Geographic Information Retrieval (GIR) community has generally accepted the thesis that both thematic and geographic aspect of documents may be useful for GIR. This paper describes a preliminary experiment exploring this thesis by seperately indexing/searching geographical relevant-terms (place names, geo-spatial relations, geographic concepts and geographic adjacetives) extracted from reference document collection. Two indexes were created one for extracted geographic relevant-terms (i.e. document footprint) and one for reference document collections. Geo-Score and ThematicScore against document collection footprint and reference document collection respectively were combined through a linear interpolation to obtained the nal score for document relevance ranking. We used several freely available geographic resources { Wikipedia, World-Gazetteer, GEOnet Name Server (GNS), and WordNet. Apache Lucene was used as an indexing and search platform while Alias-I LingPipe was used to detect geographic named entities (GNEs), and other geo-relevant concepts and terms in documents. We submitted runs for monolingual English task, and our system achieved mean average precision (MAP) of 0:1690 to 0:2194. No signi cant improvement was observed through geographic query expansion.
Geographic Information Retrieval (GIR) concerns the retrieval of information involving some kind of spatial awareness. Geographic information pervades many documents, and therefore, geographic references may be important for Information Retrieval (IR). Additionally, many documents contain geographic references expressed in multiple languages which may or may not be the same as the query language.
To perform GIR both themetic (non-geographic aspect) and geographic aspect of documents require consideration. In order to approach this thesis we derive for each document in the collection a corresponding document footprint containing place names (e.g. Uganda), geo-spatial relations (e.g. west of), geographic concepts (e.g. country) and geographic adjectives (e.g. Ugandan). The document footprint and reference document collection provided were separately indexed and searched. Geo-Score and Thematic-Score against document collection footprint and reference document collection were combined through a linear interpolation to obtained the nal score to perform document relevance ranking. Queries were performed for geographic relevant terms identi ed in topics against document collection footprints and reference document collection provided to investigate impart of geo-references and geo-relevant terms for GIR.
Freely available geographic resources (from: Wikipedia1, World-Gazetteer2, GEOnet Name Server3 (GNS), WordNet4) were consulted for query geographic reference expansion. Apache Lucene5 was used as an indexing and search platform while Alias-I LingPipe6 was used to detect geographic named entities (GNEs) and other geo-relevant concepts and terms in documents. 2
GeoCLEF evaluation track was run for the rst time at CLEF 2005 to evaluate retrieval of multilingual documents with an emphasis on geographic search [Gey et al, 2005]. As GeoCLEF 2005, GeoCLEF 2006 outline the following challenges to GIR in a multilingual environment: (1) translation of locations (e.g. Uganda (EN) to Oeganda (NL)), (2) resolution of geographic reference ambiguities (e.g. "Jack London" the author not a place; South Yorkshire and S. Yorks refer to the same place), (3) resolution of spatial ambiguity (e.g. She eld in UK or USA), (4) nding or creating suitable multilingual geographic knowledge base, and (5) combining both text and spatial retrieval methods. The speci c aims for GeoCLEF 2006 are: (1) compare methods of query translation, (2) query expansion, (3) translation of geographical references, (4) use of text and spatial retrieval methods separately or combined, and (5) retrieval models and indexing methods.
GeoCLEF 2006 consists of document collections in English, German, Portuguese and Spanish, and 25 search topics in these languages. The tasks for GeoCLEF 2006 are: (1) monolingual retrieval { retrieval where the topic and document languages are the same, and (2) bilingual retrieval { cross-language retrieval where the topic language is di erent from the document language, i.e. X ! fDE, EN, ES, PTg. For each document language, participants may submit the results of up to 10 runs: 5 monolingual and 5 bilingual. Two of these runs are required: (1) Title-Description { where the search queries are created using only the contents of the Title and Desc tags of the topic, and (2) Title-Description-Narrative { where the search queries are created using the contents of the Title, Desc and Narr tags from the topic. The Narrative tag contains a more comprehensive description of the information request de ned by the topic, including speci cs about the geography of the topic such as a list of desired cities, states, countries, rivers or latitudes and longitudes. An example search topic is depicted below: <top> <num>GC027</num> <EN-title>Cities within 100km of Frankfurt</EN-title> <EN-desc>Documents about cities within 100 kilometers of the city of Frankfurt in Western Germany</EN-desc> <EN-narr>Relevant documents discuss cities within 100 kilometers of Frankfurt am 1http://www.wikipedia.org 2http://www.world-gazetteer.com 3http://earthinfo.nga.mil/gns/html 4http://wordnet.princeton.edu 5http://jakarta.apache.org/lucene 6http://alias-i.com/lingpipe
Main Germany, latitude 50.11222, longitude 8.68194. To be relevant the document must describe the city or an event in that city. Stories about Frankfurt itself are not relevant</EN-narr> </top> 3
GeoCLEF 2005 [Gey et al, 2005] featured several approaches to GIR: (1) conventional IR systems, (2) geographic named entity recognition, classi cation and real world resolution, (3) creation and expansion of geographic knowledge base (e.g. name variants, multilingual), (4) query expansion strategies { blind feedback, addition of proper names, geographic reference expansion using hierarchical information contained in GKB, (5) geo-spatial query restriction strategies { minimum bounding box based, geo-scope based, and (6) topic translation strategies mainly employing usage of o -shelf software packages.
