PURPOSE: To attain the extraction of a clause/phrase boundary with no use of a grammar rule set with the linguistic knowledge by outputting the position information on the clause/phrase boundary through a neural net where plural categorized words are successively added and a teacher pattern is already learnt to the position information on the clause/phrase boundary.

CONSTITUTION: A neural net means 2 outputs the boundary position information through a neural net where plural categorized words are successively added and a teacher pattern corresponding to the boundary position information is already learnt. The plural words categorized by an input means 1 consist of an object word that decides a preceding or subsequent position of the boundary and preceding (n) pieces of words and subsequent (m) pieces of words centering on the object word (n = 1, m = 2, etc.) For instance, the boundary position information is outputted through a learnt neural net by means of three or four continuous words. Then a clause phrase boundary is obtained based on the boundary position information and the threshold value. Thus it is possible to extract the clause phrase boundary with no use of a grammar rule set with the philological knowledge.