In a typical expert-consultation dialogue, one participant (hereafter referred to as the executing agent or EA) has a goal that he wants to achieve and is working with the other participant (referred to as the consulting agent or CA) to construct a plan for achieving this goal. Although both the plan construction process and the con­versation are collaborative activities, this does not mean that people always believe what they are told. In fact, part of the collaborative activity of conversation is negoti­ation of conflicting beliefs. This negotiation is particularly important in task-oriented expert-consultation dialogues, since the participants must resolve any conflicting be­liefs in order to work together effectively to devise a plan that is both well-formed and addresses the executing agent's needs. Thus, a robust natural language consultation system must be able to handle negotiation subdialogues.

Even though there is wide agreement that negotiation is an integral part of multi-agent activity, previous natural language understanding systems have been unable to handle negotiation subdialogues such as the following:

f Department of Physics, Computer Science, and Engineering, Newport News, VA 23606, USA 1 For exposition purposes, we will use the masculine gender when referring to EA and the feminine gender when referring to CA.

(2) S2: Dr. Smith is teaching CS360.

To identify how speakers express doubt, we analyzed a corpus of naturally occurring dialogues in the domains of financial planning, university courses, real estate, pets, taxes, and travel. The real estate, pets, and financial planning (Harry Gross Transcripts 1982) dialogues were transcribed from radio talk shows, the taxes and travel (SRI Transcripts 1992) dialogues were transcribed from tapes of simulated interactions, and the university courses dialogues (Columbia University Transcripts, 1985) were transcribed from student advisement sessions. In the corpus we found instances in which a speaker expressed doubt at a proposition by contending that some other conflicting proposition was true. In addition, we extracted other examples of such expressions of doubt from the dialogues in novels. These kinds of expressions of doubt can be realized as surface negative questions or tag questions and are often accompanied by the cue word but, as in the following example taken from the Harry Gross financial planning dialogues (Harry Gross Transcripts 1982) in which S2's last utterance expresses doubt at Si's recommendation:

SI: I would like to see that into an individual retirement account rollover in a mutual fund group.

SI: Yes.

However, our corpus analysis also provided instances where surface negative and tag questions were used to seek verification, such as in the following excerpt from the set of financial planning dialogues:

SI: And if you have more money left after you pay the taxes what difference does it make if you pay a few bucks more in taxes?

S2: I'm telling my wife but she won't listen.

SI: Well maybe she'll listen to me. If you get 200 bucks—isn't it better to have 200 bucks and uh have 200 left than to have nothing at all?

Our recognition algorithm has only been concerned with recognizing instances in which a speaker expresses doubt by contending that some other proposition is true. However, in our corpus, speakers also expressed doubt in the following ways, and our future work will include extending our system to handle these: speaker brings into focus a feature or proposition that is already part of the dialogue context but that is intended to discredit the proposition being doubted.• Drawing attention to an inconsistent feature or proposition: The

These utterances were often realized as an elliptical fragment and included "And you're how old?", "Even though it's four more years?", and "At my age?"

• Drawing attention to violated expectations: The speaker mentions an expectation that is inconsistent with the doubted proposition. An example of this from our corpus is "You're kidding, what happened to the seventy-eight dollar fares or those sort of things?"

2 We found a very few instances in which the speaker asked the hearer if he was sure the conflicting proposition wasn't true; an example from our corpus is "Are you sure he didn't name himself as attorney for the estate?" taken from the Harry Gross financial planning dialogues (Harry Gross Transcripts 1982). Our current system does not handle such expressions of doubt.

3 Walker (1996) analyzed the Harry Gross financial planning dialogues (Harry Gross Transcripts 1982) to identify features that distinguish acceptance from rejection. However, she did not consider expressions of doubt and some of her rejections would fall into our "express doubt" category.

When a listener in a collaborative interaction does not accept the proposition that the speaker is trying to convey, a negotiation subdialogue ensues in which the participants attempt to "square away" (Joshi 1982) their disparate beliefs. Negotiation subdialogues often involve complex discourse acts that implicitly refer to some proposition that is part of the existing dialogue context. We have found that such complex discourse acts require evidence for their recognition. Since the motivation for our work is the recognition of communicative acts that occur in negotiation subdialogues, this section examines in detail how plausibility and evidence affect the recognition of one kind of complex discourse act, an expression of doubt.

For a listener to recognize a discourse action, it must be plausible that the speaker holds the requisite beliefs for performing the action. (A belief is plausible if the avail­able evidence does not refute it.) For example, in order for a listener to interpret an utterance as felicitously asking a question in order to obtain information, it must be plausible that the speaker does not already know the information and that the speaker believes that the listener may be able to provide it. Similarly, to interpret an utterance as expressing doubt at a proposition PdouU by contending P„ it must be plausible that the speaker holds certain beliefs. Consider a university setting in which each course has only one instructor and a speaker uses the proposition P„ that Dr. Brown teaches Architecture, to express doubt at the proposition Vdoubl, that Dr. Smith is teaching Ar­chitecture, as in utterance (9) of Figure 1. In order to intend (9) as an expression of doubt in a collaborative dialogue, the speaker must believe 1. that the hearer has some belief in Pdoubt

3. that if Pi is true, then Pdoubt is not

The hearer must be able to plausibly ascribe each of these beliefs to the speaker in recognizing the expression of doubt. First, it must be plausible that the speaker believes that the hearer has some belief in the proposition that is being doubted, since it is pointless in a collaborative dialogue to express doubt at something about which there is no disagreement. It must also be plausible that the speaker has some belief in P„

4 We are using "express doubt" in the sense of challenging the truth of a proposition.

3.1.1 Evidence for a Generic Discourse Act. A number of researchers (Reichman 1978, 1985; Grosz and Sidner 1986; Polanyi 1986; Cohen 1987; Hirschberg and Litman 1987; Litman and Allen 1987; Schiffrin 1987; Hinkelman 1989; Litman and Hirschberg 1990; Knott and Dale 1994; Knott and Mellish 1996; Marcu 1997) have investigated the use in discourse of special words and phrases such as but, anyway, and by the way. They found that these clue words, or discourse markers, have a number of different functions, including indicating the role of an utterance in the dialogue, conveying the relationship between utterances, suggesting shifts in focus of attention, conveying the structure of the discourse, etc.

Consider again the dialogue shown in Figure 1. If EA had followed (7)-(8) with (9)a. EA: Isn't Architecture one of our required courses?

then EA's utterance would not be interpreted as expressing doubt but would instead be understood as merely seeking information about the Architecture course. However, if this utterance is preceded by the clue word but, as in (9b) below, (9)b. EA: But isn't Architecture one of our required courses?

then the utterance is expressing doubt, though we have difficulty ascertaining the reason for this doubt—perhaps EA believes that Dr. Smith does not teach courses that students are required to take!

Thus, clue words comprise one source of evidence in the recognition of discourse acts. In particular, a clue word can provide evidence for a generic discourse act, such as Express-Doubt, but it remains for other sources to resolve what is being doubted.

World knowledge in the form of stereotypical beliefs is another source of evidence that the speaker holds the requisite beliefs for a particular discourse act. For example, world knowledge can provide evidence for the third speaker belief, that if P, is true, then Pdoubt is not. Suppose that it is stereotypically believed that prestigious fellow­ships are awarded for sabbaticals, that faculty on sabbatical do not teach, and that faculty only teach in their area of expertise. Consider the dialogue shown in Figure 2. After (13), there are two propositions that have been conveyed by CA but not yet completely accepted by EA: the proposition that Dr. Smith is not on sabbatical and the proposition that Dr. Smith is teaching CS360, communicated by utterances (13) and (11), respectively. A subsequent utterance might express doubt at one of these propositions or might forego the opportunity to doubt them, perhaps by pursuing some discourse act unrelated to either of the propositions. Consider the following three possible continuations of the dialogue:

b. EA: Isn't Dr. Smith a theory person?

c. EA: Isn't Dr. Smith an excellent teacher?