Larson [
Ferres et al [
Hughes [
Buscaldi et al [
Berkeley group 2 [Gey and Vivien, 2005] retrieval strategy involved query augmentation with blind feedback. Another feature of their approach is the augmentation of query information by inclusion of location-speci c tags and expansion of geographic references (e.g. Europe to individual country names). The blind feedback approach adds 30 top-ranked terms to the query from the top 20 ranked documents of intial ranking. Manual expansion of geographic references proved disastrous to retrieval performance. Addition of concepts and location imformation improved retrieval precisions across. Most improvement was achieved with blind feedback by adding mostly proper names and word variations and very few irrelevant words that won't distort the search towards another direction. 4
We are participating in GeoCLEF evaluation track at CLEF 2006 for the rst time. The main
motivation for our participation is to experiment with both thematic and geographic aspect of a
document for GIR. In this section we describe our approach, system architecture and resources
used. Our appraoch borrows techniques from (Larson [
We used the World Gazetteer, GEOnet Names Server (GNS), Wikipedia and WordNet as the bases for our Geographic Knowledge Base (GKB) for several reasons { free availability, multilingual (English, Germany, Portuguese and Spanish), most popular and major places, etc. Volcano active region, European river, Atlantic Ocean ports/coast and European Wine processing region information were speci cally gathered from the Wikipedia. 4.1.2
Alias-I LingPipe was used to detect named entities (location, person and organisation), geographic concepts (continent, region, country, city, town, village, etc.), spatial relations (near, in, south of, north west, etc.) and locative adjectives (e.g. Ugandan).
We used a simple appraoch to geo-code identi ed geographic named entities (GNEs) presented in CLIN 2005 [Andogah, 2005]. The approach exploits location type (e.g. city, mountain) and hierarchy information integrated in GKB to ground GNEs. 4.1.4
Apache Lucene is a high-performance, full-featured text search engine library written entirely in Java. It is a technology suitable for nearly any application that requires full-text search, especially cross-platform. Lucene's default similarity measure is based on vector space model7 (VSM). 4.2
Documents were pre-processed using the Alias-I LingPipe to detect place names (e.g. Kampala), geographic concepts (e.g. city), spatial relations (e.g. west of) and adjectives referring to things or people or language connected to a place (e.g. Ugandan).
Candidate locations for detected place names is obtained from our GKB. Place names are resolved to their respective locations using a simple geo-coding approach exploiting location type and hierarchical information present in GKB. The preliminary experimental result of geo-coding approach used here was reported in [Andogah, 2005]8. However, due to time limitation geo-coding task was not experimented as planned, instead we assume that all geo-relevant terms detected 7The vector space model (VSM) is an algebraic model used for information ltering and information retrieval. It represents natural language documents in a formal manner by the use of vectors in a multi-dimensional space. http : ==en:wikipedia:org=wiki=V ector space model
8http://www.science.uva.nl/events/CLIN2005/Program/Abstracts/abstract-andogah.html in a document will some-how relate or point to a speci c geographic region/scope or geographic concept in the discourse. 4.3
Footprint document collection repository derived from document collection was created. Footprint documents contain geo-relevant terms such as place name, geographic concepts, spatial relations, locative adjectives plus their respective term frequency as depicted below.
<GeographicTermFrequency docid="GH950102-000006"> <GT name="east" tf="2" gtt="SPR" /> <GT name="america" tf="1" gtt="LOC" /> <GT name="new york" tf="2" gtt="LOC" /> <GT name="bu alo" tf="3" gtt="LOC" /> <GT name="orlando magic" tf="1" gtt="LOC" /> <GT name="american" tf="4" gtt="GAD" /> <GT name="texas" tf="1" gtt="LOC" /> <GT name="bu alo jills" tf="1" gtt="LOC" /> <GT name="city" tf="1" gtt="GCO" /> </GeographicTermFrequency>
Derived footprint documents were indexed using Lucene along side index of reference document collection provided for the experiment (see [Table 1] for details).
