File Information
File: 05-lr/acl_arc_1_sum/cleansed_text/xml_by_section/abstr/90/w90-0118_abstr.xml
Size: 24,810 bytes
Last Modified: 2025-10-06 13:47:04
<?xml version="1.0" standalone="yes"?> <Paper uid="W90-0118"> <Title>Reassessing Rhetorical Abstractions and Planning Mechanisms 1</Title> <Section position="1" start_page="0" end_page="141" type="abstr"> <SectionTitle> Abstract </SectionTitle> <Paragraph position="0"> The utihty of rhetorical abstractions and certain text planning mechanisms were assessed from the standpoint of accounting for how an explainer chooses and structures content under multiple perspectives to meet knowledge communication goals. This paper discusses ways in which they were found to be inadequate, argues for greater emphasis on an epistemological level of analysis, and proposes a mixed architecture matching computational mechanisms to the explanation planning subtasks they are suited for.</Paragraph> <Paragraph position="1"> Introduction Our research is concerned with explanation in its broad sense, as &quot;the act or process of making plain or comprehensible; elucidation; clarification&quot; (American Heritage Dicgionary). In many physical science domains, an explanation can be based on a variety of models of the phenomenon being explained. These models differ on the type of properties and relationships emphasized; what ontology is used and whether the phenomenon is described at a macroscopic, microscopic, or atomic granularity; what factors are ignored or assumed to be constant; the use of general statements vs. concrete examples; and in general on what concepts function as primitives, providing the basis for understanding the topic phenomenon. Selection of an appropriate model of the topic is an important aspect of content selection which impacts on the interlocutor's comprehension of the explanation and on its appropriateness for his or her purposes. We use the term perspeetlve to refer to an abstract characterization of the kind of knowledge provided by a class of models. Our notion of perspective is a composite of distinctions made by Falkenhainer & Forbus (submitted), Stevens & Collins (1980) Stevens & Steinberg (1981) and White & Frederiksen (1989).</Paragraph> <Paragraph position="2"> 1 Copyright @ 1990, Daniel D. Suthers Permission granted to copy for non-profit academic purposes.</Paragraph> <Paragraph position="3"> Most existing work in explanation and text planning has been directed at other problems, and hence has utilized single-perspective knowledge bases to simplify the research. (Notable exceptions include McCoy, 1989 and McKeown et al. 1985.) A number of such research efforts emphasize rhetorical abstractions for the analysis of natural explanations and for expressing a theory of explanation (Hovy, 1988; Mann & Thompson, 1986; Maybury, 1988; McKeown, 1985; Moore, 1989). A variety of mechanisms for selecting and organizing the content of explanations have also been explored. This includes schema filling (McKeown 1985) and structure matching (Hovy 1988), graph traversal algorithms (Paris & McKeown 1986), and top-down expansion planning (Cawsey, 1989; Moore, 1989).</Paragraph> <Paragraph position="4"> Our own work on choosing explanatory content from multiple-perspective knowledge bases has uncovered some limitations of rhetorical abstractions, and led us to question previous applications of the computational mechanisms listed above. The purpose of this paper is to present our perception of the roles and limitations of these items, and suggest some alternatives. (We do not emphasize our work on perspective and content organization here: see Suthers & Woolf, 1990.) We begin with an example, used to illustrate some of the problems.</Paragraph> <Paragraph position="5"> An Example The following example explanation will be used to illustrate our points. The domain is elementary electricity. We emphasize the communication of an understanding of concepts such as &quot;charge&quot;, &quot;current&quot;, and &quot;capacitance&quot; within the context of qualitative reasoning about the behavior of simple circuits and their components. The explanation is an edited version of a human protocol.</Paragraph> <Paragraph position="6"> QI: How does a capacilor sC/ore charge? El: A capacitor E2: can be thought of as two fiat metallic plates</Paragraph> <Paragraph position="8"/> <Paragraph position="10"> E7: then positive charge builds up on one plate and negative charge on the other, E8: until the voltage across the capacitor is equal to the voltage of the battery.