While (14a) and (14b) seem to be expressing doubt, (14c) is simply seeking further information about Dr. Smith. The reason for this difference in interpretation is that in the case of (14a) and (14b), evidence from world knowledge suggests that EA believes that P; (the proposition that EA contends is true) implies that one of the two open propositions is false, whereas no such evidence exists in the case of (14c). In the case of (14a), since it is stereotypically believed that prestigious fellowships are awarded for sabbaticals, EA's utterance should be interpreted as expressing doubt at the proposition that Dr. Smith is not on sabbatical. In the case of (14b), since Dr. Smith being a theory person is an alternative to Dr. Smith being a systems person and it is stereotypically believed that being a systems person is necessary for teaching CS360 (a systems course), EA's utterance would instead be interpreted as expressing doubt at the proposition that Dr. Smith is teaching CS360. Thus, world knowledge in the form of stereotypical beliefs is another source of evidence that the speaker holds the requisite beliefs for performing a particular discourse act.

If EA had uttered (14c), EA's utterance would be interpreted as merely seeking new information since there is no domain knowledge suggesting that EA believes that Dr. Smith being an excellent teacher contributes to determining whether Dr. Smith is (11) CA: Dr. Smith is teaching CS360.

A dialogue with two open propositions.

on sabbatical or to identifying the instructor of CS360. Note that (14c) demonstrates why plausibility alone is insufficient for recognition. Although there is no evidence that EA believes that Dr. Smith being an excellent teacher implies that Dr. Smith is on sabbatical or that Dr. Smith is not teaching CS360, there is also no evidence to the contrary, and thus it is plausible that EA believes that Dr. Smith being an excellent teacher indicates that he is on sabbatical or that he is not teaching CS360. This is not sufficient, however, to interpret (14c) as an expression of doubt.

An agent can infer from a dialogue many of the beliefs of the other participant. These acquired beliefs about the other participant's beliefs form one kind of contex­tual knowledge that can be used as evidence for the beliefs listed above. In addition, contextual knowledge determines the salience (or degree of prominence) of propo­sitions at the current point in the dialogue, and salience is a factor that constrains the interpretation of coherent discourse actions. Consider the first three utterances in the dialogue shown in Figure 2. EA's acceptance of CA's telling of the proposition that Dr. Smith is teaching CS360 establishes the mutual belief that CA believes that Dr. Smith is teaching CS360 and thus provides evidence for the first belief; in addi­tion, the proposition that Dr. Smith is teaching CS360 becomes salient and is added to the dialogue context. Thus, while an utterance such as (12a) might be used following (11) to express doubt at the statement that Dr. Smith is teaching CS360, it cannot be used following (11) to express doubt at the proposition that Dr. Smith teaches CS410 because 1) there is no reason for EA to believe that CA has any belief in the proposition that Dr. Smith teaches CS410, and 2) the proposition that Dr. Smith teaches CS410 is not salient at this point in the dialogue.

In addition, contextual knowledge plays two other roles in the recognition of discourse acts. First, in the case of expressions of doubt, contextual knowledge dis­tinguishes propositions that have not yet been accepted by the speaker and thus are open for rejection. Consider again the dialogue in Figure 2. After (13), there are two propositions that have not yet been accepted by EA and are thus open for rejection by EA. If EA were to continue with (14b), repeated below, then EA would again be expressing doubt at the proposition that Dr. Smith is teach­ing CS360 and would have implicitly conveyed acceptance of the proposition that

5 Note that here EA is only accepting CA's felicitous telling of the proposition, but EA is not adopting the proposition as one of his own beliefs.

Grosz and Sidner (1986) claim that a robust model of understanding must use multi­ple knowledge sources in order to recognize the complex relationships that utterances have to one another. We have developed an algorithm that combines linguistic, world, and contextual knowledge, such as that identified in Section 3, in order to recognize complex discourse acts, including one kind of expression of doubt. Linguistic knowl­edge consists of clue words and the surface form of the utterance; world knowledge includes a set of stereotypical beliefs that users generally hold and recipes for perform­ing discourse acts; and contextual knowledge consists of a model of the user's beliefs acquired from the preceding dialogue, the current structure of the discourse, the ex­isting focus of attention (that aspect of the task on which the participants' attention is currently centered), and the relative salience (degree of prominence) of propositions in the discourse.

The remainder of this section presents the core ideas of our algorithm. Section 4.1 shows why it is necessary to capture varying degrees of belief in a proposition and presents the multistrength belief model used in our system. Section 4.2 describes our representation of recipes for actions and shows how our recipe for an Inform action refrains from assuming that the listener will adopt the communicated proposition as part of his own beliefs; it also presents a recipe for expressing doubt and shows how constraints on the speaker's beliefs are captured in the recipe's applicability con­ditions. Section 4.3 gives an overview of our dialogue model. Section 4.4 describes how chaining is used to hypothesize a sequence of higher-level discourse acts that a speaker may be performing; Section 4.5 introduces the notion of a discourse action that requires evidence for its recognition; Section 4.6 discusses how our model recog­nizes implicit acceptance of a discourse act; and Section 4.7 presents our recognition algorithm that uses a combination of linguistic, world, and contextual knowledge in recognizing discourse acts.

As argued in Section 3, if a speaker is expressing doubt at a proposition Pdoubt by contending some other proposition P„ then the speaker must have some belief in P,. Evidence for this belief is often provided by the surface form of the speaker's utterance, such as an utterance of the form "Isn't Pi?"—for example, "Isn't Dr. Brown teaching Architecture?" But if we treat such an utterance as conveying certain belief that P, is true, then we cannot handle situations in which an utterance such as this is merely requesting verification since a speaker cannot felicitously seek verification of a proposition that he already knows is true. Therefore, since modeling only ignorance and certainty is inadequate for recognizing complex discourse acts, it is necessary to model the strength of an agent's beliefs.

4.1.1 Representing Varying Degrees of Belief. We use a multistrength model of belief, which captures not only ignorance and certainty about the truth of a proposition but also several degrees of belief in between. Utterances of the form "Isn't Pi?" are treated as conveying a strong (but uncertain) belief in P,. In this way, our system is able to handle instances in which an utterance of the form "Isn't Pi?" is used to request verification (since the utterance is viewed as conveying uncertainty about P,) as well as instances in which it is used to express doubt (since the utterance is viewed as conveying some belief in P,).

Our multistrength belief model maintains three degrees of belief: certain belief (a belief strength of C); strong but uncertain belief, as in "Isn't Dr. Brown teaching Architecture?" (a belief strength of S); and weak belief, as in "I think that Dr. Cayne might bean education instructor" (a belief strength of W). Three degrees of disbelief (indicated by attaching a subscript of N, such as Sn to represent strong disbelief and Wn to represent weak disbelief) are also maintained, and one degree indicating no belief about a proposition (a belief strength of 0). We adopted three degrees of positive and negative belief in our model because that was the minimum number of degrees required for modeling the beliefs communicated in the dialogues that we examined and in the negotiation subdialogues that our system is intended to handle.

Although an agent has some specific strength of belief in a proposition, the other agent may not always know precisely what that strength of belief is but may be able to bound it—for example, he may be able to say that the first agent has some belief in a proposition. Our belief model uses belief intervals to capture this, where a belief interval specifies the range of strengths within which an agent's beliefs are thought to fall.

