This paper describes the approach of the University of Groningen to GeoCLEF task for CLEF 2007. We used geographic scope based approach to rank documents. This paper describes non-geographic similarity, geographic similarity and combined similarity measures employed to approach GeoCLEF task for CLEF 2007. The motivation for our participation was to test geographic scope (geo-scope) based relevance ranking for geographic information retrieval (GIR). We participated in monolingual English task and our evaluation result shows no significant improvement for geo-scope based approach. Geographic Knowledge Base. We used the World Gazetteer1, GEOnet Names Server2 (GNS), Wikipedia3 and WordNet4 as the bases for our Geographic Knowledge Base (GKB) for several reasons: free availability, multilingual, coverage of most popular and major places, etc. Geographic Tagger. Alias-I LingPipe5 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).
2.1
Lucene Search Engine. Apache Lucene6 is a high-performance, full-featured text search engine library written entirely in Java. Lucene’s default similarity measure is derived from the vector space model (VSM). Lucene was used to index and search both indexes. 2.2
The reference document collection provided for experimentation was indexed using Lucene. Document HEADLINE and TEXT contents were combined to create document content for indexing (see Table 1 for details). Before indexing, the documents were processed with the Porter stemmer and the default Lucene English stopword list.
brary to perform non-geographic search. Lucene’s default similarity measure is derived from the vector space model (VSM). The Lucene similarity score formula combines several factors to determine the document score for a query [4]: N onSim(q, d) =
X tf (t in d) . idf (t) . bst . lN (t.f ield in d)
t in q
(
Geographic Similarity Measure. As in [2, 5], we chose to use geographic scopes assigned to queries and documents to perform geographic relevance ranking of documents. Our geographic scope resolver [1] assign multiple scopes to documents ranking the scopes from the most relevant to the least relevant, thereby associating a document with multiple scopes or associating a scope with several documents. We limit geographic scopes (geo-scopes) of population centers 6 http://jakarta.apache.org/lucene to continent, continent-directional (e.g. western Europe, eastern Africa, eastern Europe, etc.), country, country-directional (e.g. north-of Italy), province7, and province-directional (e.g. northern California) level.
Equation 2 depicts our geographic similarity measure formula between query q and document d: where;
GeoSim(q, d) = (SF × W T S 0 if SF > 0 otherwise otherwise W T S =
X pwt(q,s) × log(1 + wt(d,s)) and, where; Nq is the number of scopes in the query scope set, Nd is the number of scopes in the document scope set, N(d,q) is the number of document scopes present in query scope set, wt(q,s) is the weight assigned to scope s in query q by the scope resolver and wt(d,s) is the weight assigned to scope s in document d by the scope resolver. For a given query, Nq is invariable whilst Nd and N(d,q) vary per document retrieved. The motivation for designing Equation 3 is to mitigate effects of arbitrarily large variations of Nd and Nq to a reasonable level. The values of Equation 3 are within 0.0 < SF ≤ 0.5. Equation 4 as arranged provides the best overall performance for Equation 2 (that is, applying square root weighting to wt(q,s) and logarithmic weighting to wt(d,s)).
Geo-IR Similarity Measure. The final similarity score formula is directly derived from Equation 1 and Equation 2 similarity score formulae.
Sim(q, d) = λT N onSim(q, d) + λG GeoSim(q, d) λT + λG = 1 where; λT is the non-geographic interpolation factor and λG is the geographic interpolation factor. Before the ranked list for non-geographic and geographic relevance ranking are linearly combined, their respective scores are normalized to [0, 1]. 2.4
This section describes the University of Groningen official runs for GeoCLEF
2007. In particular we describe which topic components are used for query
formulation and which similarity measures were used to perform relevance ranking.
7 Here a province represents first order administrative division of a country.
(
We performed geographic expansion on the following members of GROUP1 – 51, 59, 60, 61, 63, 65, 66, 70. The motivation for geographic expansion on these topics is that they lack sufficient geographic information or the geographic information provided are too ambiguous. For example, topic 59 is expanded by adding the names of major cities in Bolivia, Columbia, Ecuador and Peru. The Lucene boost factor of 0.45F is assigned to placenames for geographic query expansion while the boost factor for placenames in the original query is left at the default value of 1.0F.
Members of GROUP2 are topics – 56, 67, 68, 72. These topics fall under geographic subject with non-geographic restriction with exception of topic 72 which is more complex. Resolving geographic scope of these topics to a specific place is a non trivial undertaking. The most reasonable scope for these topics is geographic subject scope such as lake, river, beach, city, etc. For example, topic 56 concern documents with scope lake.
for these runs are formulated by the content of topic TITLE (T), TITLE-DESC (TD) and TITLE-DESC-NARR (TDN) tags respectively. GROUP1 topics are ranked according to Equation 5 with λT = 1.0 and λG = 0.0. GROUP2 topics are ranked according to Equation 1. However, the query for CLCGGeoEETDN00 was mistakenly formulated by the content of topic TITLE instead of TDN.
CLCGGeoEETDN00P is CLCGGeoEETDN00 with query formulated by topic TDN tag content.
CLCGGeoEETDN01. The query for this run should have been formulated by the content of topic TDN tags, however, the official result submitted erroneously used TITLE tag content. GROUP1 topics are ranked according to Equation 5 with λT = 0.85 and λG = 0.15. GROUP2 topics are ranked according to Equation 1.
CLCGGeoEETDN01P is CLCGGeoEETDN01 with query formulated by topic TDN tag content.
CLCGGeoEETDN01B. The query for this run should have been formulated by the content of topic TDN tags, however, the official result submitted erroneously used TITLE tag content. GROUP1 topics are ranked according to Equation 5 with λT = 0.85 and λG = 0.15.
For GROUP2 topics, we scan each document retrieved and ranked according to Equation 1 for geographic types (geo-types) as well as determine the geotypes of geographic names found in the documents. Each geo-type found in the document is assigned a weight. For documents containing query geo-type, we add the geo-type weight to Lucene score and then re-rank documents based on the new score.
CLCGGeoEETDN01BP is CLCGGeoEETDN01B with query formulated by topic TDN tag content. 3
Our future work will focus on investigating: (
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”.