</Paragraph> <Paragraph position="11"> PS9: The charge stays there when you disconnect the battery. null One possible rhetorical analysis of the first explanation, using relations from McKeown (1985) and Mann & Thompson (1986), is given in Figure 1. (Whether or not this is an optimal analysis is not the point; rather we are concerned with the role of rhetorical abstractions and various characterizations of the content planning task * in accounting for the explanation.) The analysis does point out some features of interest. For example, we note that the explanation starts with Background material (El-E4) before proceeding to the primary explanation (E5-E9). The Background relation in this case describes the high level organization of the explanation. The question is how this relation should be manifest in the mechanism for generating the explanation. Also of interest are the explainer's determination that a process account of how the capacitor carries out its function is appropriate; use of an abstract structural model of the capacitor, simplified to what is needed to support the process account; and use of a concrete situation for the enablement condition of the process.</Paragraph> <Section position="1" start_page="137" end_page="138" type="sub_section"> <SectionTitle> Problems with Rhetorical Abstractions </SectionTitle> <Paragraph position="0"> Rhetorical relations were an important development in explanation research, since they provided a first pass at abstractions for a general description of explanatory structure, and in bringing various roles of the parts of explanation into the foreground, pointed out phenomena in need of further study. They are also useful via their &quot;relational propositions&quot; (Mann & Thompson 1983), for conveying propositional information implicitly in the structure of the text, hence reducing its length and redundancy. However, we claim that rhetorical abstractions fail to make the necessary distinctions for further advances in a theoretical understanding of explanation.</Paragraph> <Paragraph position="1"> Potpourri. Rhetorical abstractions are descriptive of explanatory tezt, i.e. the end product of some expla- null nation generation process, and so describe with one device structure due to a variety of distinct knowledge sources bearing on such a process. These knowledge sources operate on different levels of information, and hence need to be separated in a theory of explanation generation. For example, (drawing on relations in McKeown, 1985 and Mann & Thompson~ 1986) some correspond directly to the fundamental structure of domain objects and processes (e.g. Constituency and Causality), while others are about derived relations between concepts which may vary according to the context (e.g.</Paragraph> <Paragraph position="2"> Comparison). Relations such as Amplification, Background, Evidence, and Illustration are primarily about relationships between propositions which arise in part out of consideration of what the interlocutor knows and needs to know to better grasp a point. Illocutionary acts are involved in the relations as well (most blatantly, Concession; others are not themselves illocutionary acts but only make sense in the context of certain such acts).</Paragraph> <Paragraph position="3"> Finally, relations such as Topic and Conclusion appear to be due to conventions governing writing style which direct focus of attention. Grosz & Sidner (1986) made similar criticisms from the standpoint of characterizing discourse coherence. They suggested that each rhetorical relation combines domain information with certain general relations between propositions, between actions, and between intentions.</Paragraph> <Paragraph position="4"> Implicit Features Unaccounted For. Rhetorical abstractions are also inappropriate for a theory of content selection because, in describing the final text, they fail to identify important relations between the chosen content and external material, such as what is known about the user, or information left o~$ of the explanation. For example, E5-E6 is more specific than is necessary and includes a concrete example. A more general and accurate way to state the condition for initiation of the charging process would be &quot;If a voltage is applied to the plates ...'. However, the explainer has opted to replace this with one of the many particular configurations which meet the condition. A rhetorical analysis of the text canno$ even $ell us $ha~ ~his has happened, let alone why, because it does not describe relations between the contents of the text and what is no~ included, viz., other models of the process being described. It can only report that an illustration is being used. Another example is provided by E9. Retention of charge when the voltage is removed is what is meant by &quot;storage&quot; in this case, so the fact expressed in E9 is essential to answering the question. Rhetorically, we can only identify relationships E9 has to the rest of the text, e.g. that it is a Conclusion. This does not illuminate the relationship between its content and the goal of the explanation.