The LogCLEF track was launched in 2009 with aim to analyse and classify the user's queries and had two tasks LAGI (Log Analysis and Geographic Query Identification) and LADS (Log Analysis for Digital Societies). In this edition from 2009, we built a system in order to participate in the LAGI task. The system uses GATE or Wikipedia like external resources in order to identify geographical entities in user's queries. Because, the results obtained using these resource are comparable, the main advantage of using GATE resources comes from the short duration of execution in comparison with using of Wikipedia resources. A brief description of our system is given in this paper.
1 Introduction 1. 2.
Tumba! - a Portuguese web search engine
The European Library (TEL) - on line search for materials in various
libraries in Europe.
Our system searches only in the subset in English logs (which are the majority of the
logs). Organizers force participating groups to respect some rules. The rules respected
by our system were:
1. A query is a geographical query if and only if it is bounded geographically.
2. A place term can be any country, a city or town, mountain, province or region
from GATE
Wikipedia. 4. A place term can occur in a title (of a book, movie, team, etc.), but the title itself (if a different text span from the place) is not to be tagged. 5. Capitalization (upper and lower case) in the query is ignored, as it is used inconsistently in the queries. 6. Wildcards ('*') are ignored. 7. If some words of a query can be interpreted as forming a phrase, this will be preferred over interpreting those words as isolated words put in the same query.
In order to participate in LAGI, additional to using Wikipedia resources offered by organizers, we built another resource with geographical name entities starting from GATE resources. This resource was loaded by our program in cache and it is used after that in identification of geographical resources. The Figure 1 presents the system architecture.
GATE resources
Portuguese Wikipedia
English Wikipedia
The test data consists from two files, one from Tumba! with 152 entries and one from TEL with 108 entries. In comparison the training files had 7 entries for Tumba! file and 9 entries for TEL file. The format for input data and for requested output data are presented below in Table 1 and in Table 2 and are taken from TEL training data. 875336 & 5431 & ("central europe") 828587 & 12840 & ("sicilia") 902980 & 482 & (creator all "casanova") 196270 & 5365 & ("casanova") 528968 & 190 & ("iceland*") 470448 & 8435 & ("iceland") 712725 & 5409 & ("cavan county ireland 1870") 875336 & 5431 & ("<place>central europe</place>") 828587 & 12840 & ("<place>sicilia</place>") 902980 & 482 & (creator all "casanova") 196270 & 5365 & ("casanova") 528968 & 190 & ("<place>iceland</place>*") 470448 & 8435 & ("<place>iceland</place>") 712725 & 5409 & ("<place>cavan county ireland</place> 1870") How we can see in above tables the aim is to add <place> </place> tags to user queries for geographical elements.
In order to build our resource with geographical entities we start from GATE resources and additional, we search on the web in order to add new similar entities. In separated runs, we load our resources or resources provide by organizers: page titles from Portuguese and English Wikipedia.
From GATE
Organizers offered to the participants’ two files with titles from Portuguese and English Wikipedia. In comparison with GATE, the number of entries from file with Portuguese titles from Wikipedia was 934.395 and the size on disk was 35.3 Mb, and the number of entries from file with English titles from Wikipedia was 6.996.744 and the size on disk was 282 Mb. In Table 4 are few lines from these files.
Examples <title>AmericanSamoa</title> <title>AppliedEthics</title> <title>AccessibleComputing</title> <title>Anarchism</title>… <title>Astronomia</title> <title>Astronomia e astrofísica</title> <title>América Latina</title> <title>Albino Forjaz de Sampaio</title>… 2.3
The main module load successively the GATE or Wikipedia resources in cache using a hash map in which the key is named entity itself, and the value is the number of words from initial named entity. After that using this cache we will try to identify in TEL and Tumba test data the geographical entities.
From beginning when we load our geographical resources in cache, we transform all characters from these entities in lower case. Additional, we split every name entity in components words and load also these separated words in our cache, but we specify the number of words from initial entity (in this way we will know if this key from our hash map come from a simple name entity or from a composed name entity). We will see how we will use this value when we try to identify geographical entities in user query. Even if we lose much time at resource loading, after that the operations of identification are very fast regardless of the number of user’s queries that we want to process.
In order to identify in test data geographical entities from our cache, we perform pre-processing of test data. The most important steps are: 1) In first step we parse the current line from test data and extract only the relevant text (actually this is the initial user query). 1.
2) 3)
Second step has the aim to ignore special characters like +, (, ), *, “, ” or white spaces (tab, space, return) in initial user query.
Third step transform all characters extracted in previous step in lower case characters. The result is a new form of user query, called from now new query.
The most important operation of main module is identification of the geographical entities in this new form of user query. From now the main question is: How we identify the geographical entities in this new query?
Initial we try to see if we have in our cache the new query itself. If YES, then we finished the process of identification of geographical entities and we skip to the next line in test data file. This is the case of the following line from Tel test data: 4752 & 11759 & ("portugal)" for which we have in our cache the all user query which is “Portugal” from GATE countries file.
If NO, then we try to split new user query in separated words if this is possible. When we have only one word in the new query we automatically skip to the next line in test data file. When we have more than one word, we apply the following steps: At step 1 every individual word is searched in hash map. If current word comes from a simple named entity we simply add “place” tags to it. This is the case of below line: 4892 & 5670 & mountain ranges)" ("climbing on the
Himalaya and other for which we have in our cache the separated word “Himalaya” from GATE mountains file.
At step 2 for every word searched in hash that comes from a composed named entity we look successively in it left and it right in order to combine more words with the same value in hash map. 2.1. If we have like neighbors these types of hash keys then we create a common tag. This is the case of the following line: 13128 & 11516 & ("peter woods)" for which we have in our cache separated keys “Peter” and “Woods” from GATE cities file (first one from “Peter Tavy” and second one from “Harper Woods”). Because both have the same value in hash (2 which represents the number of words from initial entity) we create a common tag for both words. 2.2. If we haven’t like neighbors these types of hash keys, then we eliminate the all these tags.
During all previous steps the stop words are ignored. We submitted two pairs of runs: one in which the main module loaded GATE like external resource, and one in which Portuguese or English Wikipedia are loaded like external resources. This paper presents the UAIC system which took part in the LogCLEF 2009 competition in LAGI task. The system uses like external resources GATE files and two files offered by organizers with titles from Portuguese and English Wikipedia.
Initial we load external resources in cache using a hash map. After that, at preprocessing part we obtain a new query from the initial user query. This new query is used by main module in order to identify geographical entities. In this process we searches in cache the initial query, and after that words components, with aim to have in the end the most comprehensive succession of geographical entities.
The results show how the results obtained with GATE or with Wikipedia like external resources are comparable like quality.
The author would like to thank to the students Victor Chircu and Alexandru Cristea and their colleagues from group 4A, second year, for their help and support at different stages of system development.