next up previous
Next: Conclusion Up: Sources of Flexibility in Previous: A sample domain:

Parameters of flexibility in hypertext generation

We indicated in Section 1.1 that an NLG system can exploit a number of techniques to generate hypertext. In this section we will attempt to isolate a number of parameters which affect the flexibility of the resulting dynamic hypertext. In each case, we assume that the parameter influences the quality of the text, and we ask whether NLG needs to investigate the parameter in greater detail in order to produce good hypertext applications.

Granularity of information representation

A fundamental factor influencing generator flexibility is the form of representation of the information to be expressed. At one extreme, the information can actually be stored as text, in which case there is no scope at all for flexibility of expression; this is essentially the scenario for `canned' text. Scope for contextual variability is introduced by establishing an independent format for information storage, and a flexible mapping algorithm from this to surface text.

In fact, the `degree of cannedness' of a system's output is on a sliding scale. A static hypertext is made up entirely of canned text. In a slightly more flexible system, individual sentences or paragraphs may be canned, while their organisation and inclusion is constrained by some theory of discourse. (For instance, a system could store canned sentences organised in terms of an RST tree; the generator could then use the RST information to select and linearly organise the canned segments.) More flexible still, a system could use canned text only at the phrasal level, for example in a template-based sentence generation system, or in a phrasal lexicon. Finally, a system could dispense with canned text altogether, and use full NLG technology.

Concern with the appropriate degree of cannedness for NLG systems is of course not limited to hypertext generation. It is currently the focus for much debate in all areas of NLG, especially where the development of practical applications is an issue (see e.g. [Reiter et al. 1995]). But as the dimensions of flexibility we are interested in depend on this issue, it is important to discuss it.

ILEX-0 represents information at two distinct levels on the cannedness scale. For fine-grained detail, the system uses a database made available by the museum, containing entries for the 120 or so jewellery items in the Modern Jewellery exhibit. These entries have been parsed in and represented in a KB-accessible form, as shown in Figure 4.

  


   (def-jewel 
     :case '7-10 
     :isn 342040
     :code 'a.1987.284
     :short-title "a brooch on pyramid mount" 
     :date "1976"
     :class 'brooch 
     :class-text "brooch with mount"
     :materials '(acrylic yellow-material
                  white-metal) 
     :designer 'morris01
     :dimensions "5.20\" L"
     :properties '(mounted)
     :place "ENGLAND: London")

Figure 4: A Sample Database Entry

The generation of sentences based on this information is currently effected via a simple template-based mechanism.

Secondly, the transcriptions of the curator tours have been hand-processed to extract canned stories about individual jewellery items, the designers of the items, and more often, classes of jewellery (e.g. jewellery made in a particular style). Figure 3 shows one example.

Given the limitations of current NLG systems, canning offers the most reliable way of ensuring high-quality text output. And, since it is much easier to enter information in the form of text than using a complicated knowledge representation formalism, canning also facilitates the authoring task. But of course, canning fails to support genuine flexibility. In the ILEX-0 system, we have sought to allay this problem by including in the canned text various types of annotation which allow a certain amount of local customisation. The prototype allows two forms of customisation:

Further types of annotation will be added as the ILEX system develops. We plan to explore whether, for some applications, the medium of `annotated text' can be a useful compromise between canned text and full NLG.

Sensitivity to Discourse History

At any point in the dialogue, a dynamic hypertext system should be able to take into account the discourse which has already taken place. This can be exploited in several places; we here discuss two.

Referring Expressions

The readability of a text is partially determined by the appropriate use of referring expressions.The continual use of an entity's full name produces bad text, just as does the use of a pronoun without any preceding referent. Since dynamic hypertext is being produced chunk-by-chunk, the system needs to keep track of prior reference forms for an entity, and choose a form for the current mention which is appropriate. For instance, if we have recently mentioned an entity of type T, and we now need to introduce an object of the same type, then we could refer to it as another T. There are already several systems which incorporate sophisticated algorithms for the generation of referring expressions such as these; see for instance IDAS and [Dale 1992], [Horacek 1995]. However, IDAS did not track the discourse history beyond the current hypertext node being generated. We need a system which can keep track of reference across nodes.

In particular domains, we might expect other kinds of sensitivity to preceding context. For instance, if the hypertext session is intended to simulate a spatio-temporal domain, then it makes sense to talk about `returning to' locations which have already been visited. But if the tour is of a more conceptual space, this is not what is needed. One of the domains in IDAS is bicycle mechanics; clearly it makes no sense to talk about returning to the wheel rim of a bicycle.

