The CLEF-IP track was launched in 2009 to investigate IR techniques for patent retrieval. It is part of the CLEF 2009 evaluation campaign. Also, in 2009 we built a system in order to participate in the CLEF-IP competition. Our system has three main components: filtering module, indexing module, and searching module. Because the process of indexing of all 75 G of input documents with patents took too much time we decided to work in a peerto-peer environment with four computers. In this paper we will see how we managed to built this system.
Our system has three main modules: module one responsible with extracting of relevant fields from XML files, module two that indexes the relevant fields, and the third module that does the searching. The Figure 1 presents the system architecture.
Indexing Module
In what follows we will see details about these modules.
The aim of this step is to reduce the amount of XML files that must be indexed and to work only with relevant tags from XML files. Why? Our initial tests demonstrate that in order to index ~150Mb we need around 94 seconds. Removing the irrelevant fields the time need for indexing was reduced to 82 seconds.
We agreed on the tags we will keep (<invention-title>, <claim text> and <abstract> tags. If <abstract> tag would not be found we will search and keep the <description> tag). These fields were determined after a review of initial preprocessed documents.
In order to index the corpora we use Lucene
Because the process of indexing of all 75 G of documents took too much (for our initial indexing we need around 80 minutes) we decided to work in a peer-to-peer environment. Thus, we split the initial corpora on four machines and we create separate indexes on each of them. In this way the time necessary for indexing was reduced to 20 minutes.
This component allows performing searches in index created at previous step. Starting from patent files with size of 1 M, we extract the same tags used in indexing part and built a Lucene query in order to search in Lucene index.
When we create the Lucene query from input patent we use different boost factors for used tags. Thus, we have the following cases: i. The greater boost value is used when we find one tag from patent in the same corresponding field from Lucene index. For example, this is the case when we search words from invention-title tag of current patent in the invention-title field of Lucene index. ii. The boost values depend by current tag name. Thus, boost values are in descending order starting from invention-title, claim text, abstract and description tags. iii. The lower boost values are used when we have cross-searches between tag from patent and field from Lucene index. For example, if we search words from invention-title tag from patent in abstract field from Lucene index.
Fourteen groups submitted 70 runs at this track. We submitted one run. Details from official evaluation are presented below: The CLEF-IP was a new track in CLEF 2009, which utilized a collection of more than 1M patent documents from English, French and German languages with at least 100,000 documents in each language.
The UAIC system, which took part in the CLEF-IP 2009 competition, has three main components: filtering module, indexing module, and searching module. The filtering module has the aim to reduce the amount of XML files that must be indexed and to work only with relevant tags from XML files. The indexing module used Lucene and because the process of indexing of all 75 G of documents took too much we worked in a peer-to-peer environment. The searching module component allows performing searches in index created at previous step. When we create the Lucene query from input patent we use different boost factors for used tags.
We want to give a “thank you” to those who helped from the beginning of the project: students from second year group 5A.