A Large-Scale Japanese CFG Derived from a Syntactically Annotated Corpus and Its Evaluation

Although large-scale grammars are prerequisite for parsing a great variety of sentences, it is difficult to build such grammars by hand. Yet, it is possible to build a context-free grammar (CFG) by deriving it from a syntactically annotated corpus. Many such corpora have been built recently to obtain statistical information concerning corpus-based NLP technologies. For English, it is well known that a CFG derived from the Penn Treebank corpus (tree-bank grammar) can parse sentences with high accuracy and coverage although the method for deriving a CFG is very simple [1]. Actually, there have been quite a few studies concerning this kind of grammars. For Japanese, however, CFGs cannot be derived using the Charniak’s method since there is no large-scale syntactically annotated corpus such as the Penn Treebank corpus 1. Therefore such corpus needs to be developed to enable derivation of a large-scale CFG. However, even if a large-scale, syntactically annotated corpus were already available, a CFG derived from it can be unsatisfactory, in as it creates a great number of possible parses (in average more than 1012, according to our preliminary experiment). Too many parse results do not only reduce the parsing accuracy and parsing speed, but also require larger memory to parse and store long sentences. Although Charniak has removed some CFG rules (e.g. rules occurring only once