next up previous contents
Next: Interpersonal decisions Up: Variation: how to describe Previous: Variation: how to describe   Contents

Ideational decisions

The most obvious source of decisions for controlling lexicogrammatical decisions--particularly from the perspectives of logic, natural language understanding, or applications of NLG--is the `propositional content' of individual sentences: the traditional ``who did what to whom (when and how)''. This can be used to guide the lexicogrammatical selection of basic units (such as, often, clause vs. noun phrase vs. adverbial, etc.) and the basic grammatical structures used within these categories (e.g., intransitive, transitive or bitransitive clause structures). For example, a `logical' form

study' (Albers', art', t-i)

that is intended to represent a predicate of studying that relates Anni Albers to the discipline art at some time t-i, might drive the lexicogrammar to select a clause with a transitive structure such as Albers studies art. A full representation of this aspect of a clause's meaning would at least require additional information concerning temporal relations (particularly with respect to the time of speaking in order to control grammatical tense decisions), logical scopes of various quantifiers, etc. Such information, which will here be termed ideational, must be available for a lexicogrammar to construct a content-appropriate sentence (or any other grammatical unit).

Semantic descriptions in fact offer many features desirable for an NLG input representation and are sometimes adopted as such. Figure 1 shows a skeleton semantics for the second sentence of the first Bauhaus text above as an example. This is written using the `Sentence Plan Language' notation (SPL: [Kasper: 1989]) originally developed for representing inputs to the grammatical component of the Penman text generation system [Mann: 1983b,Matthiessen and Bateman: 1991] and subsequently adopted in a number of generation projects. The form of such specifications is simple. A specification consists of one or more semantic instances (identified by variables: e1, x1, x2, etc.); each may be accompanied by a designated semantic type or list of types (bearing, female, etc.), separated from the variable with a ``/'', and be followed by a list of attributes or roles (e.g., :actee, :spatial-locating, etc.) each of which is filled by another semantic instance or set of instances. Thus the example in the figure shows a particular semantic instance (e1) that is of the semantic type of bearing and which has three semantic roles: one :actee (itself consisting of the semantic instance Anni-Albers of the semantic type female), and two related to time (a date) and place (a named city); other examples of generator input specifications are compared below.

 
  
(e1 / bearing
:actee (Anni-Albers / female)
:spatial-locating (x2 / city :name Berlin)
:temporal-locating (x3 / date :day 12 :month 6 :year 1899))
Example skeleton semantics for:``She was born in Berlin on 12 June 1899''

Semantic expressions at this level of abstraction are particularly designed for representing the semantics of natural language expressions in a manner appropriate for NLG. They are compatible with a wide range of possible surface expressions. The SPL specification shown in Figure 1, for example, is compatible with a range of possible sentences that include `Anni Albers was born there then', `It was on the 12th of June 1899 in Berlin that Anni Albers was born', the variants found in the second and third Albers texts above, and many others, as well as with several non-sentential realizations such as `the birth of Anni Albers in Berlin...'. Deciding among such alternatives can only be done in the context of a specific text and of particular communicative goals. Maintaining a semantic specification such as that shown therefore avoids awkward problems of `massaging', or post-editing, that arise whenever pre-built text fragments, strings, or templates stored directly as text building blocks are to be combined to form a `text' that exhibits appropriate textuality. Semantic-based inputs therefore allow a system to abstract away from particular sentences as they might need to appear in the concrete texts generated and so provide more support for variation. Particular choices of words, grammatical selections such as active or passive, word orders, pronominalizations, tense, focusing constructions, etc. that make a crucial contribution to the coherence and cohesion of a text all remain uncommitted.

Regardless of the detail of the propositional content representation, this information can never sufficiently constrain the options for expression that a lexicogrammar typically possesses. It is also necessary to go further to consider other kinds of meaning in order to gather sufficient constraints for guiding a lexicogrammar to select the particular surface strings appropriate for a particular context of use, textual position and communicative intention. This necessity of considering a wider spectrum of meaning types than that traditionally addressed in natural language processing has always been a distinguishing feature of NLG. Approaches to analysis, and approaches to generation derived from such approaches (including most `bidirectional' systems) have not yet incorporated this wider spectrum of meaning:3generation within such frameworks often relies on restricted lexicogrammatical resources, statistical distributions, defaults, or restrictions to the search space for producing sentences that may or may not be entirely appropriate to their context of use; one good overview of some of the additions that are needed in moving from analysis to generation within a theoretically bidirectional account using a larger grammar is available for the CLE system [Alshawi and Pulman: 1992, pp268--275]. In understanding the differences between analysis and generation, it is useful to bear in mind that whereas the semantic logical form representing a sentence is usually the primary target for analysis, in NLG it occupies a position relatively near the surface form of a text. Most of the difficult questions and research tasks in generation revolve around the problem of generating such a logical form on the basis of the communicative functional needs of the NLG system or its driving application: they are, in other words, upstream of the logical form.

This has resulted in some positions that appear, at first glance, irreconcileable. While some (e.g. [McDonald: 1993]) argue that the problem of mapping logical form to surface sentence is, actually, `trivial' or solved, and the main body of NLG lies elsewhere, others (e.g. [Shieber: 1993,van Noord and Neumann: 1997]) describe the logical form to surface sentence problem as unsolved with no convincing large-scale accounts available within NLG systems. These differences result from very different aims and requirements; potential users or developers of generation systems need to be aware of the restrictions adherence to either position brings. If the aim is formally guaranteed complete generation on the basis of any grammar, then this changes the task being considered. NLG has traditionally been concerned with the generation of any sentences that are appropriate for their contexts of use, rather than with decontextualized generation of all (and only) the possible sentences compatible with some input. Although theoretically attractive, complexity issues combined with large-scale lexicogrammars currently make this latter task prohibitively expensive. Consequently, if the aim is to generate more fluent and natural texts in a wider range of situations, the analysis-oriented bidirectional grammars have not yet demonstrated sufficient coverage or control of available variation and seem, at best, more suited to single sentence generation.

For the generation of more connected text, it is necessary to provide a richer kind of `functional semantics' than is typically found in logical form. We illustrate two further shades of functional semantics here: interpersonal meaning and textual meaning, both essential for adequate control of lexicogrammatical options during NLG. It is also occasionally suggested that such extended semantics might improve the status of logical form representations and provide necessary further constraints on their relationship with surface forms even without considering the particular needs of NLG (cf. [Shieber: 1993]).


next up previous contents
Next: Interpersonal decisions Up: Variation: how to describe Previous: Variation: how to describe   Contents
bateman 2002-09-21