<?xml version="1.0" standalone="yes"?>
<Paper uid="W90-0121">
  <Title>Rule Similarity Specialization I SpecialiT~fion Similarity Specialization Specialization Deduction the Contradiction in the Sample Network Inference \[ Possible Impairment \[DL apply-to Numbers has-goal BrE\]\] \[DL apply-to LT has-goal BrE\] ~ Insufficient Learning</Title>
  <Section position="4" start_page="157" end_page="162" type="ackno">
    <SectionTitle>
AND
</SectionTitle>
    <Paragraph position="0"> B a context \[PROCa--OBJm--GOALt \] with links \[PROCo apply-to OBJm\] (MB = kam) and \[PROC= has-goal GOALt\] (MB = ktam) } THEN (with certainty P2 ) Add a context \[PROCa---OBJ,--GOALt\] with links \[PROCa apply-to OBJ,\] of type I</Paragraph>
    <Paragraph position="2"/>
    <Section position="1" start_page="157" end_page="157" type="sub_section">
      <SectionTitle>
Common-sense Inference Rule
Generating Peripheral Rhetorical Devices
</SectionTitle>
      <Paragraph position="0"> The generation of Peripheral RDs is performed by applying a procedure called Impairment-Invalidate. This procedure simulates the effect of an IM on our network model to anticipate possible impairments which may inhibit a listener's acquisition of this IM, and then proposes Peripheral RDs, such as Contradictions, Revisions and Motivations, to invalidate these impairments. To this effect, it activates three mechanisms: (1) Propagation of the effect of a message, (2) Recognition of impairments, and (3) Selection of Peripheral RDs. In the following subsections, we describe these mechanisms and then discuss procedure Impairment-Invalidate.</Paragraph>
      <Paragraph position="1"> Propagation of a Message Propagation simulates the alterations taking effect in a network representing a listener's beliefs due to inferences drawn from a message. These inferences, which are drawn by activating applicable Common-sense Inference Rules, result in changes in the MBs of existing links or addition of new links and nodes to the network.</Paragraph>
      <Paragraph position="2"> For instance, a Contradiction such as &amp;quot;You cannot always simplify inside the brackets of algebraic terms,&amp;quot; presented to weaken a listener's belief in the applicability of bracket simplification to Algebraic Terms, may cause the listener to weaken his/her correct belief in the applicability of bracket simplification to Like Algebraic Terms or even to Numbers, and also to weaken his/her incorrect belief in the applicability of bracket simplification to Unlike Algebraic Terms and in the applicability of addition and subtraction to Algebraic Terms. This effect is simulated by the propagation of the Contradiction \[BrS ~apply-to AT\] in the sample network in Figure 2, which produces the inferences \[Br$ -,apply-to L'I'\] and \[BrS ~apply-to UT\] through the application of a specialization rule, \[BrS ~apply-to Numbers\] through the application of the similarity-based rule R2 (see Figure 3), and \[+/- ~apply-to AT\] through the application of a deductive inference rule. These inferences contradict the beliefs represented by existing links, thereby lowering the MBs associated with these links. However, their effect ultimately depends on the strength of the rules in question and on the MBs of the links, i.e., the impact of an inference on a weakly believed link is more pronounced than for a strongly believed one.</Paragraph>
    </Section>
    <Section position="2" start_page="157" end_page="161" type="sub_section">
      <SectionTitle>
Recognition of Impairments
</SectionTitle>
      <Paragraph position="0"> The Recognition mechanism anticipates possible impairmerits to a listener's comprehension process by examining changes in a network representing the listener's beliefs due to a message or an inference. It does not guarantee that a certain impairment will occur, rather, it conjectures that an impairment is likely to affect a partitular link, denoted a culprit link.</Paragraph>
      <Paragraph position="1"> We have characterized in terms of our network model two main types of impairments which lead to undesirable effects commonly encountered in a knowledge acquisition setting: (1) Affect-related impairments which elicit negative affective responses such as Confusion and Loss of Interest, and are caused by a conflict between a message or inference and a belief held by a listener; and (2) Belief-related impairments such as Mislearning, Insufficient Learning or Insignificant Change in a listener's knowledge status, which are caused by a discrepancy between a listener's belief in a proposition (possibly as a result of an inference) and the belief the speaker intends him/her to have in this proposition. Let us first consider the Affect-related impairments.</Paragraph>
