Analyzing Parallelism and Domain Similarities in the MAREC Patent Corpus

Statistical machine translation of patents requires large amounts of sentence-parallel data. Translations of patent text often exist for parts of the patent document, namely title, abstract and claims. However, there are no direct translations of the largest part of the document, the description or background of the invention. We document a twofold approach for extracting parallel data from all patent document sections from a large multilingual patent corpus. Since language and style differ depending on document section (title, abstract, description, claims) and patent topic (according to the International Patent Classification), we sort the processed data into subdomains in order to enable its use in domain-oriented translation, e.g. when applying multi-task learning. We investigate several similarity metrics and apply them to the domains of patent topic and patent document sections. Product of our research is a corpus of 23 million parallel German-English sentences extracted from the MAREC patent corpus and a descriptive analysis of its subdomains.