Text Generation: The Problem of Text Structure

Overview

What is text generation?¹ The long term view is that it is the process of creating a technology for building computer programs that can function as authors or speakers. I call this the long term view because in the text generation programs in existence today there is very little that deserves the title of “author” or “speaker.” Writing and speaking are rightly regarded as complex arts capable of high refinement and great intellectual achievement. In contrast, our programs reflect only fragments of the most basic skills.

Text generation has been studied seriously in computational linguistics only in the last five or ten years, and so it is still sorting out its goals and identifying its problems.

Part of the diversity of approaches in text generation will surely come from a problem that we face now: It is by no means clear how authors do what they do. Even though we are all exposed to text nearly every day, and manipulate it successfully, there is very little explicit knowledge of how text works.

One of the central issues involves text organization. It is evident that natural text is organized, and that its organization is essential to its function. Text has parts, arranged systematically. But what is the nature of text organization or structure? What are the parts, and what are the principles of arrangement?

We must have answers to such questions if we are to create text generators. However, there are no widely accepted answers to these questions. The answers that are available from outside of computational linguistics are partial and complex. There are logicians’ answers, grammarians’ answers and so forth, often representing mostly the favorite methods and assumptions of their developers – – a priori selectivity rather than comparative results.

Also, crucially for text generation, there are no accounts of text organization at a level of detail sufficient to support computer programming. Far more detail is needed.

As a result, text generation has been inventing its own answers to these questions. It has had to.²

To explore the nature of text structure, we focus in this paper on two of the energetic attempts within Text Generation to describe text structure in a way that is sufficiently detailed and general to serve as a basis for programming. They are:

• 1. The TEXT system, and

• 2. Rhetorical Structure Theory

The TEXT system was developed by Kathy McKeown at the University of Pennsylvania, as the centerpiece of her PhD dissertation, and is being followed up at Columbia University and elsewhere [Paris 86]. Rhetorical Structure Theory (RST) was initially defined by Sandra Thompson, Christian Matthiessen and the author; it is under active development at USC Information Sciences Institute.³

This paper describes each of these lines of research in its own terms; then they are compared as text structure descriptions. The comparison is extended to the related construction processes, and finally conclusions are drawn about text structures in future text generation work.