</Paragraph> <Paragraph position="5"> Epistemological Analysis. The success of an explanation is primarily a function of choice and organization of knowledge. Hencc, to account for how these choices further knowledge communication goals, one must examine explanation in part from an epistemological standpoint. Such an analysis examines how explanations are guided by: * the types of knowledge in a given domain, and its logical and etiological structure (Rissland, 1978); * the types of knowledge which, in principle, could fulfill a given request for information; * the role of an individual's knowledge in understanding new concepts and situations, and hence in understanding a given explanation (Paris, 1987); and * the ways in which individuals are willing or able to undertake conceptual change (Goldstein, 1979; Hewson, 1981; White & Frederiksen, 1989; van-Lchn, 1987).</Paragraph> <Paragraph position="6"> As discussed in Suthers (1989), most previous work has offered solutions to the subproblems of explanation in the form of mechanisms and data structures which are in part the result of, rather than the expression of, epistcmological considerations. Epistemological problems have been avoided through direct selection of content based on well-formulated queries; the simplicity of the knowledge bases used, which only permit one way of discussing each topic; and through implicit conflation of epistemological constraints on organization of the explanation with those of rhetorical and linguistic origin. A major goal of our research (Suthers & Woolf, 1990) is an explicit theory of the epistemological structure of the activity of explaining.</Paragraph> </Section> <Section position="2" start_page="138" end_page="141" type="sub_section"> <SectionTitle> Roles of Computational Mechanisms </SectionTitle> <Paragraph position="0"> In this section we illustrate how inclusion of background material and the use of multiple perspectives pose problems for various mechanisms in the literature, and suggest a mixed architecture solution.</Paragraph> <Paragraph position="1"> Structure Matching. By &quot;structure matching&quot; we mean methods where abstract descriptions of the structure of explanations are matched to a collection of propositions (or other content to be expressed) in order to organize this material. This includes bottom-up composition of structural units (Hovy, 1988) and schema filling (McKeown, 1985). We question their adequacy for accounting for the prerequisite structure of explanations, such as the Background relation of the example. In our view, the explanation is organized this way because the explainer recognized in his process model (expressed in E5-Eg) concepts the interlocutor may not be familiar with (the parts of a capacitor), and then added material prerequisite to understanding the process explanation (the structural description expressed in El-E4). The background material is not automatically part of the relevant knowledge pool for this question, and structure matching methods leave choice of content and perspective to other mechanisms. Suppose, then, that some other mechanism accounts for inclusion of the background in the pool. It is included as background by virtue of relationships between the interlocutor's assumed knowledge state and the conceptualizations contained in the first attempt at a knowledge pool. Pattern matching techniques which are ignorant of such relationships and see only the composite pool would be unable to identify the part of the knowledge pool which plays the role of &quot;background&quot;, and account for placement of background before primary material.</Paragraph> <Paragraph position="2"> Graph Traversal. These are algorithms for selectively following links in a knowledge base, with the path so traced out providing the content and structure of the explanation (Paris & McKeown 1986). Such methods model how an explanation exploits the structure of knowledge. They implicitly embody the heuristic that it will be easier for the interlocutor to reconstruct the knowledge if it is presented such that each unit is introduced in relation to the previous unit. For example, parts of an object are introduced in relation to containing or adjacent parts, and process descriptions organized to follow temporal and causal relations in the forward direction. However, graph traversal is limited to modeling local organization, or global organization which is a composite of local choices. Traversal methods don't naturally extend to global organization which occurs at a higher level of abstraction than the links followed, e.g. the presentation of a coherent structural model before a process account begins.