<?xml version="1.0" standalone="yes"?>
<Paper uid="J99-3004">
  <Title>Interpreting and Generating Indirect Answers</Title>
  <Section position="5" start_page="398" end_page="410" type="metho">
    <SectionTitle>
4. Interpretation
</SectionTitle>
    <Paragraph position="0"> This section describes the interpretation process. In our model, implicated answers are derived by an answer recognizer. Algorithms for the answer recognizer are described in Section 4.1. Of course, dialogue consists of more than questions and answers.</Paragraph>
    <Paragraph position="1"> Section 4.2 describes the role of the answer recognizer in a discourse-processing architecture. Finally, Section 4.3 discusses how this model relates to previous models of conversational implicature.</Paragraph>
    <Section position="1" start_page="398" end_page="405" type="sub_section">
      <SectionTitle>
4.1 Answer Recognizer
</SectionTitle>
      <Paragraph position="0"> The inputs to the answer recognizer include: * the set of discourse plan operators and coherence rules described in Section 3,  answer to the questioned proposition p), * the semantic representation of p, and * for each utterance performed by R during R's turn, the type of communicative act signaled by its form (e.g., to inform), and the semantic representation of its content. 25 Answer recognition is performed in two phases. The goal of the first phase is to derive a set of candidate discourse plans plausibly underlying R's response. The first phase makes use of two subcomponents: one that we refer to as the hypothesis generation component, and a theorem prover. The output of the first phase of answer recognition is a set of candidate discourse plans since there may be alternate interpretations of R's response. The goal of the second phase of answer recognition is to evaluate the relative plausibility of each candidate discourse plan. The final output of answer recognition consists of a partially ordered set of the candidates ranked by plausibility. Plan recognition is primarily top-down, i.e., expectation-driven. More specifically, Q26 attempts to interpret the response as having been generated from a discourse plan constructed from the discourse plan operators for full answers. The problem of reconstructing R's discourse plan has several aspects (to be described in more detail shortly):  * Instantiating discourse plan operators with the questioned proposition and appropriate propositions from R's response.</Paragraph>
      <Paragraph position="1"> 25 The turn in question need not be the turn immediately following Q's asking of the question, as discussed in Section 4.2. Also, we make the simplifying assumption that R's answer is given within a single turn.</Paragraph>
      <Paragraph position="2"> 26 For convenience, we refer to the answer recognizer component as Q, and to the answer generator as R. 400 Green and Carberry Indirect Answers * Consistency checking: determining whether the beliefs and goals that would be attributed to R by virtue of ascribing a particular discourse plan to R are consistent with Q's beliefs about R's beliefs and goals. * Coherence evaluation: determining whether a putative satellite of a  candidate plan is plausibly coherent, i.e., given a candidate plan's (or subplan's) nucleus proposition p, putative satellite proposition q, and the putative satellite's coherence relation CR, determining whether Q believes that (CR q p) is mutually plausible. Coherence evaluation makes use of the coherence rules described in Section 3.1.</Paragraph>
      <Paragraph position="3"> * Hypothesis generation: hypothesizing any &amp;quot;missing parts&amp;quot; of the response that are required in order to assimilate acts in R's response into a coherent candidate plan. Hypothesis generation also makes use of the coherence rules.</Paragraph>
      <Paragraph position="4"> Initially, the header variable of each &amp;quot;top-level&amp;quot; answer discourse plan operator 27 is instantiated with the questioned proposition p, i.e., all occurrences of the header variable are replaced with p. Next, consistency checking is performed to eliminate any candidates whose applicability conditions or primary goals are not consistent with Q's beliefs about R's beliefs and goals. For all remaining candidates, the answer recognizer next attempts to recognize an act from R's turn as the nucleus of the plan, i.e., to check whether R gave a direct answer. If no acts in R's turn match the nucleus, then the nucleus is marked as hypothesized. For all remaining acts in R's turn, the answer recognizer attempts to recognize all possible satellites, as specified in each remaining candidate plan. In the model the discourse plan operators do not specify a required ordering of satellites. 2s The subprocedure of satellite recognition is described in more detail in Section 4.1.2.</Paragraph>
      <Paragraph position="5">  ing an instance of a satellite of a candidate plan. The inputs consist of: * sat-op, a discourse plan operator for a possible satellite, * the proposition p conveyed by the nucleus of the higher-level plan (i.e., the plan whose satellites are currently being recognized), * act-list, a list of acts in R's turn that have not yet been assimilated into the candidate plan, * cur-act, the current act (inform s h q) in act-list, where s is the speaker, h is the hearer, and q is the propositional content of the act.</Paragraph>
      <Paragraph position="6"> The output is a set (possibly empty) of candidate instances of sat-op. To give a simplified, preliminary version of the algorithm, first, the header variable and existential variable of sat-op are instantiated with p and q, respectively. Then, coherence evaluation and consistency checking are performed. If successful, cur-act is recognized to 27 Five of these are defined in our model, corresponding to the five types of answers illustrated in Table 1. 28 The operators do specify a preferred order, however, which is used in generation. Also, our process model includes a structural constraint on satellite ordering. During interpretation, only instances satisfying this constraint are considered. That is, the constraint eliminates interpretations which, in our judgment, are not plausible due to incoherence. For a description of the constraint, see Green (1994). We expect that other such constraints may be incorporated into the process model.  Candidate discourse plan with hypotheses.</Paragraph>
      <Paragraph position="7"> be the nucleus of sat-op, and for each remaining act in act-list, satellite recognition is performed for each satellite of sat-op.</Paragraph>
      <Paragraph position="8"> However, the satellite recognition algorithm as described so far would not be able to handle R's response in (11), since there is no plausible coherence relation in the model directly relating (11)iv to (11)ii (or to any other direct answer that could be recognized in the model).</Paragraph>
      <Paragraph position="9"> (11) i. Q: Are you driving to campus tonight? ii. R: \[No.\] iii. \[My car's not running.\] iv. My car has a broken timing belt.</Paragraph>
      <Paragraph position="10"> Whenever the answer recognizer is unable to recognize cur-act as the nucleus of sat-op, a subprocedure we refer to as hypothesis generation is invoked. Hypothesis generation will be described in detail in the following section. It returns a set of alternative hypothesized propositions, each of which represents the content of a possible implicit inform act to be inserted at the current point of expanding the candidate plan. 29 In this example, the proposition conveyed in (11)iii would be returned as a hypothesized proposition, which is used to instantiate the existential variable of a Use-obstacle satellite, thereby enabling satellite recognition to proceed. Then, (11)iv can be recognized (without hypothesis generation being required) as a satellite of (11)iii. Ultimately, the plan shown in Figure 5 would be inferred. (Only the hierarchical structure and communicative acts are shown. By convention, the left-most child of a node is the nucleus and its siblings are the satellites. Labels of sentences in (11) that could realize a leaf node are used to label the node. Hypothesized nodes are indicated by square brackets.) The complete satellite recognition algorithm, employing hypothesis generation, is given in Figure 6.</Paragraph>
      <Paragraph position="11"> 29 Thus, hypothesis generation may provide additional inferences, i.e., more than just the implicated answer. Hinkelman (1989) refers to such implicatures, licensed by attributing a plan to an agent, as plan-based implicatures.</Paragraph>
      <Paragraph position="12">  p: proposition from nucleus of higher-level plan cur-act: current act, (inform s h q), to be recognized act-list: list of remaining acts in R's turn op: discourse plan operator (Use-CR s h ?p) sat-cand-set: set of candidate instances of op underlying part of R's response 1. Instantiate header variable ?p of op with p.</Paragraph>
      <Paragraph position="13"> 2. Instantiate existential variable ?q of op with q of cur-act.</Paragraph>
      <Paragraph position="14"> a. Prove that it is plausible that q and p are related by CR.</Paragraph>
      <Paragraph position="15"> If not, go to step 2c.</Paragraph>
      <Paragraph position="16"> b. Check consistency. If not consistent, then go to step 2c; else go to step 3a.</Paragraph>
      <Paragraph position="17"> c. Try substituting each q returned by hypothesis generation for ?q: Check consistency and coherence as in steps 2a and 2b.</Paragraph>
      <Paragraph position="18"> For each q passing both checks, proceed with step 3b.</Paragraph>
      <Paragraph position="19"> If none pass, then fail.</Paragraph>
      <Paragraph position="20"> 3. a. Mark cur-act as used. Go to step 4.</Paragraph>
      <Paragraph position="21"> b. Mark nucleus as hypothesized.</Paragraph>
      <Paragraph position="22"> 4. For each unused act in act-list, attempt to recognize each satellite of op.</Paragraph>
      <Paragraph position="23">  Figure 6 Satellite recognition algorithm.</Paragraph>
