Constructing literature abstracts by computer: Techniques and prospects

摘要

The automatic generation of abstracts has always been a rather neglected problem. Text summarisation methods based on artificial intelligence approaches can only be made to work in very restricted domains; consequently, practical abstracting has been attempted by simple ad hoc methods. The early workers concentrated on techniques for extracting informative sentences from documents. Inevitably, this resulted in ‘abstracts’ which were highly disjointed, due especially to the presence of dangling anaphoric references. Subsequently, work has been carried out on the twin problems of (a) recognising anaphoric references, and (b) constructing passages in which no anaphoric references are left dangling. The correct handling of definite noun phrases however represents an unsolved problem. The question of how to achieve proper balance in auto-abstracts is a problem which has hardly been addressed hitherto; it appears that assignment of selected textual material into ‘abstract-frames’ may offer a solution. However, radical progress on the construction of satisfactory abstracts now seems to require an adequate theory of text structure.