<?xml version="1.0" standalone="yes"?>
<Paper uid="P89-1025">
  <Title>PLANNING TEXT FOR ADVISORY DIALOGUES&amp;quot;</Title>
  <Section position="4" start_page="203" end_page="203" type="metho">
    <SectionTitle>
WHY A DETAILED TEXT PLAN?
</SectionTitle>
    <Paragraph position="0"> In order to handle follow-up questions that may arise if the user does not fully understand a response given by the system, a generation facility must be able to determine what portion of the text failed to achieve its purpose. If the generation system only knows the top-level discourse goal that was being achieved by the text (e.g., persuade the hearer to perform an action), and not what effect the individual parts of the text were intended to have on the hearer and how they fit together to achieve this top-level goal, its only recourse is to use a different strategy to achieve the top-level goal.</Paragraph>
    <Paragraph position="1"> It is not able to re-explain or clarify any part of the explanation. There is thus a need for a text plan to contain a specification of the intended effect of individual parts of the text on the hearer and how the parts relate to one another. We have developed a text planner that records the following information about the responses it produces: * the information that Grosz and Sidner (1986) have presented as the basics of a discourse structure: - intentional structure: a representation of the effect each part of the text is intended to have on the hearer and how the complete text achieves the overall discourse purpose (e.g., describe entity, persuade hearer to perform an action).</Paragraph>
    <Paragraph position="2"> - attentional structure: information / about which objects, properties and events are salient at each point in the discourse. User's follow-up questions are often ambiguous.</Paragraph>
    <Paragraph position="3"> Information about the attentional state of the discourse can be used to disambiguate them (cf. (Moore and Swartout, 1989)).</Paragraph>
    <Paragraph position="4"> * in addition, for generation we require the following: - rhetorical structure: an agent must understand how each part of the text relates rhetorically to the others. This is necessary for linguistic reasons (e.g., to generate the appropriate clausal connectives in multi-sentential responses) and for responding to requests for elaboration/clarification. null * assumption information: ad'vicegiving systems must take knowledge about their users into account.</Paragraph>
    <Paragraph position="5"> However, since we cannot rely on having complete user models, these systems may have to make assumptions about the hearer in order to use a particular explanation strategy. Whenever such assumptions are made, they must be recorded.</Paragraph>
    <Paragraph position="6"> The next sections describe this new text planner and show how it records the information needed to engage in a dialogue. Finally, a brief comparison with other approaches to text generation is presented.</Paragraph>
  </Section>
  <Section position="5" start_page="203" end_page="203" type="metho">
    <SectionTitle>
TEXT PLANNER
</SectionTitle>
    <Paragraph position="0"> The text planner has been developed as part of an explanation facility for an expert systern built using the Explainable Expert Systems (EES) framework (Swartout and Smoliar, 1987). The text planner has been used in two applications. In this paper, we draw our examples from one of them, the Program Enhancement Advisor (PEA) (Neches et al., 1985). PEA is an advice-giving system intended to aid users in improving their Common Lisp programs by recommending transformations that enhance the user's code. 1 The user supplies PEA with a program and indicates which characteristics of the program should be enhanced (any combination of readability, maintainability, and efficiency). PEA then recommends transformations. After each recommendation is made, the user is free to ask questions about the recommendation.</Paragraph>
    <Paragraph position="1"> We have implemented a top-down hierarchical expansion planner (d la Sacerdoti (1975)) that plans utterances to achieve discourse goals, building (and recording) the intentional, attentional, and rhetorical structure of the generated text. In addition, since the expert system explanation facility is intended to be used by many different users, the text planner takes knowledge about the user into account. In our system, the user model contains the user's domain goals and the knowledge he is assumed to have about the domain.</Paragraph>
  </Section>
  <Section position="6" start_page="203" end_page="205" type="metho">
    <SectionTitle>
THE PLAN LANGUAGE
</SectionTitle>
    <Paragraph position="0"> In our plan language, intentional goals are represented in terms of the effects the speaker intends his utterance to have on the hearer.</Paragraph>