Allen and Perrault (1980) noted the need to represent an agent's wanting to know the referent of a term in a proposition, without having to specify what that referent was. For example, if EA asks CA "Who is teaching CS360?", we cannot represent CA's belief that EA wants to know the teacher of CS360 as since this representation says that CA believes that EA wants to believe that the teacher is Dr. Smith (but EA, in asking the question, may not be predisposed to any such belief and may in fact reject "Dr. Smith" as the answer to the question). Allen and Perrault addressed this with knowref and knowif predicates, which represented an agent's know­ing the referent of a term in a proposition and knowing whether a proposition is true. Thus CA's belief that EA wants to know the teacher of CS360 in the above example might be represented as

In our multistrength belief model, knowref is treated as being certain about the referent of the term in the specified proposition and knowif is treated as being certain about whether a proposition is true or false.

As the dialogue progresses, the belief model must capture the changing beliefs of the user. When a discourse act has been successful, its goals can be used to update the belief model. For example, if the user explicitly accepts the proposition conveyed by the utterance "Dr. Smith is teaching Architecture" (perhaps by saying "Yes, I'll accept that"), then the system can update its belief model to include the belief that the user himself believes that Dr. Smith is teaching Architecture. However, explicit acceptance is less common than implicit acceptance; Section 4.6 discusses implicit acceptance and its recognition.

Since we do not currently have a response generation component, our system processes the utterances of both participants, alternating between playing the role of CA and the role of EA. Note that this differs from playing the role of a third-party observer—when the system plays the role of EA, the system has access to EA's beliefs (including EA's beliefs about the current dialogue model and EA's beliefs about CA), and when the system plays the role of CA, it has access to CA's beliefs. However, whenever the system assumes the role of a participant and processes a new utterance, it is assumed that this participant has correctly interpreted previous utterances and has a correct model of the preceding dialogue.

4.1.2 Related Work on Modeling Belief. Young (1987) built a model in which the beliefs of the user are part of an explicit, missing, or stereotype module (he used the system's beliefs as a stereotype). Although this system provides needed differ­entiation among beliefs that the system knows the user holds, those that the system has attributed to the user, and those about which the system has no knowledge, this model still does not contain degrees of partial belief that are essential for modeling discourse acts such as expressing doubt. Ballim and Wilks (1991) developed a nested belief model that captures an agent's beliefs about other agents' beliefs. Their system combines belief ascription based on stereotypes with belief ascription based on per­turbations of the system's own beliefs, but they do not represent how strongly a belief is held.

In previous work, we noted the need to differentiate among domain, problem-solving, and discourse actions (Lambert and Carberry 1991; Elzer 1995). In task-oriented con­sultation dialogues, the participants are constructing a plan for achieving some domain goal, such as owning a home, and the resultant plan will consist of domain actions such as applying for a mortgage. In order to construct the domain plan, the partici­pants pursue problem-solving actions such as evaluating alternative domain actions or correcting an action in the partially constructed domain plan. Domain and problem-solving actions have been investigated by many researchers (Allen and Perrault 1980; Perrault and Allen 1980; Wilensky 1981; Litman and Allen 1987; van Beek and Cohen 1986; Ramshaw 1989; Carberry 1990).

Discourse actions are communicative actions that are executed by the dialogue participants in order to obtain or convey the information needed to pursue the problem-solving actions necessary for constructing the domain plan. Examples of very different discourse actions include answering a question, informing, and expressing doubt. Al­though our system models domain, problem-solving, and discourse actions, this paper is only concerned with recognizing discourse acts, particularly complex discourse acts such as expressing doubt.

Our system's knowledge about how to perform actions is contained in a library

6 We would like to thank one of the anonymous reviewers and Ingrid Zukerman for bringing this work to our attention.

#Address-Believability(_agentl, _agent2, .proposition) Effects: believe(.agent2, want(.agentl, believe(_agent2, .proposition, [C:C])),

#Address-Understanding(.agentl, _agent2, .proposition) Effects: told-about(_agentl, _agent2, .proposition)

Recipes for Inform and Tell discourse acts.

either be ignorant about _agent2's belief in .proposition or have some belief (possi­bly certain) that _agent2 is not already certain that .proposition is true, i.e., .agentl does not believe that _agent2 is already convinced that .proposition is true. In de­termining the belief strengths specified in the applicability conditions, we examined the beliefs that an agent must hold and tried to identify the minimum and maximum strength of belief that would make the discourse act reasonable to pursue.

We have divided the effects of an action into two subclasses: 1) the results of correctly performing the action, which are labeled effects, and 2) the desired effects of the action (over which the agent may lack control), which are labeled goals. For example, in the case of domain actions, the effect of applying for graduate study is that one has applied, while the goal is that one be accepted for graduate study. This distinction between effects and goals is particularly important in the case of discourse actions, where the agent cannot be assured that an action will have its intended result.

Variables in recipes are represented as lowercase strings preceded by an under­score, with the string reflecting the variable's type; for example, .coursel and _course2 refer to variables of type course.

In Allen's seminal model of plan recognition (Allen 1979), the bodies of opera­tors could contain either goals to be achieved or action names with parameters. In our system, preconditions are represented as goals to be achieved, while the bodies of recipes specify actions. Since the recipe for each action in our recipe library con­tains a single goal in its goal field, this suffices—the goal makes clear the purpose of the action in a plan. However, in a richer domain where an action could be used to achieve several different goals, it would be necessary to specify the intended goal in the recipe body and chain from it to the desired action in order to capture the moti­vation for performing the action. During plan recognition, our system matches goals

10 For example, one might read a book to gain knowledge or to entertain oneself.

Action: Express-Doubt(_agentl, _agent2, .proposition!, _proposition2)

{jigentl expresses doubt to jxgentl about .propositionl by contending that -proposition! is true}

Appl Cond: believe(_agentl, believe(_agent2, .propositionl, [S:C]), [S:C]) believe(_agentl, _proposition2, [W:S]) believe(_agentl, _proposition2 —» ->.propositionl, [S:C])

Body: Convey-Uncertain-Belief(_agentl, _agent2, _proposition2) Effects: believe(-agent2, believe(.agentl, .propositionl, [Sn:Wn]), [S:C])

believe(.agent2, believe(_agentl,_proposition2 —> -i.propositionl, [S:C]), [S:C]) believe(.agent2, want(.agentl, Resolve-Conflict(_agent2, .agentl,

.propositionl, _proposition2)), [S:C]) Goal: want(.agent2, Resolve-Conflict(.agent2, .agentl, .propositionl, _proposition2))

A recipe for an Express-Doubt discourse act.

Express-Doubt discourse act is that the listener also wants to resolve the conflict. The mutual desire for conflict resolution resulting from a successful Express-Doubt discourse act leads to a negotiation subdialogue (initiated by the Express-Doubt action).

We maintain a structure called a dialogue model that captures the system's beliefs about the existing dialogue context. The discourse level of the dialogue model contains a tree structure called the discourse tree. Each node of the discourse tree represents a discourse or communicative act that has been initiated by one of the dialogue par­ticipants, and the children of a node represent discourse acts that are being pursued in order to perform the parent action. For example, asking and answering a ques­tion (Ask-Ref and Answer-Ref) are part of obtaining information (Obtain-Info-Ref) in the discourse tree in Figure 5. The lowest uncompleted action in the discourse tree is marked as the focus of attention; it represents the first expectation for subsequent utterances. In Figure 5, the focus of attention is the Tell action. The active path consists of the sequence of actions along the path from the action that is the focus of attention to the root node. The actions on the active path provide successive expectations about the role of the next utterance in the dialogue; actions closer to the current focus of attention are regarded as more salient than those further back on the active path.For example, in the discourse tree of Figure 5, the first expectation is that if EA does not understand CA's previous statement, then EA will now choose to address her understanding of it and thereby contribute to the Tell act that is the existing focus of attention. The next expectation is that if EA has any doubt about the proposition conveyed by CA, then EA will choose to address its believability, thereby contributing to the Inform discourse act that is the next action on the active path. (As shown in Figure 3, Address-Understanding is a subaction in the recipe for the Tell discourse act, and Address-Believability is a subaction in the recipe for the Inform discourse act.)