Content Selection

It is also desirable that the content of the text produced is sensitive to what has been said in the preceding discourse. We assume that the system aims to tell the seeker something interesting---but what is interesting depends partly on what has already been said. Some things become interesting only because of prior context. For instance, the system might have a very informative comparison to make between two classes of objects in the domain being browsed; but if the seeker has not already come across instances of each class of objects, the comparison might not be interesting or even intelligible. The system's decision about whether to include the comparison might thus be strongly influenced by prior context.gif

On the other hand, some things cease to be interesting if they have already been discussed, due to repetitivity. This problem, called here the redundancy problem, affects static hypertext systems: since a series of hypertext pages can be browsed in a variety of sequences, a static system cannot know at any point what the user has already been told. Information essential to the interpretation of the present text may need to be repeated at each point it is needed. As a result, the reader may become bewildered by the mass of information presented, unable quickly to determine what is new and what is a restatement of what he has already read. A dynamic system however can keep track of what has been told to the user, and not repeat such information (except, naturally, for rhetorical reasons).

Of course, when we decide not to re-express old information, we can still provide a backward-pointing link. If a story has already been given (an earlier piece of jewellery had the same property or designer), then repeating it verbatim can have a serious effect of the user's perception of text quality. Because of this, the second time a story is triggered in ILEX-0, the story is not printed, but rather, a pointer is given back to the prior piece where the description occurred. For instance:

Like the earlier piece of jewellery (240384), this piece is also a piece made for Liberty & Co.gif
A third content-suppressing mechanism implemented in ILEX-0 is for handling terminology definition. The system keeps a list of terms defined for the current user, and avoids re-defining these terms.

The question arises as to whether hypertext applications demand more work on the discourse history parameter---supporting referring expressions and content selection---than has hitherto arisen in NLG. The tentative answer---arising particularly from the need dynamically to suppress content---is that this is an area where further work is required.

The User Model

A useful kind of flexibility in any NLG system involves keeping a model of the reader of the text. Texts produced for an expert will be very different from those produced for a novice, even if they are on the same subject. Likewise, texts for an adult audience will differ from those intended for children. In the domain of guided tours, user models will clearly be important; in fact, the National Museums of Scotland have a ramified visitor classification scheme, with four categories of visit (including tourist, formal educational), fifteen types of visitor (including 9-year-old, tourist with Scottish connections), and four levels of visitor interest (including baby browser and specialist). Many NLG systems have taken such dimensions of variation into account: see for instance [Paris 1993].

An important dimension of variability in hypertext generation relates to control over the user model. Changes to the user model can be under the control of the system or the user. We might envisage a form-based preferences page where the user can configure his own user model, either before browsing begins, or at any point within browsing. The ILEX-0 prototype allows the user to specify the degree of interest he has in a set of themes. These preferences are then used to drive content-selection. We also plan to experiment with user-changeable parameters, such as `longer/shorter texts', `more/fewer comparisons' etc. These choices are roughly analogous to meta-comments in ordinary discourse.

Alternatively, updates to the user model can be system-initiated, and driven by observations about the options chosen by the user during browsing. If the user often selects the tell me more about this button, then the quantity of text given at each point can be increased. A user who often asks for terms to be defined can be offered more definitions within the text. A user model is thus built up by behavioural observation. This approach can also capture changes in the user's interest as they occur.

ALFRESCO maintains an interest model for its user which can affect what is generated in the first textual response. However, a system of this kind which sets up a network of hyperlinks in its entirety in response to a single query from the user, will not be able to take full advantage of a dynamic user-model. The ideal may be a system where pages and links are created one by one, and the user has the opportunity for input at every stage.

The theory of user modelling in NLG is still at a relatively early stage of development. However, state-of-the-art techniques should be sophisticated enough to deal with many of the requirements of dynamic hypertext mentioned above.