    <Paragraph position="1"> Following Hovy (1988a), we use the terminology for expressing beliefs developed by Cohen and Levesque (1985) in their theory of rational interaction, but have found the need to extend the terminology to represent the types of intentional goals necessary for the kinds of responses desired in an advisory setting.</Paragraph>
    <Paragraph position="2"> Although Cohen and Levesque have subsequently retracted some aspects of their theory of rational interaction (Cohen and Levesque, 1987), the utility of their notation for our purposes remains unaffected, as argued in (Hovy, 1989). 2 a PEA recommends transformations that improve the 'style' of the user's code. It does not attempt to understand the content of the user's program.</Paragraph>
    <Paragraph position="3"> 2Space limitations prohibit an exposition of their terminology in this paper. We provide English paraphrases where necessary for clarity. (BR8 S II x) should be read as 'the speaker believes the speaker and hearer mutually believe x.'  Rhetorical structure is represented in terms of the rhetorical relations defined in Rhetorical Structure Theory (RST) (Mann and Thompson, 1987), a descriptive theory characterizing text structure in terms of the relations that hold between parts of a text (e.g., CONTRAST, MOTIVATION). The definition of each RST relation includes constraints on the two entities being related as well as constraints on their combination, and a specification of the effect which the speaker is attempting to achieve on the hearer's belids. Although other researchers have categorized typical intersentential relations (e.g., (Grimes, 1975, Hobbs, 1978)), the set of relations proposed by RST is the most complete and the theory sufficiently detailed to be easily adapted for use in generation.</Paragraph>
    <Paragraph position="4"> In our plan language, each plan operator consists of: an effect: a characterization of what goai(s) this operator can be used to achieve. An effect may be an intentional goal, such as persuade the hearer to do an ac~ionorarhetorical relation, such as provide motivation for an action.</Paragraph>
    <Paragraph position="5"> a constraint list: a list of conditions that must be true before the operator can be applied. Constraints may refer to facts in the system's knowledge base or in the user model.</Paragraph>
    <Paragraph position="6"> * a nucleus: the main topic to be expressed. The nucleus is either a primitive operator (i.e., speech acts such as inform, recommend and ask) or a goal intentional or rhetorical) which must be ther expanded. All operators must contain a nucleus.</Paragraph>
    <Paragraph position="7"> * satellites: subgoal(s)that express additional information which may be needed to achieve the effect of the operator.</Paragraph>
    <Paragraph position="8"> When present, satellites may be specified as required or optional.</Paragraph>
    <Paragraph position="9"> Examples of our plan operators are shown in Figures 1 and 2. The operator shown in Figure 1 can be used if the speaker (S) intends to persuade the hearer (H) to intend to do some act. This plan operator states that if an act is a step in achieving some domain goal(s) that the hearer shares, one way to persuade the hearer to do the act is to motivate the act in terms of those domain goals. Note that this plan operator takes into account not only the system's knowledge of itself, but also the system's knowledge about the user's goals, as embodied in a user model. If any domain goals that satisfy the constraints are found, this operator will cause the planner to post one or more MOTIVATION subgoals. This plan operator thus indicates that one way to achieve the intentional goal of persuading the hearer to perform an action is by using the rhetorical means MOTIVATION.</Paragraph>
    <Paragraph position="10">  value to any generalized-variable. Serq can only be used to assign a value to a simple-variable. A generalized-variable is a storage location that can be named by any accessor function.</Paragraph>
    <Paragraph position="11"> What is a generalized variable? For example, the car and cdr of a cons are generalized-variables, named by the accessor functions car and cdr. Other examples are an element of an array or a component of a structure.</Paragraph>
    <Paragraph position="12">  Plans that achieve intentional goals and those that achieve rhetorical relations are distinguished for two reasons: (1) so that the completed plan structure contains both the intentional goals of the speaker and the rhetorical means used to achieve them; (2) because there are many different rhetorical strategies for achieving any given intentional goal. For example, the system has several plan operators for achieving the intentional goal of describing a concept. It may describe a concept by stating its class membership and describing its attributes and its parts, by drawing an analogy to a similar concept, or by giving examples of the concept. There may also be many different plan operators for achieving a particular rhetorical strategy. (The planner employs selection heuristics for choosing among applicable operators in a given situation (Moore and Swartout, 1989).) Our plan language allows both general and specific plans to be represented. For example, Figure 2 shows a plan operator for achieving the rhetorical relation MOTIVATION.</Paragraph>