</Paragraph> <Paragraph position="3"> Top-Down Goal Expansion. Top-down expansion of a discourse goal (Cawsey, 1989; Moore, 1989) is a stronger candidate for a uniform mechanism for explanation, integrating content selection and structuring. The background problem can be handled with preconditions on plan operators, as Cawsey does. It simplifies modeling explanations such as our example if one can specify the kind of background knowledge required in preconditions at the highest level of plan operators. To illustrate, consider a rhetorical plan operator which contains an optional satellite for Background but does not specify what constitutes background knowledge. Expansion of the satellite would have to be predicated on a comparison of the content selected by expansion of the nucleus with the user model. This decision could not be made at the time the operator is selected by the planning system, since the knowledge pool for the nucleus would not have been selected yet. Instead, one would have to place the satellite decision on hold, expand the nucleus, collect together the knowledge selected at the leaves of its expansion, and perform the comparison before deciding on the satellite. (One could make the decisions concerning the need for background locally to each leaf, avoiding the need for a high level decision depending on knowledge selected at many localities. But then one could not model the structure of explanations such as this one, where the need for prerequisite material is anticipated and provided in advance as a coherent model, rather than as an interruption to the flow of the process description.) Then, if it was decided that some background was required, expansion of the satellite would have to occur under a binding of some variable in the satellite to the concepts for which background is required.</Paragraph> <Paragraph position="4"> With regards to perspective, some problems emerge.</Paragraph> <Paragraph position="5"> Content selection in top-down expansion can be influenced by the current perspective if some mechanism for sharing perspective decisions across the expansion tree is provided. However, neither the choice of perspective nor the actual mechanism by which it influences content selection are modeled appropriately by top-down expansion. Choice of perspective tries to balance the (sometimes conflicting) constraints of adequacy and comprehensibility. Adequacy constraints come from examination of the informative goals (McKeown et al.</Paragraph> <Paragraph position="6"> 1985). Comprehensibility requires answering the question: given the concepts which have been used in the dialogue so far, and/or which the explainer has evidence the interlocutor is familiar with, what other concepts are also likely to be familiar? This suggests a strengthof-association mechanism (McCoy 1989), which we comment on further in the next section. As Hovy (1990) points out, top-down planning is prescriptive, and does not handle conflicting goals easily. The influence of perspective on content selection involves what Hovy calls restrictive planning: some perspective goals, once generated, remain active throughout a span of discourse, and operate as preferences applied to choice points in content selection. They cannot be erased once satisfied, and they may change dynamically. Finally, McDonald & Pustejovsky (1985) point out that a uniform mechanism may not be desirable, as it incorrectly implies that all information is equally available during each point of the text planning process. A principled match of distinct computational mechanisms to each subtask of the explanation planning t&sk is one way to express one's theory of what kind of process explanation is. For these reasons, we postulate a separate level of planning for co-ordinating perspective decisions.</Paragraph> <Paragraph position="7"> A Mixed Architecture. In our current approach, planning occurs at two granularities: selection and retrieval of coherent packages of knowledge similar to Suthers' (1988a,b) &quot;views&quot; or Souther, Acker, Lester, & Porter's (1989) &quot;viewpoints&quot;; and editing and ordering the propositions and examples which make up these views. Planning at the granularity of views is concerned with purely epistemological constraints on identification of appropriate content, including choice of perspective and prerequisite explanations. At the finer granularity, further epistemological constraints governing the comprehensibihty of the explanation are apphed, and rhetorical and linguistic constraints play a role as well.</Paragraph> <Paragraph position="8"> The question at hand is: wA~z~ sort of t~zsl~ is cor~ter~ plar~r~ir~g ? We postulate several, and comment on possible computational mechanisms for each. Figure 2 gives the rough relationships between the different tasks. The process is reminiscent of case-based design, where one identifies and refines design specifications, and retrieves and reconfigures a previous solution to fit the circumstances. null An explainer must first identify at least some of its goals. These are of two types, as discussed above: prescriptive informative goals, and restrictive goals such as comprehensibihty. We mentioned that the latter suggests a concept association mechanism. This could be done by activating weighted links between concepts, or though intermediate frames of attribute weights, as in McCoy (1989). However, one would have to install a potentially combinatorial number of associations between concepts. Because we wish to make an epistemological theory explicit and implement it with an abstract interface to the knowledge base, we are investigating use of a mechanism akin to prototype induction, to generate an abstract description of the desired perspective from the user model and dialogue history.