Using the English Resource Grammar to extend fact extraction capabilities

Fact extraction from unstructured sources is key to supporting human-machine cognitive tasks, but requires natural language processing (NLP) to provide more comprehensive and formal expression of facts in terms of an agreed conceptual model. We describe BPP13 research to extend the BPP11 work on Controlled English (CE) for output, configuration and rationale in fact extraction, by integrating DELPH-IN linguistic resources; specifically a comprehensive English grammar,(the English Resource Grammar, or ERG), an efficient parser (PET) and a formalism for semantics (Minimal Recursion Semantics or MRS). Integration requires the representation of structures in the ERG/PET/MRS system within CE, based upon a mapping between the underlying Typed Feature Structure representation and the BPP11 common linguistic model in CE. Entries in the lexicon must be mapped into concepts in the domain model; grammar rules must be mapped into CE rules and other CE structures, so that a user can tailor and understand linguistic reasoning on unstructured sources for their own domain; sentence semantics in MRS must be mapped into CE domain semantics in order to extract CE facts for further reasoning. We explore initial mechanisms for these mappings, as we seek to understand how semantic processing by NLP can be deepened to take more account of domain semantics, expressed in CE, improving quality of fact extraction.