next up previous contents
Next: Bibliography Up: ATG01 Previous: Evaluation   Contents

Conclusions: Problems, bottlenecks, and outlook

In this introduction to the field of NLG, we have seen something of the variety of the approaches adopted, as well as the overall structure of the problem of the automatic production of natural language texts. We have also seen that there is substantial potential for application of natural language generation technology, although that potential remains as yet largely unrealized. Along with some genuinely difficult and unsolved problems, there is a widespread lack of understanding of the relevance of NLG techniques to concerns ranging from knowledge representation to page layout (for the latter, see particularly [Reichenberger, Rondhuis, Kleinz and Bateman: 1996,Bouayad-Agha, Scott and Power: 1996,Bateman, Kamps, Kleinz and Reichenberger: 2001]). Although it may still be some time before NLG techniques begin making a substantial impact on systems in everyday use, increased awareness of the potential will certainly be the most crucial aspect in hastening the shift towards their acceptance.

The most significant bottleneck is therefore now sociological rather than technical. The potential uses of NLG are still relatively poorly understood and most system designers remain ill-informed about the technology that is being developed. Moreover, there is little knowledge concerning what is required of an application system if it is to adequately support the opportunities that NLG offers. The task as a whole is generally seen in a simplified light, resulting in non-realistic assessments of the investment necessary to achieve sophisticated and compelling natural language performance. Partly to counter this viewpoint, this review has tried to make it clear that the text surface of sequences of characters or sounds, which most would associate with the meaning of the word `text' itself, in fact represents only the most visible (audible) tip of an information iceberg. Many of the submerged layers of this iceberg have important consequences for other, at first glimpse unrelated, problems; this information needs to be incorporated into knowledge-based systems at the outset in order to avoid re-engineering or damaging compromises when NLG functionality is subsequently required.

One good example area here is domain modelling. Although knowledge representations equally optimal for all uses probably cannot be found, there are some basic modelling principles for application domains developed within NLG that can improve the adequacy of domain knowledge representations. Organizations of the kind provided by an Upper Model (see above) appear often to have beneficial side-effects for domain modelling by enforcing a consistent `ontological' modelling style that is difficult to re-impose after modelling has begun. This not only eases the transition from application system terms to NLG terms should language generation be desired, but also becomes a worthwhile investment in its own right. The maintenance of appropriately structured domain information can then serve a variety of functions, both during the development of that information--for example, by supporting design rationale generation [Gruber and Russell: 1995] or by making the consequences of the formal definitions given more readily apparent to users [Frölich and van de Riet: 1997,Aguado, Bañón, Bateman, Bernardos, Fernández, Gómez-Pérez, Nieto, Olalla, Plaza and Sánchez: 1998,Hoppenbrouwers, van der Vos and Hoppenbrouwers: 1996]--as well as after, by providing improved and wider access.

The addition of multilingual NLG also adds significantly to the re-usability of the information maintained by any system and is beginning to provide an appropriate alternative to translation [Kittredge: 1992,Hartley and Paris: 1995,Hartley and Paris: 2000]; the ability of NLG to tailor its information presentation to particular audiences also widens accessibility still further. While these aspects alone can offset the extra effort required to develop the richly structured information sources required, there are also tasks where the construction of a detailed and explicit body of domain knowledge is itself motivated independently of any need for language generation--examples here are the growing use of conceptual models in expert system design (cf. [Swartout: 1985]) and computer-aided software engineering (CASE: cf. [Wieringa: 1989]). Proposals for incorporating generation have been made both for explainable expert systems [Moore and Paris: 1993] and for CASE conceptual modelling applications [Gulla: 1996,Hoppenbrouwers, van der Vos and Hoppenbrouwers: 1996].

While the initial move from non-NLG based technology to NLG is a large one, once taken there are several further possibilities that can be realistically targetted. For example, the provision of a rich text structure also supports a range of increased functionalities for information presentation. Rhetorical organization can motivate hyperlinks in hypertext as well as textual realization with particular conjunctions. The deployment of well-established rhetorical relations is then one obvious way of reducing the perceived complexity of hyperdocuments: adding generation techniques to tailor the linked document fragments can take this still further. This is particularly important given the current drive towards an information society with `digital libraries' and `information highways'. NLG systems are accordingly now beginning to appear on the World Wide Web (cf. [Gruber, Vemuri and Rice: 1995,Milosavljevic and Dale: 1996b,Alexa, Bateman, Henschel and Teich: 1996,Dale, Oberlander, Milosavljevic and Knott: 1998]). The pervasiveness of this new media gives fresh impetus to previously more academic explorations of automatic hypertext generation; many aspects of NLG--including text planning, multimodal information presentation, and tactical (surface) generation are now combining in the construction of (possibly partially) synthetic hyperdocuments. This area is certain to be highly significant in increasing the acceptance and use of NLG technology, as well as in driving future NLG research.

