A precision grammar is a formal grammar designed to distinguish ungrammatical from grammatical sentences. This is in contrast to large treebank-induced grammars which often accept ungrammatical input [Charniak 1996]. While high coverage is required, it is difficult to increase coverage without also increasing the amount of ungrammatical sentences that are accepted as grammatical by the grammar. Most publications in grammar-based automatic error detection focus on locating and categorising errors and giving feedback. Existing grammars are re-used [Vandeventer Faltin 2003], or grammars of limited size are developed from scratch [Reuer 2003].
The ParGram English LFG is a hand-crafted broad-coverage grammar developed over several years with the XLE platform [Butt 2002]. The XLE parser uses OT to resolve ambiguities [Prince and Smolensky 1993]. Grammar constraints resulting in rare constructions can be marked as ``dispreferred'' and constraints resulting in common ungrammatical constructions can be marked as ``ungrammatical''. The use of constraint ordering and marking increases the robustness of the grammar, while maintaining the grammatical / ungrammatical distinction [Frank 1998]. The English Resource Grammar (ERG) is a precision Head-Driven Phrase Structure Grammar (HPSG) of English [Copestake and Flickinger 2000, Pollard and Sag 1994]. Its coverage is not as broad as the XLE English grammar. Baldwin et al. 2004 propose a method to identify gaps in the grammar. Blunsom and Baldwin 2006 report ongoing development.
There has been previous work using the ERG and the XLE grammars in the area of computer-assisted language learning. Bender et al. 2004 use a version of the ERG containing mal-rules to parse ill-formed sentences from the SST corpus of Japanese learner English [Emi 2004]. They then use the semantic representations of the ill-formed input to generate well-formed corrections. Khader et al. 2004 study whether the ParGram English LFG can be used for computer-assisted language learning by adding additional OT marks for ungrammatical constructions observed in a learner corpus. However, the evaluation is preliminary, on only 50 test items.