    <Paragraph position="13"> This is a very specific operator that can be used only when the act to be motivated is a replacement (e.g., replace sezq with sezf).</Paragraph>
    <Paragraph position="14"> In this case, one strategy for motivating the act is to compare the object being replaced and the object that replaces it with respect to the domain goal being achieved. On the other hand, the operator shown in Figure 3 is general and can be used to achieve mutual belief of any assertion by first informing the hearer of the assertion and then, optionaUy, by persuading him of that fact. Because we allow very general operators as well as very specific ones, we can include both domain-independent and domain-dependent strategies.</Paragraph>
  </Section>
  <Section position="7" start_page="205" end_page="208" type="metho">
    <SectionTitle>
A DETAILED EXAMPLE
</SectionTitle>
    <Paragraph position="0"> Consider the sample dialogue with our system shown in Figure 4, in which the user indicates that he wishes to enhance the maintainability of his program. While enhancing maintainability, the system recommends that the user perform the act replace-I, namely 'replace setq with serf', and thus posts the intentional goal (BMB S H (GOAL H Evenzually(DONE H replace-I))). This discourse goal says that the speaker would like to achieve the state where the speaker believes that the hearer and speaker mutually believe that it is a goal of the hearer that the replacement eventually be done by the hearer.</Paragraph>
    <Paragraph position="1"> The planner then identifies all the operators whose effect field matches the discourse goal to be achieved. For each operator found, the planner checks to see if all of its constraints are satisfied. In doing so, the text planner attempts to find variable bindings in the expert system's knowledge base or the user model that satisfy all the constraints in</Paragraph>
    <Paragraph position="3"> the constraint list. Those operators whose constraints are satisfied become candidates for achieving the goal, and the planner chooses one based on: the user model, the dialogue history, the specificity of the plan operator, and whether or not assumptions about the user's beliefs must be made in order to satisfy the operator's constraints.</Paragraph>
    <Paragraph position="4"> Continuing the example, the current discourse goal is to achieve the state where it is mutually believed by the speaker and hearer that the hearer has the goal of eventually executing the replacement. This discourse goal can be achieved by the plan operator in Figure 5. This operator has no constraints. Assume it is chosen in this case. The nucleus is expanded first, 3 causing (RECOMMEND S H replace-l) to be posted as a subgoal. RECOMMEND is a primitive operator, and so expansion of this branch of the plan is complete. 4 Next, the planner must expand the satellites. Since both satellites are optional in this case, the planner must decide which, if any, are to be posted as subgoals. In this example, the first satellite will not be expanded because the user model indicates that the user is ca31n some cases, such as a satellite posting the rhetorical relation background, the satellite is expanded first.</Paragraph>
    <Paragraph position="5"> +At this point, (RECOMMEND S H replace-l) must be translated into a form appropriate as input.to the realization component, the Penman system (Mann, 1983, Kasper, 1989). Based on the type of speech act, its arguments, and the context in which it occurs, the planner builds the appropriate structure. Bateman and Paxis (1989) have begun to investigate the problem of phrasing utterances for different types of users. pable of performing replacement acts. The second satellite is expanded, s posting the intentional subgoal to persuade the user to perform the replacement. A plan operator for acldeving this goal using the rhetorical relation MOTIVATION was shown in Figure i.</Paragraph>
    <Paragraph position="6"> When attempting to satisfy the constraints of the operator in Figure 1, the system first checks the constraints (GOAL S ?domain-goal) and (STEP replace-1 ?domain-goal). These constraints state that, in order to use this operator, the system must find an expert system goal, ?domain-goal, that replace-I is a step in achieving.</Paragraph>
