Beauty and the Beast: What running a broad-coverage precision grammar over the BNC taught us about the grammar — and the corpus

Introduction Typically, broad-coverage precision grammars are based on grammaticality judgment data and syntactic intuition, and corpus data is relegated to secondary status in guiding lexicon and grammar development. On the other end of the scale, shallow grammars are often induced directly from treebank data and make little or no use of grammaticality judgments or intuition. This tends to cause precision grammars to undergenerate and shallow grammars to massively overgenerate. With broadcoverage precision grammars, the issue of undergeneration is solved incrementally by extending coverage. When faced with an as-yet unanalyzed sentence (from a corpus or otherwise), grammar engineers first consult the literature, their intuition, or their office mates in order to map out a space of both grammatical and ungrammatical examples, which serves as the basis for the analysis coded into the grammar. Shallow grammars, on the other hand, tend not to deal with grammaticality and focus instead on selecting the most plausible of the available parses, generally through stochastic means.