13 By uncompleted, we mean not as yet known to be completed.

14 The active path is a sequence of actions; this work has not considered multithreaded discourse, a topic that Rosé et al. (1995) have begun to investigate.

Sample discourse tree.

Surface-Say-Prop(CA, EA, Teaches(Dr.Smith, CS360)) CA: Dr. Smith is teaching CS360.

Our process model starts with the semantic representation of a new utterance and uses plan inference rules (Allen and Perrault 1980; Carberry 1988) along with constraint satisfaction (Litman and Allen 1987) to hypothesize chains of actions Ai,A2,...,Anthat the speaker might be intending to perform with the utterance. In such a chain, action A, contributes to the performance of its successor action A(+1. For example, the semantic representation of an utterance such as "Dr. Smith is teaching Architecture" is

Surface-Say-Prop(_agentl, _agent2, Teaches(Dr.Smith, Architecture))

A Surface-Say-Prop is a subaction in the recipe for a Tell discourse act, which in turn is a subaction in the recipe for an Inform discourse act. Thus chaining from the surface utterance produces a sequence of hypothesized discourse acts, each of which plays a role in the performance of its successor on the chain.

We have expanded on Litman and Allen's (1987) notion of constraint satisfaction and Allen and Perrault's (1980) use of beliefs. As described earlier, many of the appli­cability conditions in our discourse recipes are beliefs that the agent of the action must hold in order for the action to be felicitous. Our recognition algorithm requires that the system be able to plausibly ascribe these beliefs in hypothesizing a new action; if the belief ascription is implausible or if the constraints of the discourse recipe are not satisfied, the inference is rejected.

As we claimed in Section 3, actions such as the expressions of doubt in utterances (14a) and (14b) (repeated in Figure 6) require evidence for their recognition. Let us further examine why this is the case. Figure 7 illustrates a situation in which we have several discourse acts (with different degrees of salience) that the agent might be expected to pursue. The solid boxes show some of the actions that are part of the (11) CA: Dr. Smith is teaching CS360.

(13) CA: No, Dr. Smith is not on sabbatical.

A sample dialogue with surface negative questions.

I action- 1(EA, CA, PROPA) I laction-2(EA, CA, PROPBH I action-3(EA, CA, PROPCfl

Relating an inference path to the existing dialogue context.

existing dialogue context and the dashed boxes show a chain of actions inferred from the new utterance. As depicted in the figure, the process model can infer a chain of actions starting with some surface speech act surf ace-action(EA, CA, PROPD), up to some other action, action(EA, CA, PROPD), up to some other action, e-action(EA, CA, _propl, PROPD). The action e-act ion contains two propositions. One of these, PROPD, is instantiated by chaining from the ear Her action, action(EA, CA, PROPD). However, the other proposition, _propl, cannot be instantiated by plan chaining; it must be instantiated by unification with a proposition from the existing dialogue context. For example, if e-action is identified as contributing to action-3 in the existing dialogue context, then _propl might be instantiated as PROPC. On the other hand, if e-action is identified as contributing to action-1, then _propl might be instantiated as PROPA. Chaining might suggest three possibilities (e-action could contribute to action-3, or to action-1, or to neither of them), and the relative salience of the propositions in the existing dialogue context is not sufficient to identify this relationship.

As a concrete example, consider again the dialogue segment in Figure 6 consisting of utterances (10)-(13) followed by one of utterances (14a)-(14c). After utterance (13),

15 These are actions on the active path of the dialogue model; the actions that are deepest on the active path are closer to the current focus of attention and are therefore regarded as more salient.

In a collaborative task-oriented dialogue, the participants are working together to construct a plan for accomplishing a task. If the collaboration is to be successful, the participants must agree on the plan being constructed and the actions being taken to construct it. Thus, since a communicated proposition is presumed to be relevant to this plan construction process, the dialogue participants are obligated to communi­cate as soon as possible any discrepancies in belief about such propositions (Walker and Whittaker 1990; Chu-Carroll and Carberry 1995b) and to enter into a negotia­tion subdialogue in which they attempt to "square away" (Joshi 1982) their disparate beliefs.

In our earlier work (Carberry 1985,1989), we claimed that a cooperative participant must accept a response or pursue discourse goals directed toward being able to accept the response. As we noted there, this acceptance need not be explicitly communicated to the other participant; for example, failure to initiate a negotiation subdialogue conveys implicit acceptance of the proposition communicated by an Inform action. This notion of implicit acceptance is similar to an expanded form of Perrault's default rea­soning about the effects of an inform act (Perrault 1990). Our model captures this by recognizing implicit acceptance when an agent foregoes the opportunity to address acceptance of an action and moves on to pursue other discourse actions.

4.6.1 Acceptance Actions. If a statement is intended to answer a question, the listener in a collaborative dialogue must note when he believes that the statement does not suffice as a complete answer. However, doing so implies that the listener believes the statement, since it is inefficient to address a proposition's completeness as an answer if one does not accept the proposition. Similarly, if a listener does not believe a commu­nicated proposition, he must convey this disagreement as soon as possible (Walker and Whittaker 1990). But by questioning the validity of a proposition, the listener conveys that he believes that he understood the utterance. As Clark and Schaefer (1989) note, by passing up the opportunity to ask for a repair, a listener conveys that he has un­derstood an utterance. Thus we hypothesize that listeners convey their acceptance (or lack of acceptance) in a multistage acceptance phase: 1) understanding, 2) believability, 3) completeness.

Acceptance can be communicated explicitly or implicitly. We include actions that address acceptance in the body of six of our discourse recipes. These recipes were selected because they allow us to capture acceptance of a question (the recipes for Ask-Ref and Ask-If), acceptance of the answer to a question (the recipes for Answer-Ref and Answer-If), and acceptance of a statement (the recipes for Inform and Tell). In this research, we have been primarily concerned with one aspect of acceptance: believing the proposition communicated by an Inform action. For example, the actions in the body of the Inform recipe (see Figure 3) are: 1) the speaker (_agentl) tells the listener (_agent2) the proposition that the speaker wants the listener to believe; and 2) the speaker and listener address believability by discussing whatever is necessary in order for the listener and speaker to come to an agreement about this proposition.This second action, and the subactions executed as part of performing it, account for subdialogues that address the believability of the proposition conveyed in the Inform action. Other actions related to acceptance are captured in other discourse recipes. For example, the Tell action has a body containing a Surface-Say-Prop action and an Address-Understanding action; the latter enables both participants to ensure that the utterance has been understood. Similarly, the Answer-Ref action contains an Inform action and an Address-Answer-Acceptability action that ensures that the Inform action is sufficient to answer the question. Further research is needed to model the full range of actions that address acceptance and to recognize utterances resulting from them.