The merging of automatic generation techniques and the provision of hypermedia documents on the world-wide web and other linked-information contexts, such as organizational intranets, is a natural development that suggests new answers to some still unsolved problems with hyperdocument-based information presentation. For example, although often hotly contested, it is in fact the case that the early promise (and promises) of hyperlinked text have in many respects failed [Dillon: 1996,Rouet and Levonen: 1996]. Suggestions that hypertexts would favor more effective learning and allow users to deal easily with complex information have been regularly scaled down in scope and even placed in question by empirical investigation. Several critics of `simple' views of hypertexts relate this failure to the fact that users need to be able to construct coherence while manipulating and exploring information. If that coherence is not evident, then it is not possible for information to be understood and used effectively. The ability to `overcome' linearity by ignoring coherent development in a body of linked data then becomes, not a liberation, but an enormous organizational overhead that undermines the other benefits that a set of linked data might provide.

Here NLG may provide the missing link that has made the original expectations of hypertexts fall short of their mark. When moving from one node in a hyperdocument to another, a user has to establish a context of interpretation for the new information for it to be coherently related to what has gone before. This means that it is common for users to go back to their previous node, to reread this, and then to try and make sense of the new information. This may be more or less difficult since the text of the former node may not fit particularly to the text of the latter node, and vice versa. If, however, the texts were not fixed but were dynamically generated using NLG technology so as to respond appropriate to the particular `text' that the user is implicitly constructing in their traversal of the space of hypernodes, then coherence and the construction of coherent interpretations of the data is optimally encouraged. It may then turn out that the real requirement for effective hypertext lies not in the linking but in the dynamic production of the nodes so as to support the sense intended, but not created, by the links. Here we have an important connection to the rapidly growing field of adaptive hypermedia, which, while recognizing and promoting the importance of dynamic node generation, generally has not yet employed NLG technology to achieve it [de Bra and Calvi: 1998]. The growing number of generation researchers moving into this area is very likely an indication of the change to come [Stock: 1993,de Carolis: 1998,Oberlander, O'Donnell, Knott and Mellish: 1998,Lu, Paradis, Paris, Wan, Wilkinson and Wu: 2000,Bontcheva: 2001,Androutsopoulos, Kokkinaki, Dimitromanolaki, Calder, Oberlander and Not: 2001].

The influence of web-based technologies is also beneficial for the development of NLG in other ways. Until quite recently, the most significant technical bottleneck preventing the wider exploitation of NLG technology was the availability of sufficiently well-structured knowledge sources: unless such information is available an NLG system has literally `nothing to say'. In earlier work on NLG, source information was typically handcrafted--this severely restricted the added value of the NLG components. Now, however, it is becoming common for structured information to be acquired either fully automatically on the basis of automatic analysis of texts (however limited), or semi-automatically on the basis of controlled input specifications or `knowledge' markup, where not only issues relevant to logical document structure and formatting are encoded but also information concerning the content of the document [Rostek, Möhr and Fischer: 1994,Rostek: 1999,Knorz and Möhr: 1999]. The current moves towards employing XML-representations for all kinds of content-information provides a further push in this direction as the value of making content accessible becomes clear to almost any web-designer. The fact that organization and content is made explicit through various more or less standardized annotation schemes provides a far more suitable basis for employing NLG technology than hitherto widely available and simplifies the task of interfacing with content. Accordingly, there are already attempts to explore just how far the XML-based technologies that are emerging can take us towards NLG [Cawsey: 2000,Wilcock: 2001], as well as re-interpretations of document `re-purposing' as an NLG task.

To conclude, there are several other areas where the application of results from NLG can be expected to bring benefits. For example, the increasing complexity of information systems requires a matching increase in the sophistication of such systems' abilities to interact with their users (cf. [Stein and Thiel: 1993]). Here, the text planning techniques of NLG--and particularly of dialog generation--will become ever more relevant as the models of interaction approach those of natural language. Also, text structure can be used to motivate sophisticated page layouts where the setting out of information on a `page'--be this physical or electronic--is more supportive of the communicative goals of the system than is currently achieved in dynamically produced pages Information can be differentiated according to its importance and category by corresponding differences in type faces, relative positioning, etc. thereby increasing both the effectiveness and the attractiveness of the presented information.

The ability to create text has long been seen as one of the most distinctive traits of the human mind, and the range of such texts' functionalities in and for human societies knows few bounds; it should not therefore be surprising that NLG faces many difficult issues. The range of functions that language is called upon to carry leads to a phenomenon whose richness and breadth invites an unusually diverse set of approaches, and the basic phenomenon at issue--`textuality'--is still one which is relatively poorly understood. The question of what makes a text a text, rather than just a sequence of words or sentences, raises enough theoretical questions to keep research busy for many years to come. It is of more than theoretical interest, however, to establish more secure theoretical foundations for the field of NLG as a whole: better theoretical understandings of what is required will have direct consequences for improvements in practical systems. In order for NLG to begin fulfilling its potential it is essential that theoretical and practical work be combined: practical developments alone will be shortlived, while theory needs to be driven further to cover real-scale systems and functionalities. Thus, even though a sound theoretical basis for NLG still requires considerable work, there are certainly sufficient grounds for already considering wider applications of NLG technology.


next up previous contents
Next: Bibliography Up: ATG01 Previous: Evaluation   Contents
bateman 2002-09-21