Bootstrapping a Comparable Corpus from Patent Family Members
Mihai Lupu · 2012
We present a method to generate comparable corpora from different patent documents covering the same invention. We rely on the fact that many inventors apply for protection in more than one jurisdictions. Often, these jurisdictions have different publication languages, and therefore, the same invention is described in more than one language. We use this fact to generate comparable corpora in any language pair where patent documents are available. We do this at the level of the title, abstract, description and claims and present statistics for English-Spanish data thus generated. We then show that with an additional filtering step we can reduce the errors inserted in the collection by the automated procedure.