The discourse tree reflects the order of acceptance actions. As discussed above, lack of understanding should be addressed before believability. This is reflected in the discourse tree that results from a statement, such as the one in Figure 5, where the Tell action (whose recipe contains an Address-Understanding action) is a descendant of the Inform action (whose recipe contains an Address-Believability action); in addition, since the statement in Figure 5 is intended to answer a question, the Inform act is a descen-

16 Questions must also be accepted and assimilated into a dialogue. Our model has recently been expanded to address the acceptance of questions (Bartelt 1996), but we are concentrating on statements in this paper.

17 Since our system does not generate responses, we do not model what the speakers need to discuss; however, if a speaker expresses doubt at a proposition by contending that a second proposition is true, then the speaker is introducing this second proposition and its relationship to the first proposition as items that need to be discussed.

Dialogues conveying different implicit acceptance.

dant of the Answer-Ref action (whose recipe contains an Address-Answer-Acceptability action). Since the Tell is the current focus of attention, it must be completed before other actions are pursued. Thus, if the listener believes that the telling has not been successful (i.e., the listener does not fully understand the utterance), then the listener will pursue discourse acts that contribute to its Address-Understanding subaction. Once the Tell has been successfully completed, then attention reverts back to the Inform act. The Inform must be successfully completed before other higher-level acts are pursued further. Thus if the Inform has not been successful (i.e., the listener does not accept the communicated proposition), then the listener will pursue discourse acts that contribute to its Address-Believability subaction.

We have concentrated primarily on recognizing the acceptance and nonaccep-tance of propositions communicated by Inform actions: i.e., modeling negotiation sub-dialogues in which participants do not automatically believe everything that they are told. Others, Allen and Schubert (1991), Clark and Schaefer (1989), Traum and Hinkelman (1992), and Traum (1994) have investigated how understanding and lack of understanding are communicated and can be recognized.

4.6.2 Modeling Implicit Acceptance. Our system models implicit acceptance in col­laborative dialogue as passing up the opportunity to express lack of acceptance. For example, consider the two dialogue variations shown in Figure 8. Figure 5 depicts the discourse tree constructed from utterances (15) and (16) in Figure 8, with the current focus of attention, the Tell action, marked with an asterisk. In attempting to assimilate (17a) into this discourse tree, the system's first expectation is that (17a) will address the understanding of (16) if EA does not understand it (i.e., as part of the Tell ac­tion that is the current focus of attention in Figure 5). The next expectation is that (17a) will relate to the Inform action in Figure 5, by addressing the believability of the proposition conveyed by (16). The system finds that the best interpretation of (17a) is that of expressing doubt at the proposition that Dr. Smith is teaching CS360, thus confirming the secondary expectation that (17a) is addressing the believability of the proposition conveyed by (16). This recognition of (17a) as part of the Inform action in Figure 5 indicates that EA has implicitly indicated understanding, by passing up the opportunity to address understanding in the Tell action that appears at a lower level in the discourse tree and by moving instead to a relevant higher-level action; (17a) is thus (implicitly) conveying that the Tell action has been successful.

Thus, when an utterance contributes to an ancestor of an action A,- and all of A,'s applicability conditions, except those negated by the goal, are still satisfied, then Aj is assumed to have completed successfully; if that were not true, the dialogue participants would have been required to address those actions. When an action

18 By requiring that all applicability conditions, except those negated by the goal, still be satisfied in order for the action to be viewed as successful, we eliminate situations in which the agent of the Inform act

is recognized as successful, the system updates its model of the user's beliefs with the effects and goals of the completed action. For example, in determining whether (17a) in Figure 8 is expressing doubt at (16) (thereby implicitly indicating that (16) has been understood and that the Tell action has therefore been successful), the system tentatively hypothesizes that the effects and goals of the Tell action hold, resulting in the tentative belief that EA believes that CA believes that Dr. Smith is teaching CS360. If the system determines that this Express-Doubt action is the most coherent interpretation of (17a), it attributes the hypothesized beliefs to EA.

Now consider a dialogue in which utterances (15) and (16) in Figure 8 are instead followed by utterance (17b). In this case, the system finds that the best interpretation of (17b) does not contribute to any of the actions in the existing discourse tree; instead (17b) is identified as initiating an entirely new Obtain-Info-Ref action at the discourse level, resulting in a new discourse tree. Since EA has gone on to pursue some other discourse action unrelated to any of the acts that were part of the previous discourse tree, the system recognizes not only EA's understanding of (16) but also EA's implicit acceptance of the proposition conveyed by (16). That is, because the system interprets EA's utterance as foregoing the opportunity to initiate a negotiation subdialogue to address the acceptance of the proposition communicated by the Inform action, the system recognizes that the Inform action has been successful and that EA has implicitly conveyed acceptance of the proposition that Dr. Smith is teaching CS360.

Our recognition algorithm, outlined in Figure 9, assimilates a new utterance into the existing dialogue context and identifies discourse acts that the speaker is pursuing. It proceeds as follows: Start with the semantic representation of the utterance and extract from it two kinds of linguistic information: 1) clue words that might suggest a generic discourse act, and 2) beliefs that are conveyed by the surface form of the utterance. In our implemented system, possible clue words are explicitly noted in the semantic representation of the utterance, and beliefs conveyed by the surface form of an utterance are extracted from the applicability conditions of the recipe for the surface speech act. For example, the surface form of an utterance such as "Isn't Dr. Smith on sabbatical?" conveys that the speaker has a strong but uncertain belief in the queried proposition; this is captured in the applicability conditions of the recipe for a Surface-Neg-YN-Question discourse act (see the appendix).

Next, use plan inference rules to hypothesize sequences of actions A\,A\V.. -,Aid(inference paths) such that A\ is the surface action directly associated with the speak­er's utterance and Aid is an action on the active path in the existing dialogue context. By requiring that an inference path link up with an action that is already part of the existing dialogue context, we are capturing the expectation that the new utterance will contribute to an action that has already been initiated. This corresponds to a focusing heuristic that captures expectations for new utterances in an ongoing dialogue (Carberry 1990). For any inference path, if Ajd is not the focus of attention in the existing dialogue context, then A!d must be an ancestor of the action that is the focus of attention; tentatively hypothesize that each of the actions on the active path between the existing focus of attention (the focus of attention immediately prior to the new utterance) and Aid have completed successfully and use this hypothesis in reasoning

has become convinced by the other participant that the proposition he was trying to convey is really false. In such cases, the applicability condition believe(.agentl, .proposition) of the Inform will no longer be true and thus the Inform act will not be viewed as completing successfully. We have not addressed situations in which the participants cannot resolve their disagreements and agree to disagree.

LE = clue words extracted from semantic representation of utterance

D = dialogue model

B = listener's beliefs

;;Construct inference paths that link up to active path of dialogue model S^-{P, = Ax,Ah,.. .,Aidj I on-active-path^,,.,D)

;;Eliminate inference paths with unsatisfied constraints or implausible applicability conditions For each P, G S Do

Then B, <— B;U {beliefs that Af and all actions on the active path between Af and Aidhave completed successfully} If (BAj)(3Ck) Aj G Pi A is-constraint(Q,A;) A -.C* Then S <— S - P;

Else So <— SoU{P,} ;;So contains inference paths with no e-actions ;;Select inference paths containing actions with the most evidence

If S2 / 0 Then S <— S2 ;;S contains inference paths with multiple evidence

Else If Si / 0 Then S <— Si ;;S contains inference paths with evidence

Else S <— So ;;S contains inference paths with no e-actions ;;Select inference path containing the action closest to current focus of attention If S/0

Else Begin

B <— BU {beliefs that all actions on active path have completed successfully} ;; Assimilate utterance into dialogue model

Mark i4P2 as new focus of attention in D

Pseudocode outlining our recognition algorithm.