      <Paragraph position="24"> 4.1.3 Hypothesis Generation. Based upon the assumption that the response is coherent, the goal of hypothesis generation is to fill in missing parts of a candidate plan in such a way that an utterance in R's turn can be recognized as part of the plan. The use of hypothesis generation broadens the coverage of our model to cases where more is missing from a full answer than just the nucleus of a top-level operator. (From the point of view of generation, it enables the construction of a more concise, though no less informative, response.) The hypothesis generation algorithm constructs chains of mutually plausible propositions, each beginning with the proposition (e.g., the proposition conveyed in (11)iv) to be related to a goal proposition in a candidate plan (e.g., the proposition conveyed in (11)ii), and ending with the goal proposition, where each pair of adjacent propositions in the chain is linked by a plausible coherence relation. The algorithm returns the proposition (e.g., the proposition conveyed in (11)iii) immediately preceding the goal proposition in each chain. Thus, when top-down recognition has reached an impasse, hypothesis generation (a type of bottom-up data-driven reasoning) provides a hypothesis that enables top-down recognition to continue another level of growth. An example of hypothesis generation is given in Section 4.1.5.</Paragraph>
      <Paragraph position="25"> The algorithm for hypothesis generation, which is given in Figure 7, performs a breadth-first search subject to a processing constraint on the maximum depth of the search tree. Note that a chain may have a length greater than three, e.g., the chain may consist of propositions (p0, pl, P2, P3), where P0 is the proposition to be related to the candidate plan, p3 is the goal, and P2 would be returned as a hypothesized proposition. In such a case, after p2 has been assimilated into the candidate plan, if pl is not present in R's turn, then hypothesis generation is invoked again and pl would  pg: goal proposition GCR: goal coherence relation, i.e., coherence relation that must hold between hypothesized proposition and Pa S: set of coherence relations N: maximum search depth hypoth-list: list of alternative hypothesized propositions  1. Initialize root of search tree with p0.</Paragraph>
      <Paragraph position="26"> 2. Expand nodes of tree ill breadth-first order until either no more expansion is possible or maximum tree depth of N is reached, whichever happens first. To expand a node pi: a. Find all nodes pi+l such that for some relation CR in S, (Plausible (CR Pi pi+l)) is provable.</Paragraph>
      <Paragraph position="27"> b. Make each such Pi+l a child of Pi, linked by CR.</Paragraph>
      <Paragraph position="28"> c. A goal state is reached whenever Pi+l is pg and CR is identical to GCR.</Paragraph>
      <Paragraph position="29"> d. Whenever a goal state is reached, add the parent of pg to hypoth-list.</Paragraph>
      <Paragraph position="30">  be hypothesized also. 3deg Finally, the search for a proposition Pi+l in step 2a is performed in our implementation using a theorem prover.</Paragraph>
      <Paragraph position="31">  bility of the set of candidate plans output by the first phase of answer recognition. First, plausibility decreases as the number of hypotheses in a candidate increases. (Assuming that all else is equal, it is safer to favor interpretations requiring fewer hypotheses.) Second, plausibility increases as the number of utterances in R's turn that are accounted for by the plan increases. (The more of R's turn accounted for, the more coherent the turn is likely to be, although not all of the utterances in R's turn are necessarily part of the full answer.) To give an example, consider the two candidate plans shown in Figure 8, corresponding to alternative interpretations of R's response in (12). 31  (12) i. Q: Are you going to campus tonight? ii. R: \[No/Yes\] iii. \[My car's not running.\] iv. My car has a broken timing belt, v. \[so\] I'm going to take the bus.</Paragraph>
      <Paragraph position="32"> vi. Do you know how much the fare is? 30 The implementation saves the chains to avoid the expense of recomputing intermediate hypotheses. 31 This constructed example was designed to illustrate multiple aspects of the model. In our judgement, normally it would sound more coherent to give (12)v before (12)iv if (12)vi were not included.  However, when (12)vi is included, (12)v not only elaborates upon the yes, but also serves as background for (12)vi. Another possible motivation for giving (12)v after (12)iv might be to delay giving dispreferred information (Levinson 1983), e.g., if the speaker believed that a yes was an unexpected or unwanted answer to (12)i.</Paragraph>
      <Paragraph position="33">  Ranking candidate plans.</Paragraph>
      <Paragraph position="34"> By these heuristics, a yes answer would be the preferred interpretation, since the candidate Answer-yes plan uses the same number of hypotheses as the candidate Answer-no plan, and accounts for more of R's response. ((12)vi is not recognized as part of either answer.) The preference heuristics are intended to capture local coherence only. Since global information may play a role in selecting the correct interpretation, the higher-level discourse processor (described in section 4.2) must decide which plan to attribute to the speaker.</Paragraph>
      <Paragraph position="35">  indirect answers in the model by describing how the two candidate plans shown in Figure 8 would be derived from R's response of (12)iv through (12)vi. First, each of the five top-level answer discourse plan operators would be instantiated with the questioned proposition p, the proposition that R is going to campus tonight. Assuming that Q has no beliefs about R's beliefs and goals that are inconsistent with the applicability conditions and primary goals of these candidates, none of the candidates would be eliminated yet. Second, for each candidate the recognizer would check whether the communicative act specified in the nucleus was present in R's turn. In this example, since a direct answer was not explicitly provided by R, the recognizer would mark the nucleus of each candidate as hypothesized. The hypothesized nucleus of the candidate Answer-no and Answer-yes plans would be (inform s h (not p)) and (inform s h p), respectively. Next, the recognizer would try to recognize the acts expressed as (12)iv through (12)vi as satellites of each candidate plan. Assume that these acts are represented as (inform s h piv), (inform s h pv), and (inform s h pvi), respectively. To recognize an instance of a satellite, first, a satellite discourse plan operator would be instantiated. The header variable would be instantiated by unifying the satellite plan header with the corresponding act in the higher-level plan. For example, the header variable of a Use-obstacle satellite of an Answer-no candidate would be instantiated with (not p) in this example. The existential variable would be instantiated with the proposition conveyed in some utterance to be recognized as a satellite, e.g., Ply. However, before a candidate satellite may be attached to the higher-level candidate plan, the answer recognizer must verify that the candidate satellite passes the following two tests: First, the candidate satellite's applicability conditions and primary goals must be consistent with Q's beliefs about R's beliefs and goals. Second, the specified coherence relation must be plausible with respect to the beliefs that Q presumes to be shared with R, i.e., the satellite's instantiated applicability condition  Computational Linguistics Volume 25, Number 3 of the form (Plausible (CR q p)) must be provable using the coherence rules described in Section 3. For example, given the beliefs that Q presumes to be shared with R and the coherence rules provided in the model, the act underlying (12)iv could not be the nucleus of a candidate Use-obstacle satellite of the Answer-no candidate, because the recognizer would not be able to prove that cr-obstacle is a plausible coherence relation holding between Piv and (not p). On the other hand, the act underlying (12)v would be interpreted as the nucleus of a candidate Use-elaboration satellite of the Answer-yes candidate, since the above tests are satisified, e.g., the recognizer could prove that cr-elaboration is a plausible coherence relation holding between pv and p. To return to consideration of the recognition of the Answer-no candidate, upon finding that the act underlying (12)iv cannot serve as a satellite, hypothesis generation would be attempted. Recall that the goal of hypothesis generation is to supply a hypothesized missing act of the plan so that top-down recognition can continue. Hypothesis generation would search for a chain of plausibly related propositions, beginning with the proposition (i.e., Piv) to be related to the candidate Answer-no plan, and ending with the goal proposition (i.e., (not p)). As mentioned in Section 4.1.3, each pair of adjacent propositions in the chain must be linked by a plausible coherence relation. In this example, hypothesis generation would construct the chain (piv, piii, (not p)), where both pairs of adjacent propositions would be related by cr-obstacle and Piii is the hypothesis that R's car is not running. Hypothesis generation would return the proposition immediately preceding the goal proposition in this chain, i.e., pill. Thus, piii would be used to instantiate the existential variable of a Use-obstacle satellite of the Answer-no candidate plan, and satellite recognition would proceed. (The nucleus of this satellite would be marked as hypothesized.) Then, the recognizer would recognize piv (without requiring hypothesis generation) as a Use-obstacle satellite of this Use-obstacle satellite. No remaining utterances in R's turn can be related to the candidate Answer-no plan, resulting in the candidate shown on the left in Figure 8.</Paragraph>