User Freedom

In a hypertext system, the user's requests for follow-up information are mainly in the form of `mouse clicks' to new hyperlinks. From the user's point of view, there is a degree of flexibility in the system even if the hypertext is fully generated in advance, because he has a choice about what to look at next. The two main ways in which a system can give freedom to the user are:

In some systems, all the possible texts have been anticipated explicitly in advance, even though the detailed contents may be dynamically generated. An example might be a system that offers weather forecasts for a fixed set of regions but where the individual forecasts may change to reflect up-to-the-minute information. The PIGLIT system is rather like this. More flexibility is offered if the text produced depends on a set of independent dimensions. In IDAS, for instance, there were over 5000 potentially different texts that could be generated (ignoring some dimensions of variation). PEBA-II has similar flexibility, in that its comparison texts depend on two animals that are chosen by the user independently. ILEX-0 is nearer to the PIGLIT model, because its primary mode of access simulates the actual layout of a museum gallery, where the user has the simple choice of which of a set of objects to look at next. Of course, as soon as user modelling, sensitivity to discourse history, etc. are reflected in the texts, then the number of possible texts grows quickly (indeed, the above IDAS figure takes into account some user modelling dimensions). This potential freedom may not be obvious or available to the user, however.

Just as a speaker must allow a hearer opportunity to respond, it is important for the system to generate a fair selection of new links to explore, to ensure that the reader finds something to his liking. In ILEX-0, the user is able to pick any object in the gallery to look at next, which is intended to allow the same kind of random movement that a user might make in a real museum gallery. On the other hand, it might also be important not to generate too many links to follow up, as this might make navigation confusing. In IDAS, for instance, an explicit decision was made to limit users to following certain `hyperschemas', rather than giving them the full choice. Where the set of possible texts has not been anticipated in advance, a more complex (perhaps menu-based) query system to allow the user to choose a text may be provided (as in PEBA-II).

Thus, the ability to specify a restricted subset of potentially very large set of user follow-up actions is a particular feature of dynamic hypertext. It follows that the whole issue of `user freedom', and ways of anticipating user actions will require further work in NLG. Users' freedom---or their perception of freedom---is intimately bound up with the ways in which systems realise their own goals, and it is to this that we now turn.

The System's Goals

For a hypertext system to be successful, it needs not only to succeed from the user's point of view (to be interesting), but also to succeed from the provider's point of view (to be informative).

In general, the provider has her own agenda in creating the hypertext, for instance to persuade the user to buy something, or to educate him in a particular way. In a browsing situation, however, the provider cannot plan ahead towards a complete text which satisfies her goals, since the reader's actions cannot be predicted in advance. Goals must thus be pursued opportunistically---the system achieves aspects of its goals as the dynamic environment permits. If the primary goal is educational, for instance, then the important points to be made must be incorporated into descriptions of the objects in which the browser is interested. The texts produced by the system are thus influenced by an interaction of the user's and the system's goals.

As the system is able to respond to the user's initiatives, it can be accurately termed reactive; however, it provides a rather different kind of reactivity than is normally discussed in the context of text planning systems. The most commonly discussed type of reactivity (see e.g. [Moore and Paris 1993]) is where a system allows the user the option to ask follow-up questions, to help clarify portions of the text which have not been fully understood. In such a case, the user has control over the way the system executes its goals, but not, ultimately, over the goals themselves. In the present system the situation is just the reverse: the user is able to determine the high-level goals about which objects are described, but the system is in control of the execution of these goals so as to include (if possible) information on its own agenda.

In the ILEX-0 prototype, system goals are represented as a set of messages it plans to communicate. In the case of the Jewellery domain, these messages are seen as the educational goals of the display curator. Some messages from this domain are:

  1. Jewellery need not be made of valuable materials.
  2. Contemporary jewellers make innovative use of modern synthetic materials.
  3. Jewellers frequently borrow techniques from other crafts such as pottery and weaving.
  4. The distinction between `art' and `jewellery' is often blurred.

Each of these messages is ranked on a scale ranging from important to unimportant. Attached to individual pieces of jewellery or classes of jewellery are canned stories that can be told if those items are chosen by the user. Each story in the story-database is tagged by the message or messages it works towards. Stories are thus ranked in relevance on the basis of the messages they serve, and will be expressed or not on the basis of this relevance. Thus does content-selection serve the system's goals.

If a number of stories have been given supporting a particular message, an explicit description of that message may be triggered; e.g., Here again, we see that jewellery need not be made of valuable materials.

The model of opportunism in ILEX-0 is very simple, but existing NLG planning architectures do not seem to allow even this to be captured naturally. Support for opportunistic satisfaction of system goals is therefore an area in which further NLG research is required. We are currently looking to see what models of opportunistic planning (e.g. [Pryor and Collins 1994]) have to offer in this respect.



next up previous
Next: Conclusion Up: Sources of Flexibility in Previous: A sample domain:



Mick O'Donnell
Mon Feb 12 17:01:45 GMT 1996