The preceding sections have provided the key mechanisms necessary for modeling negotiation subdialogues. Our recipes differentiate between the effects and the goals of a discourse act. Thus, instead of assuming that a communicated proposition will automatically be accepted by the listener, the effect of our Inform action is only that

19 The action at the focus of attention and some of its ancestor actions may have completed successfully, which becomes evident when the participants choose not to address them further. For example, as a result of providing an answer to a question, the active path may include the discourse actions Answer-Ref, Inform, and Tell with the Tell discourse act being the current focus of attention; if the other participant then performs a discourse act that is a subaction of the Answer-Ref but not of the Inform, then he has accepted the proposition conveyed by the Inform and it has completed successfully.

20 Our system currently maintains a list of clue words and discourse acts that each clue word might suggest.

21 Due to length restrictions, we have omitted a part of the algorithm that deals with focusing heuristics that are not needed for the kinds of utterances addressed in this paper; an example of utterances needing the full algorithm is given in Lambert and Carberry (1991).

Our system for recognizing complex discourse acts and handling negotiation subdia­logues has been integrated into the tripartite dialogue model presented in Lambert and Carberry (1991). This dialogue model contains three levels of tree structures, one for each kind of action discussed in Section 4.2 (domain, problem-solving, and discourse) with links among the actions on different levels. At the lowest level, the discourse actions are represented; these actions may contribute to the problem-solving actions at the middle level which, in turn, may contribute to the domain actions at the highest level. Figure 10 illustrates the tripartite dialogue model for a situation in which CA has previously answered a question about the cost of registering for CS180, and then EA asks "When does CS180 meet?" Note that the discourse level in Figure 10 only re­flects the current query about when CS180 meets, since previous queries have already achieved their discourse goals. Since this paper is concerned almost exclusively with the discourse level of the dialogue model, we will not discuss the overall tripartite model further, except to note that the construction of a new discourse tree requires that the system identify its relationship to existing or new actions at the problem-solving and domain levels (Lambert and Carberry 1991).

Our system has been implemented in Common Lisp on a Sun Sparestation and tested in a university advisement domain. Figure 11 lists some of the beliefs included in the system's model of a stereotypical user. In our current implementation, only the clue word but is recognized as linguistic evidence for an Express-Doubt discourse act. In future work, we will expand the clue words taken into account by our system.

Figure 12 contains an extended negotiation dialogue (portions of this dialogue have been given earlier). This dialogue illustrates a number of features that our system can handle. Utterances (18) and (19) establish the initial context in which CA has pro-

22 =>■ is usually implies rather than strict implication. The semantics of this predicate is that there may be a small number of cases where the antecedent is true and the consequent is not. This is similar to a default rule. For example, the On-Campus rule might be viewed as Vi/: on-campus(y) A faculty(y) A M -■on-sabbatical(y) —* -ion-sabbatical(y). However, as with any default rule, there could be exceptions; for example, one might be on sabbatical but have returned to campus to give a colloquium.

„I........

Instantiate-Vars(EA, CA, Attend-Class(EA, _place, _time), Learn-Material(EA, CS 180, _fac))

A simple tripartite dialogue model.

vided information in response to a question from EA. Utterances (20), (23), and (27) illustrate the use of world knowledge in resolving expressions of doubt. Utterance (20) is an expression of doubt that initiates a negotiation subdialogue, indicating that EA has not accepted the proposition communicated by CA. Utterances (21) and (22) attempt to resolve this doubt by stating that Dr. Brown is not teaching Architecture and providing support for this claim. Utterance (23) expresses doubt at the supporting information; utterances (24) and (25) attempt to resolve this doubt; and utterance (26) explicitly accepts the proposition communicated by (25). Thus utterances (23)-(26) constitute a negotiation subdialogue embedded within the negotiation subdialogue of utterances (20)-(27). Although utterance (26) explicitly conveys acceptance of at least the most salient communicated proposition (the proposition that Dr. Brown was giving a University colloquium), there are still several propositions that have not yet been completely accepted by EA and are thus open for rejection. Utterance (27) again expresses doubt at the proposition that Dr. Smith is teaching Architecture, and implic­itly conveys acceptance of the propositions that Dr. Brown is on sabbatical and that

Sabbatical rule: Teachers on sabbatical usually do not teach.

On Campus rule: Teachers on campus usually are not on sabbatical. (Vy faculty(y) A on-campus(y) =>• -ion-sabbatical(y) )

One Course rule: Teachers usually teach only one course a semester.

One Professor rule: Each course usually has only one instructor.

Expertise rule: Teachers usually do not teach courses outside their area of expertise.

A few stereotypical beliefs. (19) CA: Dr. Smith is teaching Architecture.

(20) EA: Isn't Dr. Brown teaching Architecture?

(21) CA: No, Dr. Brown is not teaching Architecture.

(22) Dr. Brown is on sabbatical.

(23) EA: But wasn't Dr. Brown on campus yesterday?

(24) CA: Yes, Dr. Brown was on campus yesterday.

(25) Dr. Brown was giving a University colloquium.

(26) EA: OK.

Extended negotiation subdialogue.

Dr. Brown is not teaching Architecture.

The rest of this section works through the details of how our system processes these utterances, recognizes the discourse acts they are pursuing, and incrementally builds the discourse tree of the dialogue model. In our examples, the system will switch between playing the role of EA and the role of CA. However, when processing an utterance, the system will have access only to the beliefs of the participant whose identity it has assumed (namely, the listener), along with the correct dialogue model at the time the utterance is made.

5.2.1 Utterance (18): Establishing the Initial Context. The system first plays the role of CA (the listener) and must understand EA's utterance of (18). The semantic repre­sentation of (18) is:

Surface-WH-Question(EA, CA, .course, Teaches(Dr.Smith, .course))

The Surface-WH-Question is a subaction in the body of a recipe for a Ref-Request dis­course act; the Ref-Request is a subaction in the recipe for an Ask-Ref discourse act; and

Surface-WH-Question(EA, CA, .course, Teaches(Dr.Smith, .course)) the Ask-Ref is a subaction in the recipe for an Obtain-Info-Ref discourse act. Therefore the following chain of actions is hypothesized:

Obtain-Info-Ref(EA, CA, .course, Teaches(Dr.Smith, .course))

Ask-Ref(EA, CA, .course, Teaches(Dr.Smith, .course))

Ref-Request(EA, CA, .course, Teaches(Dr.Smith, .course))

As each of the above actions is inferred, the system checks that its constraints are satisfied and that its applicability conditions are plausible. Since this is the only chain of actions suggested by plan inferencing on the discourse level, the system recognizes these discourse actions; it then infers problem-solving actions from the discourse ac­tions and, eventually, domain actions from the problem-solving actions. As actions are recognized, the system updates its model of EA's beliefs, wants, and knowledge from the actions' applicability conditions. Figure 13 shows the initial tripartite dialogue model that is produced. Since this paper is primarily concerned with the recognition of actions on the discourse level, the remainder of our figures will only display the discourse level and will omit the problem-solving and domain level actions.

5.2.2 Utterance (19): Answering the Question. The system is now playing the role of EA (listener) and must understand CA's utterance of (19). The semantic representation of (19) is:

Surface-Say-Prop(CA, EA, Teaches(Dr.Smith, Arch))

Chaining suggests that the surface speech act might be part of a Tell action, which might be part of an Inform action since the surface speech act and the Tell act are part of the body of the Tell and Inform acts, respectively. The applicability conditions for all of these actions are plausible.