      <Paragraph position="36"> Finally, to finish consideration of the recognition of the Answer-yes candidate, since neither the act underlying (12)iv nor the act underlying (12)vi can serve as a satellite of the Answer-yes candidate or its Use-elaboration satellite, hypothesis generation would again be invoked. Hypothesis generation would provide Piii, the hypothesis that R's car is not rtmning, as a plausible explanation for why R is going to take the bus. Thus, piii would be used to instantiate the existential variable of a Use-cause satellite of the Use-elaboration satellite of the Answer-yes candidate plan, and satellite recognition would proceed. (The nucleus of the Use-cause satellite would be marked as hypothesized.) Then, the recognizer would recognize piv (without requiring hypothesis generation) as a Use-cause satellite of this Use-cause satellite. No remaining utterances in R's turn can be related to the candidate Answer-yes plan, resulting in the candidate shown on the right in Figure 8.</Paragraph>
      <Paragraph position="37"> Given the shared beliefs and the coherence rules provided in the model, none of the utterances in R's turn would be recognized as satellites of the other three top-level candidate answer plans. Candidates that do not account for any actual parts of the response are eliminated at the end of phase one. Thus the output of phase one of interpretation would be just the two candidates shown in Figure 8. Phase two would evaluate the Answer-yes candidate as more preferred than the Answer-no candidate, since the former interpretation requires the same number of hypotheses and also accounts for more of R's response.</Paragraph>
    </Section>
    <Section position="2" start_page="405" end_page="406" type="sub_section">
      <SectionTitle>
4.2 Role of the Answer Recognizer in Discourse Processing
</SectionTitle>
      <Paragraph position="0"> As discourse researchers have pointed out (e.g., Reichman 1985; Levinson 1983)) the asking of a yes-no question creates the expectation that R will provide the answer  Green and Carberry Indirect Answers (directly or indirectly), if possible. Other acceptable, though less preferred, responses include I don't know and replies that provide other helpful information. Furthermore, an answer need not be given in the turn immediately following the turn in which the question was asked. For example, in (13) the yes-no question in (13)i is not answered until (13)v, separated by a request for clarification in (13)ii and its answer in (13)iii.  (13) i. Q: Is Dr. Smith teaching CS360 next semester? ii. R: Do you mean Dr. Smithson? iii. Q: Yes.</Paragraph>
      <Paragraph position="1"> iv. R: \[No.\]  v. He will be on sabbatical next semester.</Paragraph>
      <Paragraph position="2"> In Carberry's discourse-processing model for ellipsis interpretation (Carberry 1990), a mechanism is provided for updating the shared discourse expectations of dialogue participants throughout a conversation. Our answer recognizer would have the following role in such an architecture: The answer recognizer would be invoked whenever the current discourse expectation is that R will provide an answer. (If answer recognition were unsuccessful, then the discourse processor would invoke other types of recognizers for other types of responses.) The answer recognizer returns a partially ordered set (possibly empty) of answer discourse plans that it is plausible to ascribe to R as underlying (part or all of) the turn. The final choice of which discourse plan to ascribe to R should be made by the higher-level discourse processor, since it must select an interpretation consistent with the rest of the discourse.</Paragraph>
    </Section>
    <Section position="3" start_page="406" end_page="410" type="sub_section">
      <SectionTitle>
4.3 Comparison to Previous Approaches to Conversational Implicature
</SectionTitle>
      <Paragraph position="0"> Grice (1975) has proposed a theory of conversational implicature to account for certain types of conversational inferences. According to Grice, a speaker may convey more than the conventional meaning of an utterance by making use of the hearer's expectation that the speaker is adhering to general principles of cooperative conversation.</Paragraph>
      <Paragraph position="1"> Two necessary (but not sufficient) properties of conversational implicatures involve cancelability and speaker intention (Grice 1975; Hirschberg 1985). First, potential conversational implicatures may be canceled explicitly, i.e., disavowed by the speaker in the preceding or subsequent discourse context, or even canceled implicitly given a particular set of shared beliefs. In fact, potential implicatures may undergo a change in status from cancelable to noncancelable in the subsequent discourse (Gunji 1981).</Paragraph>
      <Paragraph position="2"> Second, conversational implicatures are part of the intended meaning of an utterance. Grice proposes several maxims of cooperative conversation that a hearer uses as justification for inferring conversational implicatures. However, Grice's theory is inadequate as the basis for a computational model of how conversational implicatures are derived. As frequently noted, Grice's maxims may support spurious or contradictory inferences. To date, few computational models have addressed the interpretation of conversational implicatures.</Paragraph>
      <Paragraph position="3"> Hirschberg's model (Hirschberg 1985) addresses a class of conversational implicatures, scalar implicatures, which overlaps with the class of implicated answers addressed in our model. (That is, scalar implicatures arise in question-answer exchanges as well as in other contexts, and, not all types of implicated answers are scalar implicatures.) According to Hirschberg, a scalar implicature depends upon the existence of a partially ordered set of values that is salient in the discourse context. Her model provides licensing rules that specify, given such a set, which scalar implicatures are  Computational Linguistics Volume 25, Number 3 It is mutually plausible to the agent that (cr-contrast q p*) holds, where q is a proposition and p* is the proposition that p is partly true, if the agent believes it to be mutually believed that q is less than p in a salient partial order, unless it is mutually believed that p is true or that q is not true.</Paragraph>
      <Paragraph position="4"> It is mutually plausible to the agent that (cr-contrast q (not p)) holds, where q and p are propositions, if the agent believes it to be mutually believed that q is an alternate to p in a salient partial order, unless it is mutually believed that p is true or that q is not true.</Paragraph>
      <Paragraph position="5"> Figure 9 Glosses of two coherence rules for cr-contrast.</Paragraph>
      <Paragraph position="6"> licensed in terms of values in the set that are lower than, alternate to, or higher than the value referred to in an utterance. For example, given a salient partially ordered set such that the value for the letter from X is lower than the value for all of the letters in question, in saying (2)ii (repeated below in (14)ii) R licenses the implicature that R has not gotten all of the letters in question.</Paragraph>
      <Paragraph position="7"> (14) i. Q: Have you gotten the letters yet? ii. R: I've gotten the letter from X.</Paragraph>
      <Paragraph position="8"> In our model, the response in (14)ii would be analyzed as generated from an Answer-hedge discourse plan whose nucleus has not been explicitly given and which has a single Use-contrast satellite whose nucleus is expressed in (14)ii. 3a The coherence rules for cr-contrast, which are based upon the notions elucidated by Hirschberg, are glossed in Figure 9. 33 However, the discourse plan operators in our model also characterize a variety of indirect answers that are not scalar implicatures, i.e., indirect answers based on the other coherence relations shown in Table 2.</Paragraph>
      <Paragraph position="9"> A model such as Hirschberg's, which does not take the full response into account, faces certain problems in handling cancellation by the subsequent discourse context (&amp;quot;backwards&amp;quot; cancellation). For example, given a salient partially ordered set such that going to campus is ranked as an alternate to going shopping, Hirschberg's model would predict, correctly in the case of (15) and incorrectly in the case of (16), that R intended to convey a no.</Paragraph>
      <Paragraph position="10">  (15) i. Q: Are you going shopping? ii. R: \[no\] iii. I'm going to campus.</Paragraph>
      <Paragraph position="11"> iv. I have a night class.</Paragraph>
      <Paragraph position="12"> (16) i. Q: Are you going shopping? ii. R: \[yes\] 32 The nucleus of such a plan conveys that the questioned proposition is partly but not completely true. 33 The uppermost rule in the figure is the one applying to this example. The other rule applies to Use-contrast satellites of Answer-no plans.</Paragraph>
      <Paragraph position="13">  Green and Carberry Indirect Answers iii. I'm going to campus.</Paragraph>
      <Paragraph position="14"> iv. The bookstore is having a sale.</Paragraph>
      <Paragraph position="15"> In our model, (16) would be interpreted by recognizing an Answer-yes plan (with a Use-elaboration and a Use-cause satellite underlying (16)iii and (16)iv, respectively) as more plausible than an Answer-no plan, rather than by use of backwards cancellation. 34 In other words, in our model subsequent context can provide evidence for or against a particular interpretation, since a discourse plan may be expressed by multiple utterances. null Also, a model such as Hirschberg's provides no explanation for why potential implicatures may become noncancelable. Our model predicts that a potential implicature of an utterance becomes noncancelable after the point in the conversation when the full discourse plan accounting for that utterance has been attributed to the speaker. For example, imagine a situation in which Q and R mutually intend to discuss two job candidates, A and B. Also, suppose that they mutually believe that they should not discuss any candidate until two letters of recommendation have been received for the candidate, and further, that both letters for B have been received. Our model predicts that the scalar implicature potentially licensed in (17)ii (i.e., that R has not gotten both letters for A yet) is no longer cancelable after R's turn in (17)iv, since by that point, the participants apparently would share the belief that Q had succeeded in recognizing R's discourse plan underlying (17)iiY (17) i. Q: Have you gotten the letters for A yet? ii. R: I've gotten the letter from X.</Paragraph>
