Figure 44: Constraint Dependencies in Sample Sentence
It is evident from a quick look at Figure 28, repeated here as Figure 44, that constraints in that sentence are locally bundled. Grupo-Roche has nothing to do with Espana. The meaning of the one can be determined independently of the other. That is not to say that they cannot influence each other. For instance, Espana helps the analyzer choose the location meaning of en, which through a series of other interactions, could possibly influence the choice of meaning for adquirir, which, finally, could influence Grupo-Roche. However, this chain of dependencies is exactly what constraint-based mechanisms handles. As far as direct dependencies, though, the two words are not linked.
This is the general state of affairs in Natural Language. Government and Binding theory [Haegeman, 1991] is built on the assumption that one part of a text governs another, and interactions can only occur under this relationship. Government is restricted, among other things, to nodes that syntactically command other nodes. A node commands another node (again, among other things), if it is higher in the syntactic tree, and both are on the same path to the topmost node. This constraining property of syntax excludes non-governing relationships, which, in effect, partitions sentences into independent bundles. Smaller bundles can be combined into larger bundles as one moves higher up in the syntactic tree, where government domains become larger. This fact will be used to great advantage when constructing the subgraphs for solution synthesis, as described below.
It is fairly obvious that, at the sentence level, computational semantics tends to bundle dependencies into these subgraphs. What about larger sections of text? This research claims to be a step toward implementation of practical, large-scale, ``real'' computational semantic systems. Such systems eventually will address discourse issues. Can the claim that dependencies are locally bundled be maintained?
Yes, and no. [Grosz and Sidner, 1986] identify three aspects of discourse: the linguistic structure, the intentional structure and the attentional state. The first two seek to identify segments of text and give their purpose. In function, they are very similar to Rhetorical Structures [Mann and Thompson, 1983]. The attentional state, on the other hand, is an abstraction of the participants' focus of attention. This can be either global, or local [Grosz, 1981]. Knowledge of the attentional state is needed in reference resolution (and generation).
The rhetorical structure of a text links together chunks of text and
identifies the function of the composition. In Rhetorical Structure
Theory (RST), a nuclear section of text is joined to a satellite.
Constraints between the nucleus and satellite are typically
constraints between the main events of the main clauses of the
sub-texts involved. For a constraint-based analyzer/planner, this
simply adds an extra constraint link between the two sub-texts. In
practice, this may cause final decisions at the sentence level to be
delayed
until later in the discourse.
This, of course, is a desirable situation. Often decisions cannot be
made until the global purpose of a text or sub-text has been
determined. In fact, this discourse oriented processing is a main
driving force behind a constraint-based approach. If a 23 word
sentence produces millions of combinations, how many combinations
would a thousand word text produce? Constraints must be used to
intelligently prune the space of possibilities to a minimum, limiting
interaction between sentences only to the bare minimum.
Attempts at processing the attentional state have concentrated on local focus. ``Centering'' theory [Grosz, et al. 1983] is an attempt to constrain reference resolution to the immediate context. Such efforts have proven to be effective for many texts; however, it is recognized that local focus alone cannot solve all reference problems.
[Elhadad, 1990] argues that conversations are locally constrained. He
uses a constraint-based paradigm to
generate turns in a conversation. Each turn is linked to the previous
turn by five local constraints. He cites [Sacks, et al, 1978] to
support the contention that the most important characteristic of
dialog is that it is locally managed.
In practice, local focus and local dialog constraints can be tracked independently of the main analyzer/planner. Before each sentence is processed, the focus and dialog constraints can be calculated for that sentence. These constraints can then be added to the local processing of each sentence. Tracking global focus can also be added to this independent mechanism. Thus, these phenomena do not pose a problem for CSP techniques.
On the other hand, certain aspects of Text Generation such as planning sentence length are heavily influenced by global considerations. What has come before and what comes next, the complexity of the preceding text, the surface length of the realizations of sub-parts of the current sentence, as well as global considerations of style; all these impact on sentence boundary decisions. Some of these factors can be tracked independently, similar to focus and local dialog constraints. The complexity and surface length of the current text, and of the text yet to be processed, however, are difficult to measure until the surface forms have been generated. For instance, a precondition such as ``Sentence length < 25'' cannot be satisfied by a single effect, but only from the combination of many effects. Effects could be created that increment a global variable, which is then referenced by the precondition; however, this creates a situation where one sub-plan is constrained by every other sub-plan, which destroys the computational efficiency of HUNTER-GATHERER (see the ``Classes of Problems'' section below.) In these cases, constraint-based planners hold little advantage.
The best solution to this problem probably lies in a post-processor that can examine the output text and suggest revisions based on measures of global surface features (Robin, 1994; Inui et al., 1992). Another possible solution would be to eliminate decisions based on surface features, such as the number of words planned, and replace these constraints with semantically based ones, such as the number of concepts. Or, perhaps, with a little more thought, using the global variable approach described above might be implemented in a way that does not impact efficiency. We leave these matters of handling surface constraints for future research.