The system tries to extend the inference chain from the Inform action. An Inform can be part of the recipes for several discourse actions, including Give-Background and Answer-Ref However, these actions are e-actions and, with the exception of Answer-Ref, inference of these e-actions is rejected. For example, Give-Background is an e-action because it relates the proposition in the current utterance to some other proposition, the proposition about which background is being given. The recipe for Give-Background contains a constraint that there be a particular relationship between the proposition in the Inform action in its body and some other proposition conveyed by CA. Since CA has made no previous utterances, there is no other proposition conveyed by CA and thus this constraint cannot be satisfied. Consequently, Give-Background is rejected. A full discussion of the Give-Background action and its recipe can be found in (Lambert 1993).

On the other hand, Answer-Ref (CA, EA, .term, .proposition) can be inferred from Inf orm(CA, EA, Teaches (Dr. Smith, Arch) ) and the system has evidence for its recognition. Answer-Ref is an e-action since the parameters .term and .proposition cannot be instantiated from the Inform action that precedes it on the inference chain.

23 Although CA can provide background information prior to conveying the proposition about which the background is being given, the Give-Background action in these instances will be recognized in assimilating CA's second utterance (the utterance about which the background is being given) (Lambert and Carberry 1991).

D.omaio.Level.

# I Take-Course(EA, .course)

Probjem-Solving.Level.

Instantiate-Vars(EA, CA, Learn-MateriaKEA, .course, Dr. Smith), Take-Course(EA, _course)) # Instantiate-Single-Var(EA, CA, .course, Learn-Material(EA, _course, Dr. Smith), Take-Course(EA, _course))

Discourse. Level.

! Obtain-Info-Ref(EA, CA, _course, Teaches(Dr. Smith, _course))

Surface- WH-Question(EA,CA,_course,Teaches(Dr.Smith,_course))

Tripartite dialogue model for utterance (18).

As discussed in Section 4.5.1, evidence for e-actions may take one of two forms: 1) evidence from world and contextual knowledge and the surface form of the utter­ance indicating that the applicability conditions for a particular e-action are satisfied, and 2) linguistic evidence from clue words suggesting a generic discourse action. In this case, there are no clue words, so any evidence must be from world and con­textual knowledge or the surface form of the utterance. Answer-Ref (CA, EA, .term, .proposition) can be a subaction in the body of Obtain-Info-Ref (EA, CA, .term, .proposition); unifying with the Obtain-Info-Ref action that is part of the existing discourse tree causes the parameters .proposition and .term to be instantiated as Teaches (Dr .Smith, .course) and .course, respectively, in both the Obtain-Info-Ref and Answer-Ref actions.

World and contextual knowledge provide evidence that the applicability condi­tions of Answer-Ref are satisfied with these instantiations. The third applicability con­dition in the Answer-Ref recipe captures the required relationship between the new parameter .proposition and the parameter .propanswer that appears in the Inform discourse act. It indicates that CA must believe that .propanswer (where .propanswer is instantiated from the Inform act as Teaches (Dr .Smith, Arch)) is an instance of the queried proposition, .proposition, with the queried term .term instantiated. Since the system (playing the role of EA in this case) believes that the participants have equivalent knowledge about language and how terms can be instantiated and since the system believes that the two propositions unify, the system has evidence that the third applicability condition is satisfied.

In addition, there is evidence that the other two applicability conditions are satis­fied. With the above instantiations, these applicability conditions become:

Applicability conditions for Answer-Ref:

believe(CA, want(EA,knowref(EA,.course,believe(CA,Teaches(Dr.Smith,.course), believe(CA, ->knowref(EA,.course,believe(CA,Teaches(Dr.Smith,.course),

If utterance 19 is in fact an Answer-Ref that contributes to the Obtain-Info-Ref that is part of the existing dialogue context, then CA has recognized and is responding to the Ask-Ref that is a child of the Obtain-Info-Ref in the existing dialogue context. Consequently, in considering the Answer-Ref interpretation, the system can hypothesize that CA has recognized the Ask-Ref and can tentatively attribute to CA the belief that the Ask-Ref s applicability conditions were satisfied, as shown below.

Beliefs attributed to CA (by virtue of hypothesis that CA has recognized the Ask-Ref):

want(EA, knowref(EA, .course, believe(CA, Teaches(Dr.Smith, .course), -iknowref(EA, .course, believe(CA, Teaches(Dr.Smith, .course), [C:C]))

This is equivalent to tentatively hypothesizing that CA has recognized the intentions communicated by utterance (18) and inferred the discourse level of the dialogue model depicted in Figure 13. Thus the system's model of CA's beliefs (resulting from CA's recognition of the Ask-Ref) provides evidence that the applicability conditions of the Answer-Ref discourse act are satisfied.

Since this inference chain is the only one containing an e-action for which there is evidence, the system recognizes CA's utterance as providing Architecture as the answer to EA's question about what Dr. Smith is teaching and thereby contributing to the Obtain-Info-Ref action initiated by EA. The updated discourse tree is shown in Figure 14, with the new focus of attention marked with an asterisk.

5.2.3 Utterance (20): Initial Expression of Doubt. The system is again playing the role of CA (listener) and must understand EA's utterance of (20). The semantic represen­tation of (20) is:

The surface form of (20) suggests that EA thinks that Dr. Brown is teaching Archi­tecture, but is not certain of it. This belief is captured in the applicability condition

24 This does not mean that the instantiation will result in a true proposition, only that it is a legal instantiation of the term. For example, CS180 is a legal instantiation of the .course term in the proposition Teachesfjones .course) although TeachesQones, CS180) may be false. We have not addressed the problem of misconceptions about class membership.

Discourse tree for first two utterances in Figure 12.

(19) CA: Dr. Smith is teaching Architecture.

of the recipe for a Surface-Neg-YN-Question. Since we assume a noise-free medium and well-formed utterances, surface speech acts always execute successfully and are correctly recognized. Thus, the beliefs captured in the applicability conditions of the surface speech act are immediately entered into the system's model of EA's beliefs. The most salient interpretation of (20), that it is addressing the understanding of (19) and thus contributing to the Tell discourse act that is the current focus of attention in the dialogue, is rejected.

The system can construct an inference path suggesting that the utterance con­tributes to the Inform discourse act that is the parent of the Tell act in the existing discourse tree. In particular, by chaining from subactions to parent actions (actions whose recipes contain the subaction), the system constructs an inference path contain­ing the following chain of actions:

Inform(CA, EA, .propositionl) î

Address-Believability(CA, EA, .propositionl)

Address-Unacceptance(EA, CA, .propositionl, Teaches(Dr.Brown, Arch))

Express-Doubt(EA, CA, .propositionl, Teaches(Dr.Brown, Arch))

Convey-Uncertain-Belief(EA, CA, Teaches(Dr.Brown, Arch))

If the last action on this inference path is unified with the Inform act in the existing discourse tree, then .propositionl in the recipe for Address-Believability will be instan­tiated as Teaches (Dr. Smith, Arch), indicating that EA uttered (20) in order to express

25 In our implemented system, it is rejected because there is no recipe for the Address-Understanding action that is part of the body of the recipe for the Tell discourse act, and thus it is not possible to construct an inference path from the utterance to the Address-Understanding act. In the future, our expanded system will include such recipes, and the interpretation will be rejected because of lack of evidence for the e-action on the inference path or because its constraints are not satisfied.