      <Paragraph position="16"> iii. Q: Then let's discuss B now.</Paragraph>
      <Paragraph position="17"> iv. R: O.K. I think we should interview B, don't you? Inference of coherence relations has been used in modeling temporal (Lascarides and Asher 1991; Lascarides, Asher, and Oberlander 1992) and other defeasible discourse inferences (Hobbs 1978; Dahlgren 1989). Inference of plausible coherence relations is necessary but not sufficient for interpreting indirect answers. For example, Q also must believe that there is a shared discourse expectation of an answer to a particular question. In other words, in our model, discourse plans provide additional constraints on the beliefs and intentions of the speaker that a hearer uses in interpreting a response. Another limitation of the above approaches is that they provide no explanation for the phenomenon of loss of cancelability described above.</Paragraph>
      <Paragraph position="18"> Plan recognition has been used to model the interpretation of indirect speech acts (Perrault and Allen 1980; Hinkelman 1989) and ellipsis (Carberry 1990; Litman 1986), discourse phenomena that share with conversational implicature the two necessary conditions described above, cancelablity and speaker intention. However, these models are inadequate for interpreting indirect answers, i.e., for deriving an implicated answer p from an indirect answer q. In these models, for p to be derivable from q, it is necessary 34 of course, in the case where R provides only I'm going to campus, both yes and no interpretations would be inferred as equally plausible in our model. Although prosodic information is not used in our model, it is an interesting question for future research whether it can help in recognizing the speaker's intentions in such cases. 35 In other words, it would sound as if R had changed his mind or was contradicting himself if he said In fact I've gotten both letters for A after saying (17)iv.</Paragraph>
      <Paragraph position="19">  Computational Linguistics Volume 25, Number 3 for the hearer to infer that the speaker is performing or at least constructing a domain plan relating p and q. However, q need not play such a role in the speaker's inferred or actual domain plans, as shown in (18). 36 That is, it is not necessary to infer that R has a domain plan involving the renting of a car by X in order to recognize R's intention to convey no.) (18) i. Q: X will be renting a car, won't he? ii. R: \[No.\] iii. He can't drive.</Paragraph>
      <Paragraph position="20"> In other words, these models lack requisite knowledge encoded in our model in terms of possible satellites (based on coherence relations) of top-level discourse plan operators. Also, the above plan-based models face the same problems as Hirschberg's since they do not address multiutterance responses.</Paragraph>
      <Paragraph position="21"> Philosophers (Thomason 1990; McCafferty 1987) have argued for a plan-based theory of implicature as an alternative to Grice's theory. Thomason proposes that implicatures are comprehended by a process of accommodation of the conversational record to fit the inferred plans of the speaker. According to McCafferty, &amp;quot;implicatures are things that the speaker plans that the hearer believe (and that the hearer can realize that the speaker plans that the hearer believe)&amp;quot; (p. 18). He claims that a theory based upon inferring the speaker's plan avoids the problem of predicting spurious implicatures, since the spurious implicature would not be part of the speaker's plan. Our model is consistent with this view of conversational implicature. McCafferty sketches a possible plan-based model to account for the implicated yes answer in (19). 37 (19) i. Q: Has Smith been dating anyone? ii. R: \[Yes.\] iii. He's been flying to New York every weekend.</Paragraph>
      <Paragraph position="22"> Although it was not McCafferty's intention to provide a computational model, but rather to show the plausibility of a plan-based theory of conversational implicature, some limitations of his suggestions for developing a computational model should be noted. First, his proposed rules cannot be used to derive an alternate, plausible interpretation of (19)iii, in which R scalar implicates a no. 38 Our model can account for both interpretations. The first interpretation would be accounted for by an inferred Answer-yes plan with a Use-elaboration satellite underlying (19)iii, while the latter would be accounted for by an inferred Answer-no plan with a Use-contrast satellite underlying (19)iii. More generally, his proposed rules cannot account for types of indirect answers described in our model by coherence relations whose definitions do not involve planning knowledge. Second, even if rules could be added to McCafferty's model to account for a speaker's plan to convey a no by use of (19)iii, his model does not provide a way of using information from other parts of the response, e.g., (20)iv, to help recognize the intended answer. As noted earlier, in our model such information can 36 (18) is based upon (1), modified for expository purposes. 37 (19) is from McCafferty (1987), page 67, and is similar to an example of Grice's. In a Gricean account, this implicature would be justified in terms of the Maxim of Relevance. 38 That is, an interpretation in which flying to New York is mutually believed to be an alternate to dating someone in a salient partially ordered set.</Paragraph>
      <Paragraph position="23">  Green and Carberry Indirect Answers be used to provide evidence favoring one candidate discourse plan over another. (For example, (20)iv would be accounted for by the addition of a Use-obstacle satellite to the Answer-no candidate described above.) (20) i. Q: Has Smith been dating anyone? ii. R: \[No.\] iii. He's been flying to New York every weekend.</Paragraph>
      <Paragraph position="24"> iv. Besides, he's married.</Paragraph>
    </Section>
    <Section position="4" start_page="410" end_page="410" type="sub_section">
      <SectionTitle>
4.4 Summary
</SectionTitle>
      <Paragraph position="0"> This section closes with a summary of the argument for the adequacy of our model as a model of conversational implicature. As discussed earlier, two necessary conditions for conversational implicature are cancelability and speaker intention. We have demonstrated that our model can handle forward and backward cancellation, and provides an explanation for the &amp;quot;loss of cancelability&amp;quot; phenomenon. Regarding speaker intention, in our model a conversationally implicated answer is an answer that R planned that Q recognize (and that Q recognizes that R planned that Q recognize). 39 We have demonstrated how Q's recognition of R's discourse plan (in particular, the goal to provide an answer to the question) can be performed using the knowledge and algorithms in our model. Furthermore, we argue that Q's recognition of R's intention that Q recognize R's plan follows from the role of interpretation in generation, namely, Q and R mutually believe that R will not say what he does unless R believes that Q will be able to interpret the response as intended. In our model, during generation (to be described in Section 5), R constructs a model of Q's beliefs (using R's shared beliefs), and then simulates Q's interpretation of a trial pruned response. R's decision to use the pruned response depends upon whether R believes that Q would still be able to recognize the answer after the plan has been pruned. During interpretation, given the shared discourse expectation that R will provide an answer to Q's yes-no question, Q's use of (Q's) shared beliefs to interpret the response, and Q's belief that R expects that Q will be able to recognize the answer, Q's recognition of a discourse plan for an answer warrants Q's belief that R intended for Q to recognize this intention.</Paragraph>
    </Section>
  </Section>
  <Section position="6" start_page="410" end_page="424" type="metho">
    <SectionTitle>
5. Generation
</SectionTitle>
    <Paragraph position="0"> This section describes our approach to the generation of indirect answers. Generation is modeled as a two-phase process of discourse plan construction. First, in the content planning phase, a discourse plan for a full answer is constructed. Second, the plan pruning phase uses the model's own interpretation capability to determine what information in the full response does not need to be stated explicitly. In appropriate discourse contexts, i.e., in contexts where the direct answer can be inferred by Q from other parts of the full answer, a plan for an indirect answer is thereby generated. When the direct answer must be given explicitly, the result is a plan for a direct answer accompanied by appropriate extra information. (According to the study mentioned in Section 2 \[Stenstrbm 1984\], 85% of direct answers are accompanied by such information. Thus, it is important to model this type of response as well.) While the pragmatic knowledge described in Section 3 is sufficient for interpretation, it is not sufficient for the problem of content planning during generation. Applica39 Applying McCafferty's description of conversational implicature to indirect answers.</Paragraph>
    <Paragraph position="1">  bility conditions prevent inappropriate use of a discourse plan. However, they do not model a speaker's motivation for choosing to provide extra information. Consider (21).</Paragraph>
    <Paragraph position="2"> (21) i. Q: I need a ride to the mall.</Paragraph>
    <Paragraph position="3"> ii. Are you going shopping tonight? iii. R: \[No.\] iv. My car's not running.</Paragraph>
    <Paragraph position="4"> v. The timing belt is broken.</Paragraph>