      <Paragraph position="2"> Confusion occurs when an inference decreases significandy a listener's confidence in a previous belief, causing a discomforting transition from a self-perception of possessing knowledge to one of increased uncertainty. In terms of our network model, Confusion takes place when the absolute value of an Anticipated Measure of Belief (AMB), obtained by combining the MB of a link with the MB of an inference regarding this link, is significantly lower than the absolute value of the original MB of this link. For instance, the statement &amp;quot;One cannot always add algebraic terms,&amp;quot; which yields a negative value for the MB of the link \[+/- apply-to AI'\] in Figure 2, may trigger the erroneous inference \[+/-apply-to LT\], which contradicts the link \[+/- apply-to LT\], thereby lowering its MB. The Strength of this impairment for a link L is defined as follows: Strength (Confusion ,L ) = max/( I MB (L ) I -lAMB (L ) I ),0} Loss o/Interest occurs when a listener who is initially motivated to acquire knowledge is presented with an IM s/he considers redundant. In terms of our model, this takes place if there exists a node B which subsumes a new node A, i.e., new distinguishing links incident upon A are connected to the same nodes and have MBs of compatible magnitude and sign as the corresponding  links incident upon B. For this type of impairment, a culprit link is an erroneous link incident upon B representing belief, whose reversal into disbelief results in B no longer subsuming A. This situation is illustrated in Figure 2, where we try to add the node DL, representing distributive law, and the links \[DL apply-to AT\] and \[DL has-goal BrE\] to the network representing a listener's beliefs. However, the existence of the culprit link \[BrS apply-to AT\] supports the erroneous belief that bracket simplificauon is equivalent to distributive law, thereby rendering the new procedure redundant. If all the existing links participating in an impainnent due to Loss of Interest are correct, no culprit link is identified. A characterization of Belief-related impairments must take into consideration the difference between a listener's level of expertise and a level of expertise considered satisfactory. To this effect we define the Strength of an impairment I in link L as follows:</Paragraph>
      <Paragraph position="4"> where SMB is a Satisfactory Measure of Belief representative of an adequate level of expertise with respect to a link. Its value may be obtained from a network which represents the speaker's beliefs. The relative position of Belief-related impairments and Confusion in a listener's belief space is graphically represented in Figure 4 which depicts these impairments as a function of the AMB and the previous MB of a link with SMB &gt;0. The diagram for a link with SMB &lt;0 is syrnmelric to the one in Figure 4.</Paragraph>
      <Paragraph position="5"> Mislearning takes place when an erroneous belief with a relatively high degree of certainty is produced by an incorrect inference drawn by a listener. In terms of our network model, this takes place when the AMB of a link represents a substantial incorrect belief, and if this link existed previously, the strength of the impairment has increased, i.e., the AMB of this link is farther than its previous MB from its SMB. If the absolute value of the AMB of the link in question is higher than the absolute value of its previous MB, the listener will falsely perceive him/herself as being more proficient.</Paragraph>
      <Paragraph position="6"> Insufficient Learning occurs when a correct inference yields a correct belief with a relatively high degree of certainty, but which still falls short of a desired degree of certainty representative of proficiency. In terms of our network model, this occurs when the AMB of a link represents a substantial correct belief, and ff this link existed previously, the strength of the impairment has decreased, i.e., the AMB of this link is closer than its previous MB to its SMB.</Paragraph>
      <Paragraph position="7"> Finally, an Insignificant Change in a listener's knowledge status occurs when an inference accomptishes a rather inconsequential change with respect to a previously non-existent link or with respect to a link  a Link with SMB&gt;0 with an MB representative of insufficient proficiency.</Paragraph>
      <Paragraph position="8"> This MB may represent either a correct belief or an incorrect belief.</Paragraph>
      <Paragraph position="9"> The immediate invalidation of Affect-related impairments is essential for the smooth continuation of the knowledge acquisition process, since their persistence diverts a listener's mental resources from the task of acquiring knowledge. On the other hand, the invalidation of Belief-related impairments with respect to links which are not currently in focus may be temporarily postponed due to didactic or stylistic considerations. Therefore, although Confusion may take place in conjunction with a Belief-related impairment, the recognition and subsequent invalidation of Confusion takes precedence over the detection of this impairment (see Figure 4).</Paragraph>
