Computer Science Dept., University College London, Gower Street, London WC1E 6BT
Abstract. Most natural language generation (NLG) has primarily focused upon the creation of text, and not upon the intention behind that textual generation. More recently, a number of studies have shown that eschewing intention in this way leads to serious shortcomings in the consequent power of the NLG system. This paper puts forward a new claim that argumentation relies crucially upon a competent means of representing and manipulating intentions, arising from the fact that an argument can be instigated with the sole intention to bring a particular audience to believe a particular belief. Persuasive communication itself, it is argued, is a phenomenon rooted in the intentional structures of the participants, and that therefore, to automate the construction of such discourse (and thence enter into multi-agent dialectic), it is essential to base any theory of computational argumentation on a core concept of intention.
AI research into discourse has concentrated primarily upon the structural relations operating at a number of levels, from the morphological through the syntactic and the inter-clause to the inter-sentential (see [Hovy93] for an overview). Such analyses have often been distilled into comprehensive frameworks such as Rhetorical Structure Theory, RST [Mann&Thompson86]. However, it has been recognised more recently that a text-driven, analytic theory such as RST fails to capture vital information concerning the intentions behind discourse structure, for despite the fact that intentional structure is rarely found blatantly manifest in the text, it is nevertheless crucial for participation in dialogue, due to its key role in justification, explanation, failure recovery and self referential discourse ([Moore&Paris94], [Young&Moore94]).
Recent work researching into methods by which argumentative dialogue might be generated has shown that a model of such discourse relies to an even greater extent upon the intentional structures behind the utterances ([Reed,Long&Fox96]). An argument can be defined simply in terms of the intention which motivates it (to bring a particular audience to believe a particular proposition). Frequently there will be a number of auxiliary intentions (concerning the speaker-hearer relationship, concealing or revealing vested interest, countering bias and scepticism, and a wide rage of other factors also dependent upon the context in which the argument is set), but however complex and replete an argument is with such subsidiary aims, the core intention still remains as the motivation for entering into the dialectic in the first place.
The work discussed in [Reed,Long&Fox96] is based around a hierarchical planner which goes through two distinct phases in order to create a plan for the argument (before undergoing further refinement to produce textual output): the first to generate a skeleton 'logical' structure containing both deductive and inductive reasoning; the second to convert that deep structure to an eloquent, persuasive surface structure though the application of rhetorical and stylistic operators and heuristics. The first, 'Argument Structure' (AS) level refines the initial intention (which will usually be of the form (BEL ?h ?prop) ) through the application of a number of abstraction operators, each comprised of preconditions, postconditions and a body (see [Fox&Long95] for details). The goals which represent intentions can be components of either the precondition list or the body: intentions can be manipulated both within a level of abstraction across an abstract plan, as well as between levels, permitting intentions to be 'inherited' from one level of abstraction to the next, in addition to the intentions invoked and manipulated within each level. Thus the final plan produced by the AS is maintained in such a way that each step can be justified on the basis of the intention to which it can be traced back: some intentions will therefore have a very limited scope, acting in a localised area of the plan, whereas others will cover the plan in a more global fashion (and since every action in the plan can ultimately be traced back to the initial intention, every action contributing to its ultimate fulfilment, that core intention has a scope encompassing the entire plan at every level of abstraction).
The output of the AS, then, is an intention-laden plan comprising all the necessary structure and content to produce argumentative discourse. However, this deep structure is exceedingly unlikely to be at all persuasive (and thus successful) if it were to be immediately converted into a string of utterances. Instead, it undergoes a second phase of processing, the 'Eloquence Generation' (EG) level, which should not be seen merely as effecting minor cosmetic modification to the output of the AS. Rather, it has a very large remit, ranging from complex structural modifications of subargument and premise ordering, through rhetorical figures and tropes such as repetition and alliteration, to stylistic decisions over syntax and vocabulary (see [Reed,Long&Fox96] for a more detailed examination of the functionality of the EG). It could be argued that this sort of processing (especially the reworking of the gross structure) should be conflated with the processing of the AS, but this is a mistaken view, since the AS is concerned with planning the content of an argument (including which logical relations to employ) whilst the EG contributes no new content, but rather, plans eloquent text. These are quite separate tasks, and are therefore appropriately marshalled through two distinct phases of processing.
The EG, despite planning with quite different operators, nevertheless places a similar reliance upon intentions. Rather than being straightforward intentions which specify a desired state in the world, as found at the AS level, the EG level is led chiefly by more general intentions which have influence over greater portions of the plan. For example, the AS level will often rely upon intentions concerned with the state of the hearer's beliefs, such as (BEL ?h ?prop), whereas the EG uses intentions expressing the need for brevity (which increase the frequency of enthymemes, and other contractions, whilst removing unnecessary examples and explanations). The AS will usually only have one global intention; the EG will have many (eg. to be amusing, to be eloquent, to be confusing, to use simple words, to counter scepticism, etc.). That is not to say that the EG will not employ any specific 'state-describing' intentions: those related to the ordering of premises, conclusions and subarguments, in particular, follow a similar pattern to those found at the AS level. In general, however, the EG intentions will be of a somewhat different style to those of the AS; in both cases though, their method of application and their importance is the same.
In conclusion, it can be seen that the computational generation of dialectic relies heavily upon the concept of intentions - both as a means of expressing core notions in an argument (such as the motivation for the argument's very existence), and as a means of dealing with the phenomena mentioned above: by having a plan structure which is replete with intentional information, it becomes possible to deal with justification, explanation and recovery from communication failure during the plan's execution, ie. during the dialogue. It is also interesting to note that by adopting this two stage approach to the planning of argumentative dialogue, the 'Stages of Argument Construction', as described in [Blair38] (a landmark precomputational treatise on rhetoric) are mirrored precisely, with the AS level corresponding to the stage of 'Invention', the EG to 'Arrangement' and then, finally, plan execution to 'Expression'.
[Blair38] Blair, H. Lectures on Rhetoric and Belles Lettres, Charles Daly, London, 1838
[Fox&Long95] Fox, M. & Long, D. "Hierarchical planning using abstraction", IEE Proc.-Control Theory Appl., 142 (3), 1995, pp197-210
[Hovy93] Hovy, E., "Automated discourse generation using discourse structure relations", Artificial Intelligence 63,1993, pp341-385
[Mann&Thompson86] Mann, W.C., & Thompson, S.A., "Rhetorical structure theory: description and construction of text structures" in Kempen, G. (ed) Natural Language Generation: New Results in Artificial Intelligence, Psychology and Linguistics, Kluwer, 1986, pp279-300
[Moore&Paris94] Moore, J.D. & Paris, C.L., "Planning Text for Advisory Dialogues, Capturing Intentional and Rhetorical Information", Computational Linguistics 19 (4), 1994, pp651-694
[Reed,Long&Fox96] Reed, C., Long, D. & Fox, M. "An Architecture for Argumentative Discourse Planning", presented at The 1st International Conference on Formal and Applied Reasoning (FAPR'96), to appear in Gabbay, D. & Ohlbach, H.J. Proceedings of FAPR'96, Lecture Notes in AI Vol. 1085, Springer-Verlag, 1996
[Young&Moore94] Young, R.M., Moore, J.D. "DPOCL: A principled approach to discourse planning", Proceedings of the 7th International Workshop on Natural Language Generation, Kennebunkport, Maine, 1994, pp13-20