    <Paragraph position="5"> R's reason for providing the information in (21)iv might have been to give an excuse for not being able to offer Q a ride, and R's reason for providing the information in (21)v might have been to provide an explanation for news in (21)iv that may surprise Q. Furthermore, a full answer might be too verbose if every satellite whose applicability conditions held were included in the full answer. On the other hand, at the time when he is asked a question, R may not hold the primary goals of a potential satellite. (In our model the only goal R is assumed to have initially is the goal to provide the answer.) Thus, an approach to selecting satellites driven only by these satellite goals would fail. To overcome these problems, we have augmented the satellite discourse plan operators, as described in Section 3, with one or more stimulus conditions. Two examples are shown in Figure 10. Stimulus conditions describe general types of situations in which a speaker is motivated to include a satellite during plan construction. They can be thought of as situational triggers, which give rise to new speaker goals (i.e., the primary goals of the satellite operator), and which are the compiled result of deeper planning based upon principles of cooperativity (Grice 1975) or politeness (Brown and Levinson 1978). 4o In order for a satellite to be included, all of its applicability conditions and at least one of its stimulus conditions must be true.</Paragraph>
    <Paragraph position="6"> 40 It was beyond the scope of our research to model recognition of stimulus conditions. We argue in Section 5.3, however, that this does not compromise our approach as a model of conversational implicature.</Paragraph>
    <Paragraph position="7">  Green and Carberry Indirect Answers Our methodology for identifying stimulus conditions was to survey linguistic studies, described in Section 5.1, as well as to analyze the possible motivation of the speaker in the examples in our corpus. The rules used in our model to evaluate stimulus conditions are given in Section 5.2. Section 5.3 presents our implemented generation algorithm, and Section 5.4 illustrates the algorithm with an example.</Paragraph>
    <Section position="1" start_page="412" end_page="415" type="sub_section">
      <SectionTitle>
5.1 Linguistic Studies
</SectionTitle>
      <Paragraph position="0"> In linguistic studies, the reasons given for including extra information 41 in a response to a yes-no question can be categorized as:  * to provide implicitly requested information, * to provide an explanation for an unexpected answer, * to qualify a direct answer, or * politeness-related.</Paragraph>
      <Paragraph position="1"> 5.1.1 Implicitly Requested Information. As mentioned in Section 2, Stenstr6m claims that the typical reason for providing extra information is to answer an implicit whquestion. Kiefer (1980) observes that several types of yes-no questions, when used to perform indirect speech acts, have the property that one or both of the &amp;quot;binary&amp;quot; answers (i.e., yes or no) used alone is an inappropriate response to them. For example, in response to (22)i, 42 when interpreted as (22)ii, an answer of (22)iii or (22)v 43 would be appropriate, but not (22)iv alone.</Paragraph>
      <Paragraph position="2"> (22) i. Q: Is John leaving for Stockholm TOMORROW? ii. Q: When is John leaving for Stockholm? iii. R: Yes.</Paragraph>
      <Paragraph position="3"> iv. R: No.</Paragraph>
      <Paragraph position="4"> v. R: No, John is going to leave the day after tomorrow.</Paragraph>
      <Paragraph position="5">  Kiefer also provides examples of cases where the other binary answer alone is inappropriate, or where either alone is inappropriate.</Paragraph>
      <Paragraph position="6"> Clark (1979) studied how different factors may influence the responder's confidence that the literal meaning of a question was intended and confidence that a particular indirect meaning was intended. In one experiment, in which subjects responded to the question, Do you accept credit cards?, about half of the subjects provided information answering an indirect request of What credit cards do you accept? Clark speculates that the half who included information addressing the indirect request in their response had some, but not necessarily total, confidence that it was intended. According to Levinson (1983), a yes-no question often may be interpreted as a prerequest for another request, i.e., it may be used in the first position of the sequence 41 We are reporting only cases where the extra information may be used as an indirect answer. 42 In this example, Kiefer's (11b), we follow Kiefer's use of capitalization to indicate that tomorrow would be stressed in spoken English. 43 Kiefer's (21b).</Paragraph>
      <Paragraph position="7">  Levinson claims that prerequests are used to check whether the planned request (in T3) is likely to succeed so that a dispreferred response to it can be avoided by Q. Another reason is that, since receiving an offer is preferred to making a request (Schegloff 1979), by making a prerequest, Q gives R the opportunity to offer whatever Q would request in T3, i.e., the sequence would then consist of just T1 and T4. In analyses based on speech act theory, in a sequence consisting of just T1 and T4, the prerequest would be referred to as an indirect speech act.</Paragraph>
      <Paragraph position="8">  viding extra information is to provide an explanation justifying a negative answer. 44 According to Levinson (1983), the presence of an explanation is a distinguishing feature of dispreferred responses to questions and other second parts of adjacency pairs (Schegloff 1972). In an adjacency pair, each member of the pair is produced by a different speaker, and the occurrence of the first part creates the expectation that the second part will appear, although not necessarily immediately following the first member. Levinson claims that dispreferred responses to first parts of adjacency pairs can be identified by structural features such as: * use of pauses or displacement, * prefacing with markers (e.g. uh or well), appreciations, apologies, or refusals, * providing explanations, and * declinations given in an indirect or mitigated manner.</Paragraph>
      <Paragraph position="9"> For example in (23), 45 the marker well is used and an explanation is given. (23) i. Q: What about coming here on the way or doesn't that give you enough time? ii. R: Well no I'm supervising here Although Levinson defines preference in terms of structural features, he notes that there is a correlation between preference and content. For example, unexpected answers to questions, refusals of requests and offers, and admissions of blame are typically marked with features from the above list.</Paragraph>
      <Paragraph position="10">  44 She found that 61% of negative direct answers but only 24% of positive direct answers were accompanied by qualify acts.</Paragraph>
      <Paragraph position="11"> 45 Levinson's example (55).</Paragraph>
      <Paragraph position="12">  Green and Carberry Indirect Answers 5.1.3 Avoid Misunderstanding. Stenstr6m notes that extra information may be given to qualify an answer. Hirschberg (1985) claims that speakers may give indirect answers to block potential unintended scalar implicatures of a yes or no alone. For example in (2), repeated below as (24), R's response is preferable to just no, since that would license the incorrect scalar implicature that R had not received any of the letters in question. However, by use of (24)ii in an appropriate discourse context, R is able to convey explicitly which letter has been received as well as to conversationally implicate that R has not gotten the other letters in question.</Paragraph>
      <Paragraph position="13"> (24) i. Q: Have you gotten the letters yet? ii. R: I've gotten the letter from X.</Paragraph>
      <Paragraph position="14"> 5.1.4 Politeness. StrenstrOm claims that extra information may be given for social reasons. Kiefer notes that extra information may be given as an excuse when the answer indicates that the speaker has failed to fulfill a social obligation. Brown and Levinson (1978) claim that politeness strategies, which may at times conflict with Gricean maxims, account for many uses of language. According to Brown and Levinson, certain communicative acts are intrinsically face-threatening acts (FTAs). That is, doing an FTA is likely to injure some conversational participant's face, or public self-image. For example, orders and requests threaten the recipient's negative face, &amp;quot;the want ... that his actions be unimpeded by others&amp;quot; (p. 67). On the other hand, disagreement or bearing &amp;quot;bad news&amp;quot; threatens the speaker's positive face, the want to be looked upon favorably by others. Further, they claim that politeness strategies can be ranked, and that the greater the threat associated with a face-threatening act, the more motivated a speaker is to use a higher-numbered strategy.</Paragraph>
      <Paragraph position="15"> Brown and Levinson propose the following ranked set of strategies (listed in order from lower to higher rank):</Paragraph>
      <Paragraph position="17"> perform the FTA. (Brown and Levinson claim that this amounts to following Gricean maxims.) perform the FTA with redressive action, i.e., in a manner that indicates that no face threat is intended, using positive politeness strategies (strategies that increase the hearer's positive face). Such strategies include:</Paragraph>
      <Paragraph position="19"> perform the FTA with redressive action, using negative politeness strategies (strategies for increasing negative face). These include: Strategy 6: giving an excuse or an apology perform the FTA off-record, i.e., by use of conversational implicature. In the next section, we provide several stimulus conditions that reflect positive politeness strategy 1 and negative politeness strategy 6. However, although politeness considerations may motivate a speaker to convey an answer indirectly, it is beyond the scope of our generation model to choose between a direct and an indirect answer on this basis.</Paragraph>
    </Section>
    <Section position="2" start_page="415" end_page="419" type="sub_section">