</Paragraph> <Paragraph position="9"> At the view granularity, content selection may be seen as refinement of explanatory goals into a specification of the appropriate addition to the relevant knowledge pool. This refinement includes consideration of what type of knowledge, in principle, could fulfill the informative goal; of the concepts the interlocutor is likely to understand; and of the models which have already been shared in the dialogue. We are attempting to treat this refinement task as top-down planning, though our work has not progressed fax enough to comment further on this.</Paragraph> <Paragraph position="10"> Retrieval requires a knowledge-base specific mechanism: its only theoretical importance to us is that it correctly operationalize the dimensions used to describe the desired view (Suthers 1988a,b). We are examining a variant of compositional modehng (Falkenhalner & Forbus, submitted) for this purpose.</Paragraph> <Paragraph position="11"> Once a knowledge pool is available, some data driven activities occur at a finer granularity, resulting in modification of the relevant knowledge pool. This includes filtering activities, such as removing particular propositions likely to be famihax to the interlocutor; and augmenting activities, such as illustrating abstract statements with examples. These are oppor~g~istic planning tasks, that can be approached with critics which match to the knowledge pool and user model, and specify replacements or deletions to be made. As shown in figure 2, data driven operators may also reinvoke content planning at the refinement level to access material in a different type of model. For example, we have seen how process propositions may involve use of structural concepts the interlocutor is not likely to understand, causing the explainer to plan a prerequisite explanation. An explicit record of the prerequisite relations between views is created, important for ordering the explanation.</Paragraph> <Paragraph position="12"> Ordering is sensitive to prerequisite and illustration links installed by the modification processes. Hence figure 2 should not be interpreted as a claim that ordering considerations do not arise during content selection. The ordering task involves two kinds of processes. One embodies epistemological constraints by ezploi~ir~g existing structure in the knowledge pool. As discussed previously, techniques for traversing links in the knowledge pool apply here. The result will likely be a partial ordering. Further ordering requires imposition of structure for rhetorical, linguistic, and pictorial reasons during realization (generation of text and graphics).</Paragraph> <Paragraph position="13"> Matching of rhetorical patterns to the knowledge pool may be more appropriate at this later stage.</Paragraph> <Paragraph position="14"> In summary, we have argued for a mixed architecture matching prescriptive, restrictive, and opportunistic mechanisms to explanation subtasks. Coordination of these diverse processes may, at the implementation level, require an agenda control mechanism as in Nirenburg, Lesser, & Nyberg (1989).</Paragraph> <Paragraph position="15"> Conclusions Single-perspective knowledge bases, i.e. those which provide only a single conceptual basis for a given description or explanation, have dominated existing work in planning expository text. This research has overemphasized rhetorical abstractions as the basis for theories of content planning, and used computational formalisms such as top-down goal expansion, traversal algorithms, and opportunistic structure matching which, taken alone, fail to fully account for the search for appropriate conceptualizations during content selection.</Paragraph> <Paragraph position="16"> We suggest a greater emphasis on epistemological abstractions, and viewing content selection in terms of identification, refinement, retrieval, modification, and ordering tasks. Rhetorical abstractions retain a place in initial analyses of explanations, to point out features in need of further study, and in the later stages of text planning, to further constrain a partial ordering of content and implicitly convey content via textual structure. Finally, the devices of top-down goal expansion, traversal algorithms, and structure matching retain potential utility for high level content planning, exploiting the structure of knowledge when ordering an explanation, and further ordering an explanation on a rhetorical basis, respectively.</Paragraph> </Section> </Section> class="xml-element"></Paper>