      <Paragraph position="10"> Selection of Rhetorical Devices The Selection mechanism proposes a Peripheral RD to address a recognized impairment. The type of this rhetorical device, its wording, and its position in the final message sequence depend on the type of the impairment, the correctness and magnitude of the previous MB of the culprit link, and the SMB of this link (see Table I). The strength of an impairment affects the need for additional explanations, such as Causal explanations and Instantiations, to convey a Peripheral RD.</Paragraph>
      <Paragraph position="11">  tion pertaining to the previous MB and the SMB of the culprit link. This information is unnecessary when addressing Mislearning and Insufficient Learning, since these impairments are completely characterized by their type. However, for the rest of the impairments, the correctness of the previous MB is the main factor in the determination of the type of a Peripheral RD. Loss of Interest is invalidated by a Contradiction of a culprit link, if such a link is identified; otherwise, a Motivation has to generated by adding nodes and links which render the new information non-redundant, e.g., &amp;quot;You already know how to solve quadratic equations by completion to square. A faster method is ... &amp;quot; (demil~ regarding the types of Motivations which are suitable for different situations and users appear in \[Zukerman 1987\]). The invalidation of an Insignificant Change in a link with an MB representative of lack of expertise, i.e., an MB whose absolute value is close to 0, is performed in a manner similar to the correction of complete ignorance.</Paragraph>
      <Paragraph position="12"> That is, if the SMB of this link is positive, this impairment is invalidated by a Revision of the relevant informarion, whereas ff it is negative, i.e., representative of disbelief, no Peripheral RD is proposed, since it may be superfluous to induce disbelief with respect to a proposition which is hardly entertained by a listener.</Paragraph>
      <Paragraph position="13"> As seen in Table 1, we distinguish between two types of Peripheral RDs according to the source of the belief being addressed, namely Peripheral RDs which address previously existing links and Peripheral RDs which address current inferences. This distinction is not always reflected in the English realization of a rhetorical device, rather, it may affect the Meta Comments which accompany it and its position in the final message sequence. For instance, the propagation of the Contradiction \[BrS ~apply-to AT\] may yield the erroneous inference \[BrS --,aBly-to LT\] which in turn may cause an impairment in the link \[BrS apply-to LT\]. If the anticipated impairment is Confusion, it will be invalidated by a Revision of this link, whereas if it is Mislearning, it will be invalidated by a Contradiction of the erroneous hfference. The most succinct realization of both Peripheral RDs is the sentence &amp;quot;You can always simplify bracketed Like Terms.&amp;quot; However, the Revision of the previous belief may appear either before or after the above Contradiction and include a Meta Comment which states the source of this belief, e.g., &amp;quot;As we saw in Section 7, you can always simplify bracketed Like Terms&amp;quot;; while the Contradiction of the erroneous inference would usually appear after the above Contradiction and would be accompanied by a Meta Comment which indicates a violation of an expectation established by the inference, e.g., &amp;quot;Bracket simplification does not always apply to algebraic terms, but it always applies to Like Terms.&amp;quot; The Peripheral RDs proposed by our Selection mechanism contain sufficient information to support the generation of these types of Meta Comments by means of mechanisms such as the ones presented in \[Zukerman and Pearl 1986\] and \[Zukerman 1989\].</Paragraph>
      <Paragraph position="14">  Impairment-Invalidate is activated with one argument which contains an IM. It returns a Message-List composed of the IM and the Peripheral RDs which were proposed to invalidate the impairments anticipated as a result of this message.</Paragraph>
      <Paragraph position="15">  We distinguish between two main stages of this procedure: (1) The preliminary stage (lines 1-4) which determines the need for a Peripheral RD related to the input message, and (2) The iterative stage (lines 5-17) which ascertains the need for Peripheral RDs pertaining to subsequent inferences.</Paragraph>
      <Paragraph position="16"> In the preliminary stage, procedure Recognize is applied in order to determine whether the IM is likely to cause an impairment. If this is the case, Recognition returns a triple (T,S,L), where T contains the Type of the impairment, S its Strength, and L indicates whether the belief represented by the link in question is correct or incorrect. Otherwise, it returns n/l and no impairment is predicted. If an impairment was anticipated, procedure Select is activated to propose a Peripheral RD. In our example, the Selection mechanism invalidates the impairment responsible for the Loss of Interest by means of the Contradiction \[BrS -,apply-to AT\] which induces disbelief in the link \[BrS apply-to AT\]. Upon completion of the preliminary stage, the proposed Peripheral RD together with the IM form the Message-List, which constitutes the input to the next phase of the impairment invalidation procedure.</Paragraph>
