There are two methods to construct linguistic knowledge-base for natural language processing. One method is to construct the knowledge-base by the native speaker's introspective evaluation and the other method is to extract the knowledge from langage database. The latter is more objective and quantitative, but has less-data problem. This is the fact that there are many knowledge data which is supposed to appear in the database by the native evaluation but actually they do not appear at all or appear unfrequently. Linguistic knowledges can be categorized by the number of words included in the knowledge. Using this categorization, the distribution of knowledge data extracted from language database is caluculated assuming that the distribution of words is Zipf's distribution. The coveradge of knowledge-base constructed from language database which is considered a sample from a population is computed.