      <SectionTitle>
5.2 Stimulus Conditions
</SectionTitle>
      <Paragraph position="0"> In this section we provide glosses of rules giving sufficient conditions for the stimulus conditions used in our model. (The rules are encoded as Horn clauses in our implementation of the model.) Table 3 summarizes which stimulus conditions appear in which discourse plan operators. As mentioned above, for an instance of a satellite operator to be selected during generation, all of its applicability conditions and at least one of its stimulus conditions must hold.</Paragraph>
      <Paragraph position="1"> 5.2.1 Explanation-indicated. This stimulus condition appears in all of the operators for providing causal explanations. For example in (1), repeated below as (25), R gives an explanation of why R won't get a car.</Paragraph>
      <Paragraph position="2"> (25) i. Q: Actually you'll probably get a car won't you as soon as you get there? ii. R: \[No.\] iii. I can't drive.</Paragraph>
      <Paragraph position="3"> R's response may contribute to greater dialogue efficiency by anticipating a follow-up request for an explanation. The rule for this stimulus condition may be glossed as: (a speaker) s is motivated to give (a hearer) h an explanation for (not p), if s suspects that h suspects that (a proposition) p is true, unless it is the case that s has reason to believe that h will accept (not p) on s's authority, s may acquire the suspicion that h suspects that p is true by means of syntactic clues from the yes-no question, e.g., from the form of the question in (25)i.</Paragraph>
      <Paragraph position="4">  causal operators as explanation-indicated, it represents a different kind of motivation.  Green and Carberry Indirect Answers A yes-no question may be interpreted as a prerequest. Thus, a negative answer to a yes-no question used as a prerequest may be interpreted as a refusal. To soften the refusal, i.e., in accordance with negative politeness strategy 6, the speaker may give an explanation of the negative answer, as illustrated in (21), repeated below in (26). (26) i. Q: I need a ride to the mall.</Paragraph>
      <Paragraph position="5"> ii. Are you going shopping tonight? iii. R: \[No.\] iv. My car's not running.</Paragraph>
      <Paragraph position="6"> v. The timing belt is broken.</Paragraph>
      <Paragraph position="7"> The rule for this stimulus condition may be glossed as: s is motivated to give h an excuse for (not p), if s suspects that h's request, (informifs h p), is a prerequest. Techniques for interpreting indirect speech acts (Perrault and Allen 1980; Hinkelman 1989) can be used to determine whether the rule's antecedent holds.</Paragraph>
      <Paragraph position="8">  5.2.3 Answer-ref-indicated. This condition appears in Use-elaboration, illustrated by (27), 46 and in Use-contrast, illustrated by (28). 47 (27) i. Q: Did you have a hotel in mind? ii. \[What hotel did you have in mind?\] iii. R: \[Yes.\] iv. There's a Holiday Inn right near where I'm working.</Paragraph>
      <Paragraph position="9"> (28) i. Q: You're on that? ii. \[Who's on that?\] iii. R: No no no.</Paragraph>
      <Paragraph position="10"> iv. Dave is.</Paragraph>
      <Paragraph position="11">  In (27), R has interpreted the question in (27)i as a prerequest for the wh-question shown in (27)ii. Thus, (27)iv not only answers the question in (27)i but also the anticipated wh-question in (27)ii. Similarly in (28), R may interpret the question in (28)i as a prerequest for the wh-question in (28)ii, and so gives (28)iv to provide an answer  to both (28)i and (28)ii. The rule for this stimulus condition may be glossed as: s is motivated to provide h with q, if s suspects that h wants to know the referent of a term t in q. As in excuse-indicated, techniques for interpreting indirect speech acts can be used to determine if the rule's antecedent holds. 48 46 From SRI Tapes (1992), tape 1.</Paragraph>
      <Paragraph position="12"> 47 Stenstr6m's (102). A no answer may be conversationally implicated by use of (28)iv alone. 48 However, following Clark (1979), the rule does not require that R be certain that Q was making an indirect request.</Paragraph>
      <Paragraph position="13">  Although Q may not have intended to use (29)i as a prerequest for the question What Verdi operas do you have?, R suspects that the answer to this wh-question might be helpful to Q, and so provides it (in accordance with positive politeness strategy 1). The rule for this stimulus condition may be glossed as: s is motivated to provide h with q, if s suspects that it would be helpful for h to know the referent of a term t in q. The rule's antecedent would hold whenever obstacle detection techniques (Allen and Perrault 1980) determine that h's not knowing the referent of t is an obstacle to an inferred plan of h's. However, not all helpful responses, in the sense described in Allen and Perrault (1980), can be used as indirect answers. For example, even if the clerk (R) at the music store believes that Q's not knowing the closing time could be an obstacle to Q's buying a recording, a response of (30) alone would not convey no since it cannot be coherently related to an Answer-no plan.</Paragraph>
      <Paragraph position="14">  (30) i. R: We close at 5:30 tonight.</Paragraph>
      <Paragraph position="15"> 5.2.5 Clarify-concept-indicated. This stimulus condition appears in Use-elaboration, as illustrated in (31). 49 (31) i. Q: Do you have a pet? ii. R: We have a turtle.</Paragraph>
      <Paragraph position="16"> In (31), R was motivated to elaborate on the type of pet R has since turtles are not prototypical pets. The rule for this stimulus condition may be glossed as: s is motivated to clarify p to h with q, if p contains a concept c, and q provides an atypical instance of c. Stereotypical knowledge would be used to evaluate the rule's antecedent. 5.2.6 Clarify-condition-indicated. This stimulus condition appears in the operator Use-condition, as illustrated by (32). 5o (32) i. Q: Um let me can I make the reservation and change it by tomorrow? ii. R: \[Yes.\] iii. If it's still available.</Paragraph>
      <Paragraph position="17">  In (32), a truthful yes answer depends on the truth of (32)iii. The rules for this stimulus condition may be glossed as: s is motivated to clarify a condition q for p to h if 1) s doesn't know if q holds, or 2) s suspects that q does not hold.</Paragraph>
      <Paragraph position="18"> 49 Example (177) from Hirschberg (1985). 50 From SRI Tapes (1992), tape 10ab.  Green and Carberry Indirect Answers 5.2.7 Clarify-extent-indicated. This stimulus condition appears in Use-contrast, as illustrated by (2), repeated below as (33).</Paragraph>
      <Paragraph position="19"> (33) i. Q: Have you gotten the letters yet? ii. R: I've gotten the letter from X.</Paragraph>
      <Paragraph position="20">  On the strict interpretation of (33)i, Q is asking whether R has gotten all of the letters, but on a looser interpretation, Q is asking if R has gotten any of the letters. Then, if R has gotten some but not all of the letters, just yes would be untruthful. However, if Q is speaking loosely, then just no might lead Q to erroneously conclude that R has not gotten any of the letters. R's answer circumvents this problem, by conveying the extent to which the questioned proposition (on the strict interpretation) is true. The rules for this stimulus condition may be glossed as: s is motivated to clarify to h the extent q to which p is true, or the alternative q to p which is true, if s suspects that h does not know if q holds, and s believes that q is the highest expression alternative to p that does hold. According to Hirschberg (1985) (following Gazdar), sentences pi and pj (representing the propositional content of two utterances) are expression alternatives if they are the same except for having comparable components ei and ej, respectively. As mentioned earlier, Hirschberg claims that in a discourse context there may be a partial ordering of values that the discourse participants mutually believe to be salient. She claims that the ranking of ei and ej in this ordering can be used to describe the ranking of pi and pj. In the above example, (33)ii is a realization of the highest true expression alternative to the questioned proposition, p, i.e., the proposition that R has gotten all the letters. 51 5.2.8 Appeasement-indicated. This stimulus condition appears in Use-contrast, as illustrated by (34). 52 (34) i. Q: Did you manage to read that section I gave you? ii. R: I've read the first couple of pages.</Paragraph>
      <Paragraph position="21"> In (34), R conveys that there is some (though not much) truth to the questioned proposition in an effort to soften his answer (in accordance with positive politeness strategy 1). More than one stimulus condition may motivate R to include the same satellite. For example, in (34), R may have been motivated also by clarify-extent-indicated, which was described above. However, it is possible to provide a context for (35) where appeasement-indicated holds but clarify-extent-indicated does not, or a context for (34) where the converse is true.</Paragraph>
      <Paragraph position="22"> (35) i. Q: Did you wash the dishes? ii. R: I brought you some flowers.</Paragraph>
      <Paragraph position="23"> The rules for this stimulus condition may be glossed as: s is motivated to appease h with q for p not holding or only being partly true, if s suspects that (not p) is undesirable 51 Recall that additional constraints on p and q arise from the applicability conditions of operators containing this stimulus condition, namely Use-contrast in this case. Thus, another constraint is that it is plausible that cr-contrast holds. The coherence rule for cr-contrast was described in Section 4.3. 52 Example (56) from Hirchberg (1985).  1. Efficiency: explanation-indicated, answer-ref-indicated 2. Accuracy: clarify-concept-indicated, clarify-extent-indicated,clarify-condition-indicated 3. Politeness: excuse-indicated, appeasement-indicated, substitute-indicated  to h but that q is desirable to h. The antecedents to this rule would be evaluated using heuristic rules and stereotypical and specific knowledge about h's desires. For example, two heuristics of rational agency that might lead to beliefs about h's desires are 1) if an agent wants you to perform an action A, then your failure to perform A may be undesirable to the agent, and 2) if an agent wants you to do A, then it is desirable to the agent that you perform a part of A.</Paragraph>
