An architecture for language processing for scientific texts

We describe the architecture for language processing adopted on the eScience project 'Extracting the Science from Scientific Publications' (nicknamed SciBorg). In this approach, papers from different sources are first processed to give a common XML format (SciXML). Language processing modules operate on the SciXML in an architecture that allows for (partially) parallel deep and shallow processing and for a flexible combination of domain-independent and domain-dependent techniques. Robust Minimal Recursion Semantics (RMRS) acts both as a language for representing the output of processing and as an integration language for combining different modules. Language processing produces RMRS markup represented as standoff annotation on the original SciXML. Information extraction (IE) of various types is defined as operating on RMRss. Rhetorical analysis of the texts also partially depends on IE-like patterns and supports novel methods of information access.