    <Paragraph position="7"> This results in several possible bindings for the variable ?domain-goal. In this case, the applicable system goals, listed in order from most specific to the top-level goal of the system, are shown in Figure 6.</Paragraph>
    <Paragraph position="8"> The last constraint of this plan operator, (BMB S H (GOAL H ?domain-goal)), is a constraint on the user model stating that the speaker and hearer should mutu~IIy believe that ?domain-goal is a goal of the hearer.</Paragraph>
    <Paragraph position="9"> Not all of the bindings found so far will satisfy this constraint. Those which do not will not be rejected immediately, however, as we do not assume that the user model is complete. Instead, they will be noted as possible bindings, and each will be marked to indicate that, if this binding is used, an assumption is being made, namely that the binding of Sin other situations, the system could choose not to expand this satellite and await feedback from the user instead (Moore and Swartout, 1989).</Paragraph>
    <Paragraph position="11"> In this example, since the user is using the system to enhance a program and has indicated that he wishes to enhance the maintainability of the program, the system infers the user shares the top-level goal of the system (enhance-program), as well as the more specific goal enhance-mainZainabilizy. Therefore, these are the two goals that satisfy the constraints of the operator shown in Figure I.</Paragraph>
    <Paragraph position="12"> The text planner prefers choosing binding environments that require no assumptions to be made. In addition, in order to avoid explaining parts of the reasoning chain that the user is familiar with, the most specific goal is chosen. The plan operator is thus instantiated with enhance-mainzainability as the binding for the variable ?domain-goal. The selected plan operator is recorded as such, and all other candidate operators are recorded as untried alternatives.</Paragraph>
    <Paragraph position="13"> The nucleus of the chosen plan operator is now posted, resulting in the subgoal (MOTIVATION replace-1 enhancemainZainability). The plan operator chosen for achieving this goal is the one that  was shown in Figure 2. This operator motivates the replacement by describing differences between the object being replaced and the object replacing it. Although there are many differences between sezq and serf, only the differences relevant to the domain goal at hand (enhance-mainzainabilizy) should be expressed. The relevant differences are determined in the following way. From the expert system's problem-solving knowledge, the planner determines what roles eezq and eezf play in achieving the goal enhance-maintainabilizy. In this case, the system is enhancing maintainability by applying transformations that replace a specific construct with one that has a more general usage. SeZq has a more specific usage than sezf, and thus the comparison between sezq and sezf should be based on the generality of their usage.</Paragraph>
    <Paragraph position="14"> Finally, since the term generalized-variable has been introduced, and the user model indicates that the user does not know this term, an intentional goal to define it is posted: (BMB S H (KNOW H generalized-variable)). This goal is achieved with a plan operator that describes concepts by stating their class membership and describing their attributes. Once completed, the text plan is recorded in the dialogue history. The completed text plan for response (3) of the sample dialogue is shown in Figure 7.</Paragraph>
  </Section>
  <Section position="8" start_page="208" end_page="209" type="metho">
    <SectionTitle>
ADVANTAGES
</SectionTitle>
    <Paragraph position="0"> As illustrated in Figure 7, a text plan produced by our planner provides a detailed representation of the text generated by the system, indicating which purposes different parts of the text serve, the rhetorical means used to achieve them, and how parts of the plan are related to each other. The text plan also contains the assumptions that were made during planning. This text plan thus contains both the intentional structure and the rhetorical structure of the generated text. From this tree, the dominance and saris/actionprecedence relationships as defined by Grosz and Sidner can be inferred. Intentional goals higher up in the tree dominate those lower down and a left to right traversal of the tree provides satisfaction-precedence ordering.</Paragraph>
    <Paragraph position="1"> The attentional structure of the generated text can also be derived from the text plan.</Paragraph>
    <Paragraph position="2"> The text plan records the order in which topics appear in the explanation. The global variable *local-contezt ~ always points to the plan node that is currently in focus, and previously focused topics can be derived by an upward traversal of the plan tree.</Paragraph>