      <Paragraph position="24"> 5.2.9 Summary. In summary, the stimulus conditions in our model can be classified according to three general principles, as shown in Table 4. The first category, efficiency, includes the motivation to provide implicitly requested information as well as to provide an explanation for unexpected information. In other words, giving this type of extra information contributes to the efficiency of the conversation by eliminating the need for follow-up wh-questions or follow-up why? or why not? questions, respectively. In the category of accuracy, in addition to the reason cited by Hirschberg (which is represented in our model as clarify-extent-indicated), we have identified two other reasons for giving extra information, which contribute to accuracy. The category of politeness includes reasons for redressing face-threatening acts using positive and negative politeness strategies.</Paragraph>
    </Section>
    <Section position="3" start_page="419" end_page="422" type="sub_section">
      <SectionTitle>
5.3 Generation Algorithm
</SectionTitle>
      <Paragraph position="0"> The inputs to generation consist of: * the set of discourse plan operators (described in Section 3) augmented with stimulus conditions, * the set of coherence rules (also described in Section 3), * the set of stimulus condition rules (described in Section 5.2), * R's beliefs (including the discourse expectation that R will provide an answer to some questioned proposition p), and * the semantic representation of p.</Paragraph>
      <Paragraph position="1"> The model presupposes that when answer generation begins, the speaker's (R's) only goal is to satisfy the above discourse expectation. R's nonshared beliefs (including beliefs whose strength is not necessarily certainty) about Q's beliefs, intentions, and preferences are used in generation to evaluate whether a stimulus condition holds. The output of the generation algorithm is a discourse plan that can be realized by a tactical generation component (McKeown 1985). 53 53 The plan that is output specifies an ordering of discourse acts based upon the ordering of coherence relations specified in the discourse plan operators. However, reordering may be required, e.g., to model a speaker who has multiple goals.</Paragraph>
      <Paragraph position="2">  Green and Carberry Indirect Answers The answer generation algorithm has two phases. In the first phase, content planning, the generator creates a discourse plan for a full answer, i.e., a direct answer and extra appropriate information. In the second phase, plan pruning, the generator determines which propositions of the planned full answer do not need to be explicitly stated. For example, given an appropriate model of R's beliefs, the system would generate a plan for asserting only the proposition conveyed in (36)v and (36)vi as an answer to (36)i. 54 (36) i. Q: Is Mark here \[at the office\]? ii. R: \[No.\] iii. \[He's at home.\] iv. \[He is caring for his daughter.\] v. His daughter has the measles, vi. but he's logged on.</Paragraph>
      <Paragraph position="3"> An advantage of this approach is that, even when it is not possible to omit the direct answer, a full answer is generated.</Paragraph>
      <Paragraph position="4"> 5.3.1 Content Planning. Content planning is performed by top-down expansion of an answer discourse plan operator. First, each top-level answer discourse plan operator is instantiated with the questioned proposition until one is found such that its applicability conditions hold. s5 Next, the satellites of this operator are expanded (recursively). The algorithm for expanding a satellite adds each instance of a satellite such that all of its applicability conditions and at least one of its stimulus conditions hold. Thus, different instantiations of the same type of satellite may be included in a plan for different reasons. For example, (36)iii and (36)vi both realize Use-contrast satellites, the former included due to the answer-ref-indicated stimulus condition, and the latter due to the substitute-indicated stimulus condition.</Paragraph>
      <Paragraph position="5"> For each stimulus condition of a satellite, our implementation of the algorithm uses a theorem prover to search the set of R's beliefs (encoded as Horn clauses) for a proposition satisfying a formula consisting of a conjunction of the applicability conditions and that stimulus condition. A proposition satisfying each such formula is used to instantiate the existential variable of the satellite operator. For example, to generate the response in (36), the following formula would be constructed from the Use-contrast operator's applicability conditions and one of its stimulus conditions, (answer-ref-indicated), where p is the proposition that Mark is at the office, and ?q is the existential variable to be instantiated.</Paragraph>
      <Paragraph position="6"> ((and (bel s (cr-contrast ?q (not p))) (Plausible (cr-contrast ?q (not p))) (answer-ref-indicated s h ?q))) The result of the search is to instantiate ?q with the proposition that Mark is at home, due to the speaker's belief that the hearer might have been using (36)i as a prerequest 54 Parts (36)i-(36)v were overheard by one of the authors in a naturally occurring dialogue, and (36)vi was added for expository purposes.</Paragraph>
      <Paragraph position="7">  Computational Linguistics Volume 25, Number 3 for the question, Where is Mark? For a more complete description of how R's response in (36) is generated, see Section 5.4.</Paragraph>
      <Paragraph position="8"> We have employed a simple approach to planning because the focus of our research was on the use of the response as an indirect answer, i.e., on aspects of response generation that play a role in interpreting the implicature. In a more sophisticated discourse planning formalism, such as argued for in Young, Moore, and Pollack (1994), it would be possible to represent and reason about other intended effects of the response. (In our model, the effects or primary goals are used in interpretation but their only role in generation is in simulated interpretation. However, their role in interpretation is important; they constrain what discourse plans can be ascribed to the speaker.) While we believe that use of more sophisticated planning formalisms is well motivated for discourse generation in general, we leave the problem of generating indirect answers in such formalisms for future research. The use of stimulus conditions to motivate the selection of optional satellite operators is sufficient for our current goals.  plan representing a full answer, is the input to the plan pruning phase of generation. The expanded plan is represented as a tree of discourse acts. The goal of this phase is to make the response more concise, i.e., to determine which of the planned acts can be omitted while still allowing Q to infer the full discourse plan. 56 To do this, the generator considers each of the acts in the frontier of the tree from right to left. (This ensures that a satellite is considered before its related nucleus.) The generator creates a trial plan consisting of the original plan minus the nodes pruned so far and minus the current node. Then, using the answer recognizer, the generator simulates Q's interpretation of a response containing the information that would be given explicitly according to the trial plan. In the simulation, R's shared beliefs are used to model Q's shared beliefs. If Q could infer the full plan (as the most preferred interpretation), then the current node can be pruned. Otherwise, it is left in the plan and the next node is considered. For example, consider Figure 11 as we illustrate the possible effect of pruning on a full discourse plan. The leaf nodes, representing discourse acts, are numbered 1 through 8. Arcs labeled N and S lead to a nucleus or satellite, respectively. Node 8 corresponds to the direct answer. Plan pruning would process the nodes in order from 1 to 8. The maximal set of nodes that could be pruned in Figure 11 is the set containing 2, 3, 4, 7, and 8. That is, nodes 2 through 4 might be inferable from 1, node 7 from 5 or 6, and node 8 from 4 or 7, but nodes 1, 5, and 6 cannot be pruned since they are not inferable from other nodesY In the event that it is determined that no node can be pruned, the full plan would be output. The interpretation algorithm (described in Section 4) would use hypothesis generation to recognize missing propositions other than the direct answer, i.e., the propositions at nodes 2, 3, 4, and 7.</Paragraph>
      <Paragraph position="9"> To comment on the role of interpretation in generation, it is a key component of our claim to have provided an adequate model of conversational implicature. Given the shared discourse expectation that R will provide an answer to Q's yes-no question, Q's use of (Q's) shared beliefs to interpret the response, and Q's belief that R expects that Q will be able to recognize the answer, Q's recognition of a discourse plan for an answer warrants Q's belief that R intended for Q to recognize this intention. In particular, R would not have pruned the direct answer unless, given the beliefs that R presumes 56 Conciseness is not the only possible motive for omitting the direct answer. As mentioned earlier, an indirect answer may be used to avoid performing a face-threatening act. However, it is beyond the scope of our model to determine whether to omit the direct answer on grounds of politeness. 57 In fact, leaves that have no satellites of their own cannot be pruned.</Paragraph>
      <Paragraph position="10">  \[He is caring for his daughter.\] His daughter has the measles.</Paragraph>
      <Paragraph position="11"> but he is logged on.</Paragraph>
      <Paragraph position="12"> Generation example: exchange.</Paragraph>