      <Paragraph position="17"> Both the IM and the proposed Peripheral RD cause modifications in a listener's degree of belief in the addressed links. These modifications in turn may lead to changes in his/her beliefs in other links. Hence, in order to prevent impairments due to inferences drawn from these messages, additional Peripheral RDs may be called for. Furthermore, Peripheral RDs may be required in order to invalidate an Insignificant Change in links related to the IM which have MBs representative of insufficient proficiency, i.e., to attain a satisfactory degree of expertise with respect to these links. To determine the need for additional Peripheral RDs, Propagate produces inferences from each message in Message-List.</Paragraph>
      <Paragraph position="18"> During this process, inferences from different messages which affect the same link are merged into one inference with a combined effect (lines 5-7). Recognition then ascertains the attributes of the impairments which are likely to be caused by these inferences (line 8). For instance, the Propagation of the input message \[DL apply-to AT has-goal BrE\] and the Contradiction \[BrS -,apply-to AT\] in the sample network in Figure 2 may yield the inferences in Table 2. In principle, each of these inferences may be responsible for an impairment.</Paragraph>
      <Paragraph position="19"> However, as stated above, the effect of these inferences ultimately depends on the p-s of the rules applied in the propagation process and on the MBs of the affected links. In our present discussion, we assume that Confusion was recognized in the links \[BrS apply-to LT\] and \[BrS apply-to Numbers\].</Paragraph>
      <Paragraph position="20"> Based on Gricean maxims of cooperative conversation (Grice 1975), we propose to generate a minimal number of rhetorical devices to invalidate impairments occurring concurrently in a number of links. To this end, we iterate over the set of impairments, selecting at each stage the culprit link with the highest ranking impairment (line 10). The impairments are ranked according to their type and strength, where impairments causing Confusion are ranked higher than Belief-related impairments. A Peripheral RD is then proposed to invalidate the impairment in the selected link, and its effect is propagated (lines 12-14). Once a Peripheral RD has been generated to address a particular link, further inferences drawn during the same activation of Impairment-Invalidate no longer affect this link. For each iterauon, the set of impairments is updated by merging the inferenccs responsible for the previously recognized impairments with the inferences resulting from the most recent propagation, and reapplying the Recognition process 0ines 15-16). In this manner, a low-ranking impairment in a given link may be spontaneously invalidated by an inference resulting from a Peripheral RD generated to invalidate an impairment in another link. For example, ff the impairment in the link \[BrS apply-to Numbers\] ranks higher than the impairment in the link \[BrS apply-to LT\], a Revision is generated to invalidate the impairment in the former link. The propagation of this Revision may invalidate the Confusion with respect to the latter link. If this Revision does not cause further impairments, the procedure terminates returning the fol- null The generation of belief Revisions tends to inhibit further impairments, since they reinforce beliefs which are likely to be consistent with the rest of a listener's beliefs, while the generation of belief Contradictions tends to foster impairments, since they disagree with  existing beliefs. However, in a knowledge acquisition setting, misconceptions are generally not allowed to pile up, hence Contradictions to existing beliefs are expected only during the initial stages of the propagation process. Therefore, although in principle our algorithm may iterate indefinitely, in practice, impairments should no longer be detected after a few iteraations, and the process should halt after proposing a few Peripheral RDs. This expectation is confirmed by tests run with our sample network (see next section). However, this network is rather small, and there may be situations in which the number of Peripheral RDs proposed by our mechanism will have to be restricted due to stylistic or pedagogical considerations. In such cases, our mechanism will have to be adjusted to invalidate only a subset of the recognized impairments such as the highest ranking impairments