TOPIC Formulation: 1. TITLE-DESC Content 2. TITLE-DESC-NARR Content Topic 3. TITLE-DESC-NARR Content geo-relevant-terms 4. TITLE-DESC-NARR Content geo-relevant-terms augmented with geo-references Mandatory runs 1 and 2 queries were formulated by topic TITLE-DESC (CLCGGeoEE1) and TITLE-DESC-NARR (CLCGGeoEE2) contents respectively. These queries were submitted to search Lucene index (Lucene eld content was searched) of GEO-CLEF 2006 document collection (see [Table 2] for index structure and [Figure 1] for system architecture). The mandatory queries perform general-purpose search of Lucene index returning the top 1,000 documents retrieve.
Our third run query was formulated by topic TITLE-DESC only (CLCGGeoEE5). The query was submitted to search Lucene index (Lucene eld nm was searched) of GEO-CLEF 2006 document collection footprints (see [Table 1] for index structure and [Figure 1] for system architecture), and the top 1,000 documents retrieved.
Our fourth run (CLCGGeoEE10) combine run 2 query result with result of querying Lucene index of GEO-CLEF 2006 document collection footprints for geo-relevant-terms extracted from topic TITLE-DESC-NARR. To combine the result of run 2 with result of querying Lucene index of document collection footprints we used the linear interpolation as described in [Leidner, 2005].
Score(d; q) = T hematicScore(d; q) + (1 )GeoScore(d; q) (1) For this experiment was set to 0:5.
Our fth run (CLCGGeoEE11) is similar to run four except that geo-revelant-terms extracted from topic TITLE-DESC-NARR were augmented with geo-references obtain from our GKB. For example, topic G033 geo-relevant-terms were augmented with the names of the major cities/towns/places within Ruhr area of Germany { Bochum, Bottrop, Dortmund, Duisburg, Essen, Gelsenkirchen, Hagen, Hamm, Herne, Mlheim, Oberhausen, Recklinghausen, Ennepe-Ruhr, Unna, Wesel, Mlheim an der Ruhr, Mulheim an der Ruhr. 5 5.1
, ,-.. ..
.. '(
! ) '( ! ) ! ! ! " ! ! " " # ! / # ' & )
+ ( ! "! " !! ! ! " " ! ! ! ! ! ! "! " ! " #$%& #$#"" %% #$#"" %% #$#"" %% #$#"" %% #$#"" %% " #$%& #$#"" %% #$#"" %% #$#"" %% #$#"" %% #$#"" %% and CLCGGeoEE5 use topic TITLE-DESC (but querying di erent document collection content), CLCGGeoEE5 performed better.
CLCGGeoEE10 & CLCGGeoEE11 schemes use linear interpolation (with set to 0:5) to combine result of query against reference document collection and document collection footprint indexes. We note that CLCGGeoEE10 performed poorly while CLCGGeoEE11 performed better. Several factors might have in uenced the performance of these schemes (CLCGGeoEE5, CLCGGeoEE10 & CLCGGeoEE11): predominance of geographic concepts and spatial-relationship quali ers such as country, city, southern, west, etc. both in the query and document footprints at expense of place names, and thereby shifting query result in wrong direction propagating irrelevant documents to the top value of 0:5 asigned to in linear interpolation [Equation (1)] above might have tilted result by asigning higher scores to documents retrieved from reference document collection or vice versa, and thereby propagating irrelevant documents to the top in the nal rank not all documents were indexed as our adopted geographic named entity tagger (Alias-i Lingpipe) reported content error for certain les while processing reference collection les. As a result 51,525 Glasgow Heralds documents were indexed out of 56,472 and 112,552 LA Times documents were indexed out of 113,005. This might have had a considerable impart on query result as 5,400 documents (which might have contained relevant documents) were left out.
The results of our submitted runs raised several pertinent questions for future investigation: extend to which geographic aspect of document in uence GIR result: (1) querying topic geographic aspect against reference document collection, (2) querying topic non-geographic aspect against reference document an appropriate value for in linear interpolation [Equation (1)] above for GIR an appropriate document collection footprint indexing strategy improve geographic named entity recognition, classi cation and real world resolution geographic query expansion strategies { blind feedback, addition of place names, expansion through hierarchical information contain in GKB. 6
We employed a strategy of separately indexing document footprint along side index of reference document, and combine query results of the two indexes through linear interpolation. Our approach yielded an average result as compared to overall GeoCLEF 2006 result on monolingual English task. A number of pertinent questions were raised for future investigation which we hope to address and integrate in our system. Analysis of individual topic performance to give further insight in our approach is under way. 7
This work is supported by NUFFIC within the framework of Netherlands Programme for the Institutional Strengthening of Post-secondary Training Education and Capacity (NPT) under project titled "Building a sustainable ICT training capacity in the public universities in Uganda".