      <Paragraph position="14"> to be shared with Q, R believes that Q will be able to recognize a chain of mutually plausible coherence relations from the actual response to the intended answer, and thus be able to recognize R's plan. Note that although stimulus conditions are not recognized during interpretation in our approach, the model does account for the recognition of those parts of the plan concerning the answer. For example, although Q may not know whether R was motivated by excuse-indicated or explanation-indicated to provide (21)iv in response to (21)ii, Q can recognize R's intention to convey a no by Q's recognition of (21)iv as the nucleus of a Use-Obstacle satellite of Answer-No.</Paragraph>
      <Paragraph position="15"> Thus, Q can thereby attribute to R the primary goal of the Answer-No plan to convey a no.</Paragraph>
    </Section>
    <Section position="4" start_page="422" end_page="424" type="sub_section">
      <SectionTitle>
5.4 Generation Example
</SectionTitle>
      <Paragraph position="0"> This example models R's generation of the response in the exchange shown in Figure 12, which repeats (36). The discourse plan constructed by the algorithm is depicted in Figure 13, where (a) through (d) refer to communicative acts that could be performed by saying the sentences with corresponding labels in Figure 12. Square brackets in the plan indicate acts that have been pruned, i.e., that are not explicitly included in the response.</Paragraph>
      <Paragraph position="1"> First, each top-level answer operator is instantiated with the questioned proposition, p, the proposition that Mark is at the office. An (Answer-no s h p) plan would be selected for expansion since its applicability conditions can be proven. To expand this plan, the algorithm attempts to expand each of its satellites as described in Section 5.3.1. The generation algorithm searches for (at most) one instance of a satellite for each possible motivation of a satellite. In this example, two satellites of (Use-contrast s h (not p)) are added to the plan. In one, motivated by the Answer-ref stimulus condi- null tion, the existential proposition is instantiated with pa, the proposition that Mark is at home. In other words, proposition, pa was found to satisfy ?q in formula 1 below.</Paragraph>
      <Paragraph position="2"> 1. ((and (bel s (cr-contrast ?q (not p))) (Plausible (cr-contrast ?q (not p))) (answer-ref-indicated s h ?q))) In the other (Use-contrast s h (not p)) satellite, motivated by the substitute-indicated stimulus condition, the existential proposition is instantiated with pd, the proposition that Mark is logged on. That is, Pa was found to satisfy ?q in formula 2 below.</Paragraph>
      <Paragraph position="3"> 2. ((and (bel s (cr-contrast ?q (not p))) (Plausible (cr-contrast ?q (not p))) (substitute-indicated s h ?q))) The former (Use-contrast s h (not p)) satellite (i.e., the one constructed using pa) can be expanded by adding a (Use-cause s h pa) satellite to it. This satellite's existential variable is instantiated with Pb, the proposition that Mark is caring for his daughter, which was found to satisfy ?q in formula 3 below.</Paragraph>
      <Paragraph position="4"> 3. ((and (bel s (cr-cause ?q pa)) (Plausible (cr-cause ?q pa)) (explanation-indicated s h pa))) Finally, this satellite is expanded using pc, the proposition that Mark's daughter has the measles, which was found to satisfy ?q in formula 4 below.</Paragraph>
      <Paragraph position="5"> 4. ((and (bel s (cr-cause ?q pb)) (Plausible (cr-cause ?q Pb)) (explanation-indicated s h Pb))) The output of phase one is a discourse plan for a full answer, as shown in Figure 14. The second phase of generation, plan pruning, will walk the tree bottom-up. The root  Plan for full answer (before pruning).</Paragraph>
      <Paragraph position="6"> of each subtree has been annotated with a sequence number to show the order in which a subtree is visited in the bottom-up traversal of the tree, i.e., 1 through 5. Since subtree 1 has no satellites (only a nucleus), the traversal moves to subtree 2. For the same reason, the traversal moves to subtree 3. Next, the nucleus of subtree 3 is tentatively pruned, i.e., a trial response consisting of the direct answer plus (a), (c), and (d) is created. Simulated interpretation of this trial response results in the inference of a discourse plan identical to the full plan as the most preferred (in fact, the only) interpretation of the trial response. Thus, (b) can be pruned, and subtree 4 is considered next. By a similar process, (a) is also pruned. Last, the tree with root labeled 5 is examined, and it is determined that the direct answer (no) can also be pruned. The final result of the traversal is that the direct answer, (a), and (b) are marked as pruned, and a response consisting of just acts (c) and (d) is returned by the generator.</Paragraph>
    </Section>
  </Section>
  <Section position="7" start_page="424" end_page="425" type="metho">
    <SectionTitle>
5.5 Related Work in Generation
</SectionTitle>
    <Paragraph position="0"> This work differs from most previous work in cooperative response generation in that the information given in an indirect answer conversationally implicates the direct answer. Hirschberg (1985) implemented a system that determines whether a yes or no alone licenses any unwanted scalar implicatures, and if so, proposes alternative true scalar responses that do not. In our model, that type of response is generated by constructing a response from an Answer-no or Answer-hedge operator having a single Use-contrast satellite, motivated by clarify-extent-indicated, as illustrated in Section 5.2.7. 58 However, Hirschberg's model does not account for other types of indirect answers, which can be constructed using the other operators (or other combinations of the above operators) in our model, nor for other motives for selecting Use-contrast such as answer-ref-indicated and appeasement-indicated.</Paragraph>
    <Paragraph position="1"> Rhetorical or coherence relations (Grimes 1975; Halliday 1976; Mann and Thompson 1988) have been used in several text-generation systems to aid in ordering parts of a text (e.g., Hovy 1988) as well as in content planning (e.g., McKeown 1985; Moore and Paris 1993). The discourse plan operators based on coherence relations in our model 58 As mentioned earlier, the coherence rules for cr-contrast as well as the rules for clarify-extent-indicated make use of notions elucidated by Hirschberg (1985).  Computational Linguistics Volume 25, Number 3 (i.e., the operators used as satellites of top-level operators) play a similar role in content planning. However, none of the above approaches model the speaker's motivation for selecting optional satellites. Stimulus conditions provide principled discourse-level knowledge (based upon principles of efficiency, accuracy, and politeness) for choice of an appropriate discourse strategy. Also, stimulus conditions enable content selection to be sensitive not only to the current discourse context, but also to the anticipated effect of a part of the planned response. Finally, none of the above systems incorporate a model of discourse plan recognition into the generation process, which enables indirect answers to be generated in our model.</Paragraph>
    <Paragraph position="2"> Moore and Pollack (1992) show the need to distinguish the intentional and informational structure of discourse, where the latter is characterized by the sort of relations classified as subject-matter relations in RST. In our model, the operators used as satellites of top-level answer discourse plan operators are based on relations similar to RST's subject-matter relations. The primary goals of these operators are similar to the effect fields of the corresponding RST relation definitions. However, our model does distinguish the two types of knowledge. In our model stimulus conditions reflect, though they do not directly encode, communicative subgoals leading to the adoption of informational subgoals. For example, the explanation-indicated stimulus condition may be triggered in situations when the responder's communicative subgoal would lead R to select a Use-cause satellite of Answer-yes, rather than a Use-elaboration satellite. Moore and Paris (1993) argue that it is necessary for generation systems to represent not only the speaker's top-level goal, but also the communicative subgoals that a speaker hoped to achieve by use of an informational relation so that, if that subgoal is not achieved, then an alternative rhetorical means can be tried. Although stimulus conditions do reflect the speaker's motivation for including satellites in a plan, it was beyond the scope of our work to address the problem of failure to achieve a subgoal of the original response. Therefore, our system does not record which stimulus condition motivated a satellite; if the stimulus condition was recorded in the final plan then our system would have access to information about the speaker's motivation for the satellite. In our current approach, if a follow-up question is asked then a response to the follow-up question is planned independently of the previous response. However, if R's beliefs have changed since the original question was asked by Q (e.g., as a result of information about Q's beliefs obtained from Q's follow-up question), then it is possible in our approach for R's response to contain different information. Furthermore, in our approach the original response may provide the information that a questioner would have to elicit by follow-up questions in a system that can provide only direct answers.</Paragraph>
    <Paragraph position="3"> Finally, our use of interpretation during plan pruning has precursors in previous work. In Horacek's,approach to generating concise explanations (Horacek 1991), a set of propositions representing the full explanation is pruned by eliminating propositions that can be derived from the remaining ones by a set of contextual rules. Jameson and Wahlster (1982) use an anticipation feedback loop algorithm to generate elliptical utterances.</Paragraph>
  </Section>
class="xml-element"></Paper>