Discourse Planning as an Optimization Process

1 I n t r o d u c t i o n Schema-based Natural Language Generation (NLG) systems, e.g., [Weiner, 1980; McKeown, 1985; Paris, 1988], determine the information to be presented based on common patterns of discourse. Goal-based planners, e.g., [Moore and Swartout, 1989; Cawsey, 1990], select a discourse operator if its prescribed effect matches a given communicative goal. If there is more than one such operator, the operator whose prerequisite information is believed by the user is preferred. However, if all the candidate operators require the generation of discourse that conveys some prerequisite information, the selection process is either random or the system designer determines in advance which operators should be preferred. In this paper, we cast the problem of planning discourse that achieves a given communicative goal as an application of an optimization algorithm. This approach supports the definition of different optimization objectives, such as generating (1) the most concise discourse; (2) the 'shallowest' discourse, i.e., discourse that requires the least amount of prerequisite information; or (3) the most concrete discourse, i.e., discourse with the most examples. The resulting mechanism is part of a discourse planning system called WISHFUL-II , which is a descendant of the WISHFUL system described in [gukerman and McConachy, 1993a]. Table 1 illustrates the discourse generated by our system using the concise and the shallow optimization objectives. Table 1. Sample Concise and Shallow 'Wallaby' Discourse i C o n c i s e Discourse Shal low Discourse W a l l a b i e s have a pouch , W a l l a b i e s a r e N a r s u p i a l s and which i s l i k e a pocket , they come f rom L u s t r a l i a . T h e y a r e l i k e k a n g a r o o s , They hop and t h e y a r e 3 f t . bu t t h e y a r e 3 f t . t a l l . t a l l . W a l l y i s a u a l l a b y . These texts were generated in order to convey the at tr ibutes type, habitat, body parts, height and transportation mode of the concept Wallaby to a user who owns a toy wallaby calhd Wally, and knows something about kangaroos, but is not familiar with the term pouch. The concise discourse conveys most of the intended information by means of a Simile between wallabies and kangaroos. The Simile also yields the erroneous inference that wallabies are the same height as kangaroos. To contradict this inference, the system asserts that wallabies are 3 ft. tall. Since the user does not know that kangaroos have a pouch, this is asserted, and since the user does not know what a pouch is, information which evokes this concept is presented. The shallow discourse, on the other hand, uses Wally (the toy wa.llaby) to convey the body parts of a wallaby without naming them explicitly. This information is complemented by Assertions about a wallaby's type, habitat , height and t ransporta t ion mode. In the next section, we present an overview of WISHFULII. In Section 3, we describe the discourse plaaming mechanism. We then discuss the results we have obtained, and present concluding remarks. 2 O v e r v i e w o f t h e S y s t e m WISHFUL-I I receives as input a conceptto be conveyed, e.g., Wallaby, a list of aspects tha t must be conveyed about this concept, e.g., habi ta t and body parts, and a desired level of expertise the user should at tain as a result of the presentation. WISHFUL-I I was used to generate descriptive discourse in various technical domains, such as chemistry, high-school algebra, animal taxonomy and Lisp. It produces multi-sentence paragraphs of connected English text. The discourse planning mechanism, which is the focus of this paper, generates a set of Rhetorical Devices (RDs), where each RD is a rhetorical action, such as Assert, Negate or Instantiate, applied to a proposition. This set of RDs is optimal with respect to a given optimization criterion, e.g., conciseness or depth. The following steps axe performed by WISHFUL-II . 1. C o n t e n t S e l e c t i o n WISHFUL-I I consults a model of the user 's beliefs in order to determine which propositions