Robert Dale
Microsoft Institute of Advanced Software Technology, Sydney, Australia
On almost every office desk there sits a PC, and on almost every PC there resides a word processing program. The business of text creation and editing represents a very large market, and a very natural one in which to ask how we might apply speech and natural language processing technologies. Below, we look at how language technologies are already being applied here, sketch some advances to be expected in the next 5--10 years, and suggest where future research effort is needed.
Information technology solutions are generally of three types: accelerative, where an existing process is made faster; delegative, where the technology carries out a task previously the responsibility of a person; and augmentative, where the technology assists in an existing task. The major developments in the next 5--10 years are likely to be of an augmentative nature, with increasingly sophisticated systems that have people and machines doing what they each do best. The key here is to add intelligence and sophistication to provide language sensitivity, enabling the software to see a text not just as a sequence of characters, but as words and sentences combined in particular structures for particular semantic and pragmatic effect.
Although language technologies can play a part in the process of
text creation by providing intelligent access to
informational resources, the more direct role is in the provision of
devices for organizing text. The degree of organizational
assistance that is possible depends very much on the extent to which
regularity can be perceived or imposed on the text concerned.
Document production systems which impose structure support text
creation; the most useful offspring here has been the
outliner, now a standard part of many word processing
systems. However, in general the model of documenthood these systems
embody is too constrained for widespread use in text creation. While
relatively structured documents are appropriate in some business
contexts, other future markets will focus on home and leisure
usage, where concerns other than structure may become relevant to the
creation of text. In the following two subsections we focus on
unconstrained text, whereas controlled languages are treated in
section 
.
No existing tools in this area embody any real language sensitivity. Much of the initial exploratory work required here has reached computational linguistics via research in natural language generation; but we are still far away from being able to automatically interpret discourse structure in any sophisticated sense. Current models of discourse structure do not mirror the sophistication of our models of sentence structure, and so the scope for assistance in text creation will remain limited until significant research advances are made.
The story is very different for text revision. Here, language technology finds a wide range of possible applications. We already have the beginnings of language sensitivity in spelling correction technology: the techniques used here are now fairly stable, although without major advances (for example, taking explicit account of syntax and even semantics) we cannot expect much beyond current performance.
Grammar checking technology is really the current frontier of the state of the art. Commercial products in this area are still much influenced by the relatively superficial techniques used in the early Unix Writer's Workbench (WWB) system, but some current commercial systems (such as Grammatik and CorrecText) embody greater sophistication: these are the first products to use anything related to the parsing technologies developed in the research field. As machines become more powerful, and as broad-coverage grammars become more feasible, we can expect to see more of the CPU-hungry techniques developed in research labs finding their way into products; IBM's Critique system gives a flavor of what is to come.
Beyond grammar checking, the next important step is stylistic analysis. Anything more than the very simple string and pattern matching techniques first used in the Unix WWB system require the substrate of syntactic analysis, and, indeed, there are many aspects of style for which semantic and pragmatic analyses are required. Here more than anywhere the problem of different perceptions of the shape of the task rears its head: style is a term used to cover many things, from the form in which a date should be written to the overall feel of a text. Some of the simpler problems here are already being dealt with in products on the market, and this is where we can expect to see most developments in the next five years.
The key to medium-term developments in this area is the productization of parsing and grammar technologies. There are a number of shifts in research focus that are needed to accelerate this process.
We have already alluded above to the scope for incorporating sophisticated theories of discourse into the creation task in writing tools; similarly, the acceleration and delegation of language-centered tasks will become increasingly viable as advances are made in speech processing and natural language generation in the longer term.
Looking more broadly, we should be concerned not only with the words themselves, but also how they appear on the page or screen. The fact that, for example, we often have to make our texts fit word limits means that we have to take account of physical space. Systems should be able to reason about graphics as well as words, and systems should know about typographic devices.
Beyond these areas, there are new categories of assistance we might expect in the longer term. Modes of writing themselves are likely to adapt to accommodate the uneven profile of ability offered by existing systems, with currently unpredictable back and forwards effects on the tools that become required. We can't easily foresee what new market possibilities for computer-based writing tools the information superhighway will lead to; but there is a strong possibility that the categories we have previously thought in will no longer be the most appropriate.