Generating Questions with Deep Reversible Grammars

One desirable approach to question generation exploits existing general-purpose grammars which are both linguistically rich and computationally effective. A generator using such a grammar can ensure high-quality output provided that the input semantic representation is reasonably coherent, and the characterization of a well-formed semantic input can be made quite precise. The Linguistic Grammars On-line (LinGO: lingo.stanford.edu) Project, founded at Stanford’s Center for the Study of Language and Information in 1994, has been developing the English Resource Grammar (ERG: Flickinger 2000) for generation (as well as parsing) within the context of a series of application-oriented projects, including German-English machine translation in Verbmobil; an NSF-funded project to develop a speech prosthesis for physically disabled users; an automated email response product developed with the Silicon Valley start-up YY Technologies; and the Norwegian-English machine translation project called LOGON. We have developed the widely used flat semantic representation called Minimal Recursion Semantics (MRS: Copestake, Flickinger, Pollard and Sag 2005), which is the input accepted by the ERG for generation, and we use the relatively efficient LKB (Linguistic Knowledge Builder: Copestake 2002) generator to produce realizations corresponding to such an input. All of the grammar and processing resources we use are distributed as open-source software through DELPHIN (the Deep Linguistic Processing with HPSG Initiative: www.delph-in.net), a growing international community of linguists and NLP developers.