguez, 2006). 6) Four state-of-the-art Textual Georeferencing approaches for informal and formal documents that achieved state-of-the-art results in evaluation benchmarks (Ferrés and Rodríguez, 2014) and posterior experiments (Ferrés and Rodríguez, 2011; Ferrés and Rodríguez, 2015b).