or impairments in links which are closest to the  in a system called WISHFUL, which was run on several instances of the network in Figure 2. These instances featured MBs which represented three types of students:  (1) Competent -- with high MBs associated with correct  with rhetorical devices people would generate under similar circumstances, and the response time was instantaneous for these rather small networks. As expected, changes in the p-s of the Common-sense Inference Rules resulted in variations in the generated RDs, with additional Contradictions due to Misleaming being generated as the p-s of unsound rules increased, and additional Revisions due to Insufficient Learning as the p-s of sound rules decreased. In addition, the number and strength of the recognized impairments increased as the ability of the students being represented in a network decreased, iadicating that additional explanations would be required to convey a message to the poorer students.</Paragraph>
      <Paragraph position="21"> In particular, for the networks representing competent students, either Revisions of inferences due to Insufficient Learning or no Peripheral RDs were proposed (depending on the p-s of the Common-sense Inference Rules); for the networks representing average students, Contradictions and Revisions were proposed in a manner similar to the explanations in this paper, and for the networks representing mediocre students most of the proposed Peripheral RDs were Revisions of information.</Paragraph>
      <Paragraph position="22"> Our mechanism has also proven successful as an analytical tool. It accounts for the presence of Peripheral RDs in over twenty texts in a variety of domains, ranging from expert domains (Telecommunications, Cognitive Science and Linguistics) through intermediate ones (high-school Algebra, Data Structures, Lisp and introductory Chess) to novice domains (Childcraft Encyclopedia and Dr. Spock's Baby and Child Care).</Paragraph>
    </Section>
    <Section position="3" start_page="161" end_page="162" type="sub_section">
      <SectionTitle>
Limitations and Future Research
</SectionTitle>
      <Paragraph position="0"> Our mechanism proposes rhetorical devices under the assumption that after it has done &amp;quot;its best,&amp;quot; the listener's beliefs addressed by the discourse will be modified in the desired direction, although not necessarily to the desired extent. This is a valid assumption for discourse generation, since one can not say more than one knows. However, for this mechanism to generate effective discourse on a continued basis, the model of the fistener's beliefs must be updated by an independent assessment of the his/her understanding.</Paragraph>
      <Paragraph position="1"> In addition, our mechanism must be implemented on different domains and larger networks to test both its response time and the adequacy and number of the proposed rhetorical devices. As stated above, this may reveal the need to adjust procedure Impairment-Invalidate to enable it to control the number of the proposed Peripheral RDs.</Paragraph>
      <Paragraph position="2"> At present, research is in progress to extend the impairment invalidation paradigm to the generation of other types of rhetorical devices, such as Descriptions, Instantiations and Causal explanations, and to devise an algorithm to sort the generated messages according to rhetorical considerations (Zukerman 1990b). In addition, an alternative representation for MBs which keeps track of the sources of an inference is being considered.</Paragraph>
      <Paragraph position="3"> Further research is required to recognize and rectify possible misconceptions in the Common-sense Inference Rules, and to characterize conditions for the generation of rhetorical devices which satisfy a number of communicative goals. Finally, the effect of the knowledge representation and the rules of inference on the types of the proposed rhetorical devices needs to be further investigated.</Paragraph>
      <Paragraph position="4"> Conclusion This paper offers a text planning mechanism which supports the generation of explanations tailored to particular types of users. Our mechanism generates Peripheral RDs which help convey an intended message by anticipating and preventing potential impairments to a listener's comprehension process. To this effect, it characterizes these impairments in terms of a model of a fistener's beliefs and inferences, and simulates a fistener's comprehension process on this model. Clearly, this protess does not dispense with the need to interact with a listener, but it addresses commonly occurring impairments, thereby focusing the interaction. Furthermore, it is envisioned that the impairment invalidation mechanism will become a useful tool to guide the generation of cogent responses to user follow-up queries, as it can point to issues which are potentially troublesome.</Paragraph>
    </Section>
  </Section>
class="xml-element"></Paper>