    <Paragraph position="3"> The information contained in the text plan is necessary for a generation system to be able to answer follow-up questions in context.</Paragraph>
    <Paragraph position="4"> Follow-up questions are likely to refer to the previously generated text, and, in addition, they often refer to part of the generated text, as opposed to the whole text. Without an explicit representation of the intentional structure of the text, a system cannot recognize that a follow-up question refers to a portion of the text already generated. Even if the system realizes that the follow-up question refers back to the original text, it cannot plan a text to clarify a part of the text, as it no longer knows what were the intentions behind various pieces of the text.</Paragraph>
    <Paragraph position="5"> Consider again the dialogue in Figure 4.</Paragraph>
    <Paragraph position="6"> When the user asks 'What is a generalized variable?' (utterance (4) in Figure 4), the query analyzer interprets this question and posts the goal: (BMB S H (KNOW H generalized-variable) ). At this point, the explainer must recognize that this discourse goal was attempted and not achieved by the  last sentence of the previous explanation. 6 Failure to do so would lead to simply repeating the description of a generalized variable that the user did not understand. By examining the text plan of the previous explanation recorded in the dialogue history, the explainer is able to determine whether the current goal (resulting from the follow-up question) is a goal that was attempted and failed, as it is in this case. This time, when attempting to achieve the goal, the planner must select an alternative strategy. Moore (1989b) has devised recovery heuristics for selecting an alternative strategy when responding to such follow-up questions. Providing an alternative explanation would not be possible without the explicit representation of the intentional structure of the generated text. Note that it is important to record the rhetorical structure as well, so that the text planner can choose an alternative rhetorical strategy for achieving the goal. In the example under consideration, the recovery heuristics indicate that the rhetorical strategy of giving examples should be chosen.</Paragraph>
  </Section>
  <Section position="10" start_page="209" end_page="209" type="metho">
    <SectionTitle>
STATUS AND FUTURE WORK
</SectionTitle>
    <Paragraph position="0"> The text planner presented is imple.mented in Common Lisp and can produce the text plans necessary, to participate in the sample ~lialogue described m this paper and several others (see (Moore, 1989a, Paris, 1988a)). We currently have over 60 plan operators and the system can answer tlie following types of  (follow-up) questions: - Why? - Why conclusion? - Why are you trying to achieve goal? - Why are you using method to achieve goal? Why are you doing act? How do you achieve goal? - How did you achieve goal (in this case)? - What is a concept? - What is the difference between concept1 and concept2? - Huh?  The text planning system described in this paper is being incorporated into two expert systems currently under development. These systems will be installed and used in the field. This will give us an opportunity to evaluate the techniques proposed here.</Paragraph>
    <Paragraph position="1"> We are currently studying how the attentional structure inherent in our text plans can be used to guide the realization process, for example in the planning of referring expressions and the use of cue phrases and pronouns. We are also investigating criteria for the expansion and ordering of optional satellites in our plan operators. Currently we use information from the user model to dictate whether or not optional satellites are expanded, and their ordering is specified in each plan operator. We wish to extend our criteria for satellite expansion to include other factors such as pragmatic and stylistic goals (Hovy, 1988a) (e.g., brevity) and the conversation that has occurred so far. We are also investigating the use of attentional information to control the ordering of these satellites (McKeown, 1985).</Paragraph>
    <Paragraph position="2"> We also believe that the detailed text plan constructed by our planner will allow a system to modify its strategies based on experience (feedback from the user). In (Paris, 1988a), we outline our preliminary ideas on this issue.</Paragraph>
    <Paragraph position="3"> We have also begun to study how our planner can be used to handle incremental generation of texts. In (Moore, 1988), we argue that the detailed representation provided by our text plans is necessary for execution monitoring and to indicate points in the planning process where feedback from the user may be helpful in incremental text planning.</Paragraph>
  </Section>
class="xml-element"></Paper>