Extraction of index words from manuals

An automatic extraction method of index words from manuals, which optimizes Rijsbergen's E calculated from the recall and precision is described and experimentally evaluated.

Abstract

We describe and experimentally evaluate an automatic extraction method of index words from manuals. Enough number of high quality index words help us read and understand the contents of manual. In Japanese, majority of index words are compound nouns. Moreover we find several simple nouns that are parts of many compound nouns which are to be index words. Based on this phenomena, we propose the several functions which show how suitable compound nouns are as index words. Using one of these functions, we propose the method to extract index words, which optimizes Rijsbergen's E calculated from the recall and precision. Our result shows 0.712 of recall and 0.650 of precision for five Japanese manuals.