Figure 17 contains a second negotiation dialogue. Due to space limitations, we will only discuss two interesting features of the dialogue and its processing by our system. Utterance (30) illustrates the use of a linguistic clue word to suggest an expression of doubt. In interpreting utterance (30), the system constructs an inference path contain­ing the action:

Although the system does not have evidence that all of the applicability conditions for this Express-Doubt action are satisfied, the linguistic clue but does provide evidence for the generic Express-Doubt act. Since this is the only inference path containing an e-action for which there is evidence, the system recognizes (30) as expressing doubt at the proposition that CS510 meets on Monday at 7PM by contending that Dr. Jones is teaching CS510. In this case, the system lacked evidence for the third applicability condition in the recipe for Express-Doubt. But having recognized that EA is expressing doubt, it attributes to EA the beliefs captured in the applicability conditions. In partic­ular, the system attributes to EA the belief that Dr. Jones teaching CS510 implies that CS510 would not meet on Monday at 7PM, though it has no idea why EA believes that this implication holds—perhaps EA believes that Dr. Jones has to be home to take care of his children at night.

When utterance (33) occurs, there are three propositions that have not yet been accepted by EA, and the system considers the possibility that EA is performing one of three express doubt actions, namely

In all three cases, the system lacks evidence that the third applicability condition in the Express-Doubt recipe is satisfied. However, the linguistic clue word but provides evidence for a generic Express-Doubt action. Since the system has equivalent evidence for all three of the Express-Doubt acts, contextual knowledge is used to choose among them. Since the proposition Teaches (Dr. Hart, CS510) is closest to the existing focus of attention in the discourse tree, it is the most salient of the three propositions that are open for rejection. Utterance (33) is therefore interpreted as expressing doubt at the proposition that Dr. Hart is teaching CS510 by contending that it is a graduate-level course. Thus contextual knowledge arbitrates when equivalent evidence is available for several specific discourse acts.

We undertook an evaluation of our prototype system both to assess whether it derived appropriate interpretations of utterances and to identify areas for further research. We obtained eight human volunteers, six of whom are not engaged in NLP research and two of whom are involved in unrelated NLP projects. The subjects were given a set of world knowledge stereotypically believed in the domain, such as that faculty on sabbatical do not normally teach. The subjects were presented with a set of dialogues and asked to analyze several utterances from each dialogue. The selected utterances did not include simple questions initiating the dialogue or straightforward answers to questions, since it seemed likely that the subjects would agree with the system's interpretation and thus the results would be biased in favor of the system. The selected utterances did include surface negative questions (both with and without a clue word but), statements interpreted by our system as support for a previous assertion or as explanations about why a proposition was not in conflict with a previous claim, and examples of implicit acceptance.

For each utterance selected for analysis, the subjects were given a suggested inter­pretation, and asked whether the suggested interpretation was reasonable and whether they could identify a better interpretation. For 15 of 20 utterances, the subjects unani­mously believed that the system's interpretation was best. It should be noted there was unanimous agreement that utterance (42) below should be interpreted as an expression of doubt but that utterance (39) should not.

There were two categories of utterances where the subjects disagreed. In the case of surface negative questions that did not express doubt, such as utterance (39) above,

27 The subjects were not told that the suggested interpretation was the one produced by our system but only that we were trying to determine how utterances in a discourse should be interpreted.

Grosz and Sidner (1986) postulated a theory of discourse structure that included lin­guistic, intentional, and attentional components, and they argued that the dominance and satisfaction-precedes relationships between discourse segments must be identi­fied in order to determine discourse structure. They also noted three kinds of infor­mation that contribute to determining the purposes of discourse segments and their relationship to one another: linguistic markers, utterance-level intentions, and general knowledge about actions and objects. Subsequently Lochbaum (1994) developed an algorithm based on Grosz and Sidner's SharedPlan model (Grosz and Sidner 1990) that recognizes discourse segment purposes and discourse structure.

We contend that, in order to understand utterances and respond appropriately, it is necessary not only to determine the structure of the discourse but also to identify the communicative acts that an agent intends to perform with an utterance. For example, if a listener does not recognize when an utterance such as "Wasn't Dr. Smith on campus yesterday?" is expressing doubt, then the listener's response might fail to address the reasons for this doubt. Our research provides a computational algorithm that uses multiple knowledge sources to recognize complex discourse acts, including expressions of doubt, and to identify their relationship to one another. This algorithm and our strategy for recognizing implicit acceptance enable us to model negotiation subdialogues, something that previous systems have been unable to handle.

Several researchers have built argument understanding systems, but none has ad­dressed participants coming to an agreement or mutual belief about a particular situa­tion, either because the researchers investigated monologues only (Cohen 1987; Cohen and Young 1991), or because they assumed that dialogue participants do not change

28 In a dialogue, Grosz and Sidner's discourse segment purpose is intended to capture the purpose of a segment consisting of a series of utterances by both participants, not the communicative intentions underlying each participant's discourse actions.

Several models of discourse have recently been built which view conversation as a kind of collaborative behavior in which speakers try to make themselves understood and listeners work with speakers to help speakers attain this goal.

Clark and Schaefer (1989) contend that utterances must be "grounded," or un­derstood, by both parties, but they do not address conflicts in belief, only lack of understanding. Walker (1992) has found many occasions of redundancy in collabora­tive dialogues, and explains these by claiming that people repeat themselves in order to ensure that each utterance has been understood. Clark and Wilkes-Gibbs (1990) propose a collaborative model of dialogue in which referring is viewed as a collabo­rative process and each conversation unit is viewed as a contribution, which consists of 1) an utterance that performs a referring action, and 2) the utterances required to understand the referent described in the utterance. Heeman (1991) implemented this model in a plan-based collaborative model of dialogue that is able to plan and recognize referring expressions and their corrections.

Other collaborative models assume that two participants are working together to achieve a common goal (Cohen and Levesque 1990a, 1991a, 1991b; Lochbaum, Grosz, and Sidner 1990; Lochbaum 1991; Grosz and Sidner 1990; Searle 1990). Searle (1990) proposes a model in which the two agents working together have a joint intention, a "we intention," instead of individual intentions. Cohen and Levesque (1990a, 1990b, 1990c, 1991a, 1991b) have developed a formal theory in which agents are jointly com­mitted to accomplishing a goal, so both parties have individual intentions to accom­plish the goal as part of their joint commitment. Grosz, Lochbaum, and Sidner (Grosz and Sidner 1990; Lochbaum, Grosz, and Sidner 1990; Lochbaum 1991) have specified a system in which two agents are working to accomplish some common goal by build­ing a "shared plan" in which each agent holds certain beliefs and intentions. These beliefs and intentions indicate that the agents intend to perform some joint action, and that they believe they can perform this action. All of these models indicate the need for modeling collaborative dialogue, but none suggests a system that can handle the

29 Another reason for repetition, she claims, is for centering (Grosz, Joshi, and Weinstein 1995), but she concentrates on repetitions that give evidence of understanding.

kind of negotiation subdialogues that people often engage in when trying to negotiate their conflicts in belief, even when they are both working towards the same goal.

We have presented a plan-based model for handling cooperative negotiation subdia­logues. Our system infers both the communicative actions that people pursue when speaking and the beliefs underlying these actions. Beliefs, and the strength of these beliefs, are recognized from the surface form of utterances and from the explicit and implicit acceptance of previous utterances. Our algorithm for recognizing discourse actions combines linguistic, contextual, and world knowledge in a unified framework. By combining these different knowledge sources, we are able to recognize complex discourse acts such as expressing doubt, to identify the relationship of utterances to one another, and to model negotiation subdialogues. Since negotiation is an integral part of multiagent activity, our process model addresses an important aspect of coop­erative interaction and thus is a step toward an intelligent and robust natural language consultation system.

This work was supported by the National Science Foundation under Grant No. IRI-9122026. The Government has certain rights in this material. We would like to thank Rachel Sacher for her help in our corpus analysis and the anonymous reviewers for their helpful comments on the manuscript.