A Vector Space analysis of Swedish patent claims, does decompounding help?

In the present study, a comparison between three different subsets of patent claims against an entire collection of 30,117 claims was performed in order to find similarities. A Vector Space Model was used as the retrieval model to calculate similarity between claims. The study was performed in a traditional laboratory environment. The Vector Space Model was evaluated with and without an additional morphological decompounding module. The decompounding module was implemented in two different ways (an exhaustive method and a more restricted method). The results indicate that decompounding will influence the performance of the retrieval model in a positive way. However, the sublanguage of patent claims and the errors made during the optical character recognition process were harmful towards the overall performance of the retrieval model. D-Level Thesis Advanced Course in Computational Linguistics May 2009 Supervisor: Magnus Sahlgren Department of Linguistics