Framing a Solution
BEL(CG, B E L (ES, BEL(CG, P) ) )
BEL(Audience, BEL(CG, BEL(ES, BEL(CG, -P))))
She knows she’s a goodguy He knows she’s a goodguy... ... but she doesn’t realise th a t... ... and he knows she doesn’t realise... ... or at least that's what the audience thinks! After JM has left tnwever, CG gets up - the shooting fiad been faked. CG and ES must have been in league with each other after alt. At this point the belief held must be changed to
BEL (Audience, BMB(CG, ES, P)))) They both know she’s a goodguy, and
both know that ttiey both know it.
Hkchcock envisaged the whole sftuation and realised ttiat it w as unusual and interesting. There were in total, five levels o f nested beliefs.
Figure 33 Deep nesting of beliefs
The ability to nest beliefs of a Cohen and Levesque form has been implemented in Wilks’s VIEWGEN system, (Lee and Wilks, 1997), (Ballim and Wilks, 1991), upon which the current work draws quite heavily. The characterisation of belief employed in the current work, however, diverges a little from the basic view espoused by Cohen and Levesque, and implemented in VIEWGEN, in permitting the expression of more than just the presence of existence of belief. Using negation as failure, the basic model supports a representation of an agent’s belief in a proposition (or its negation) or absence of belief in a proposition (or its negation) using B E L (A g , P) and -B E L (A g , P ) ,
respectively. However, it also becomes crucial to represent a situation in which an agent is undecided
about the validity of a claim - that is, the agent is aware of a fact but has not yet concluded whether the belief should be in the claim or its negation. This is in contrast to the state -b e l (A g , p ) , where an agent has no knowledge with regard to a claim. The state of ‘undecided’ is represented as B E L (A g ,
? P ) .
A straightforward use of Cohen and Levesque’s modal operators simplifies the integration of the belief models with the planning process used to guide the construction of argument. Plan operators used in argument generation (and NLG in general) based on advanced planning architectures rely heavily upon pre- and post-conditions composed, in part, of epistemic facts: Maybury’s communicative acts, for example, employ BELIEVE and KNOW-ABOUT (Maybury, 1992); Reithinger’s idiosyncratic MB, BGP, etc. (Reithinger, 1991); and Hovy’s influential work (Hovy, 1991)
operationalising Rhetorical Structure Theory (RST) relations (Mann and Thompson, 1988) using Cohen and Levesque’s BEL and BMB (and this approach is characteristic of all systems which draw upon RST - (Moore and Paris, 1989), (Paris, 1991), etc.). Although the current work follows this last precedent (though with RST playing a rather different role, as discussed in detail above in §3.1.4), the integration with the planning mechanism is somewhat different. Following (Moore and Paris, 1994), the distinction is drawn between communicative goals and linguistic goals - the former are characterised by their use of epistemic pre- and post-condition specifications, and represent intentions of the speaker. The latter, which arise from fulfilment of communicative goals, represent rhetorical relations or speech acts which can be directly realised into text.
The flat structure - mapping a communicative goal directly to a linguistic goal - is expanded in 3(fietO Tica, such that communicative goals form the substrate of the planning process at an abstract
level. Thus one communicative goal can be fulfilled by one or more further communicative goals, and so on. At the stage where conventional approaches such as (Moore and Paris, 1994) resolve linguistic goals, ü (fie to T ic ahas an intermediate layer of communicative goals which express the intention of the
speaker to make particular facts s a l i e n tto the hearer. Further planning is then invoked to decide upon appropriate linguistic goals to achieve the saliency; the subsequent realisation of those linguistic goals is beyond the scope of the current work for reasons put forward in §3.1.3, above.
Introducing goals of saliency as an interface between other communicative goals and linguistic goals offers a number of advantages over more traditional approaches. In the first place, it neatly accommodates the need for avoiding redundancy in discourse. In addition to generic dicta for conciseness such as Grice’s maxim of quantity (Grice, 1975), there are also specific considerations for argument generation. The Modus Brevis form of argument (Sadock, 1977) whereby the inference step in a Modus Ponens is omitted (discussed in relation to argument structure in (Cohen, 1987)), is an example of generic enthymematic contraction discussed by Aristotle (1926). In a syllogism, consisting by definition of two premises and one conclusion, it is frequent in natural language to find one component omitted (and, moreover, that including the omitted component would render the text cumbersome and repetitive). By introducing an intermediate level between communicative and linguistic goals, it is possible to effect heuristic control over which parts of an argument to make salient and which to leave implicit. Examples of the phenomenon and its production are given below in chapters four and five. Secondly, the approach also admits the generation of the reverse phenomenon. Marcu (1996a) notes that “Contrary to NLG wisdom, a system capable of generating persuasive text will also have to generate information that is known to the hearer”. Thus a communicative goal of persuading the hearer o f p could be fulfilled by making several facts - including some already known to be in the model of the hearer’s beliefs - salient to the hearer. Equally, it is quite possible for several goals of saliency to refer to the same proposition at various stages in the text, causing repetition of information in the text. Introducing redundancy into text has been recognised for discourse in general (Walker, 1992) - informationally redundant utterances (IRU’s) functioning either to offer evidence or to centre propositions. It is thus an important advantage that the model encompass generation of IRU’s. Finally, as explored in detail in the next two chapters, the division of communicative goals into those of belief and those of saliency facilitates precise control over the ordering of components, and in
m . FRAMING A SOLUTION 63
particular, enables the conclusion o f an argument to be ordered with respect to its prem ises (an important factor in determ ining persuasive effect, as discussed in (H ovland, 1957) inter alia).
3.2 Architecture
The key com ponent o f the IRhetorica system is the hierarchical planner based on A bN LP (Fox and Long, 1995). The planning process is conceptually divided into four stages o f abstraction, as show n in Figure 3.4. O perators ( AS Level 'N L ogical F a lla c ie s ^ Inductive y r EG Level S tylistic P e r s u a s i v e J RST Level R S T O p e r a to rs G ram mar, S yntax & Morphology
Hierarchical P lan n e r Beliefs
Model of H 's plan A S level S ystem Knowledge Model of H R S T /eve; Evaluator Analysis for knowledge update Analysis for h earer model update S y n ia c h c level
H
Figure 3.4 Architecture of the ^Rfietorica system
At the low est levels, the syntactic and m orphological structure is determined. This functionality com es within the remit o f the LOLITA system (Sm ith, 1994), which takes as input a Sem N et sem antic network (Shiu et a l , 1996). The structure o f this network can be directly mapped from an Rhetorical Structure (Mann and Thom pson, 1988) tree, with appropriate enhancements (including tags expressing various stylistic parameters such as those discussed in (H ovy, 1990), and b elow in chapter four). Consider, for exam ple, the RS tree in Figure 3.5(a) - w hich can be mapped on to the Sem N et representation in Figure 3.5(b) (the exam ple represents an amalgam ation o f those presented in (R eed et a l , 1997a) and (Mann and T hom pson, 1988)). The details o f the m echanism by w hich this can be accom plished, and evaluation o f its flexibility and expressive pow er, are beyond the scop e o f this work, since for the reasons previously put forward in §3.1.3, the current work is focusing entirely on the higher levels o f structure determination (and as explained in §3 .1 .2 , the assumption that this low er level functionality can be decoupled in this w ay is fully justified).
(a)
NONVOLinONAL CAUSE