File Information
File: 05-lr/acl_arc_1_sum/cleansed_text/xml_by_section/metho/02/w02-0215_metho.xml
Size: 25,920 bytes
Last Modified: 2025-10-06 14:07:55
<?xml version="1.0" standalone="yes"?> <Paper uid="W02-0215"> <Title>Issues Under Negotiation</Title> <Section position="3" start_page="0" end_page="0" type="metho"> <SectionTitle> 2 Sidner's theory of negotiative </SectionTitle> <Paragraph position="0"> dialogue As the title says, Sidner's theory is formulated as \an artiflcial discourse language for collaborative negotiation&quot;. This language consists of a set of messages (or message types) with propositional contents (\beliefs&quot;). The efiects of an agent transmitting these messages to another agent is formulated in terms of the \state of communication&quot; after the message has been received. The state of communication includes individual beliefs and intentions, mutual beliefs, and two stacks for Open Beliefs and Rejected Beliefs. Some of the central messages are does not believe belief, which has been ofiered as a proposal 2A more in-depth description of Sidner's account and its relation to the GoDiS model, including a reformulation of Sidner's artiflcial negotiation language in terms of GoDiS information state updates, can be found in (Cooper et al., 2001).</Paragraph> <Paragraph position="1"> Philadelphia, July 2002, pp. 103-112. Association for Computational Linguistics. Proceedings of the Third SIGdial Workshop on Discourse and Dialogue, In addition, there are three kinds of acknowledgement messages, the most important being AcknowledgeReceipt (AR agt1 belief agt2), which may occur after a ProposeForAccept message and results in belief being pushed on the stack for Open Beliefs. Acknowledgement indicates that a previous message from agt2 about belief has been heard; the agents will not hold belief as a mutual belief until an AcceptProposal message has been sent.</Paragraph> <Paragraph position="2"> While we will not give a detailed analysis of the efiects of each of these acts, some observations are important for the purposes of this paper. Speciflcally, a counter-proposal (CO agt1 belief1 agt2 belief2) is analyzed as a composite message consisting of two PFA messages with propositional contents. The flrst proposed proposition is belief2 (the \new&quot; proposal), and the second is (Supports (Not belief1) belief2), i.e. that belief2 supports the negation of belief1 (the \old&quot; proposal). Exactly what is meant by \supports&quot; here is left unspecifled, but perhaps logical entailment is at least a simple kind of support.</Paragraph> <Paragraph position="4"> Sidner's analysis of proposals is only concerned with propositional contents. A Request for action is modelled as a proposal whose content is of the form (Should-Do Agt Action).</Paragraph> <Paragraph position="5"> A question is a proposal for the action to provide certain information. This brings us to our flrst problem with Sidner's account.</Paragraph> </Section> <Section position="4" start_page="0" end_page="0" type="metho"> <SectionTitle> 3 Problem 1: Negotiation vs. </SectionTitle> <Paragraph position="0"> utterance acceptance In Sidner's theory, all dialogue is negotiative in the sense that all utterances (except acceptances, rejections, and acknowledgements) are seen as proposals. This is correct if we consider negotiation as possibly concerning meta-aspects of the dialogue. Since any utterance (content) can be rejected, all utterances can indeed be seen as proposals.</Paragraph> <Paragraph position="1"> (Clark, 1996) provides a \ladder&quot; with four levels of comprehension involved in grounding of natural language utterances in dialogue.</Paragraph> <Paragraph position="2"> 1. A attends to B's utterance 2. A perceives B's utterance 3. A understands B's utterance 4. A accepts or rejects B's utterance So in one sense of \negotiative&quot;, all dialogue is negotiative since assertions (and questions, instructions etc.) can be rejected or accepted. But some dialogues are negotiative in another sense, in that they contain explicitly discussions about difierent solutions to a problem. Negotiation, on this view, is distinct from Clark's level 4. There is thus a stronger sense of negotiation which is not present in all dialogue. A minimum requirement on negotiation in this stronger sense could be that several alternative solutions (answers) to a problem (question or issue) can be discussed and compared before a solution is flnally settled on. Sidner is aware of this aspect of negotiation, and notes that \maintaining more than one open proposal is a common feature of human discourses and negotiations.&quot; What we want to do is to flnd a way of capturing this property independently of grounding and of other aspects of negotiation, and use it as a minimal requirement on any dialogue that is to be regarded as negotiative.</Paragraph> <Paragraph position="3"> On our view, utterances realizing proposalmoves are moves on the same level as other dialogue moves: greetings, questions, answers etc., and can thus be accepted or rejected on this level. Accepting a proposal-move on the grounding level merely means accepting the content of the move as a proposal, i.e. as a potential answer to a question. This is difierent from accepting the proposed alternative as the actual solution to a problem (answer to a question).</Paragraph> <Paragraph position="4"> To give a concrete example of these difierent concepts of negotiativity, we can compare the dialogues in Figures 1 and 2. The type negotiation in 1 concerns acceptance-level grounding of the utterance and its content. By contrast, the type of negotiation in 2 concerns domain-level issues rather than some aspect of grounding. We won't have much to say about grounding-related negotiation in this paper, but see (Lewin et al., 2000) for an account of negotiation related to utterance grounding.</Paragraph> </Section> <Section position="5" start_page="0" end_page="0" type="metho"> <SectionTitle> 4 Problem 2: Alternatives and </SectionTitle> <Paragraph position="0"> counterproposals When analyzing a travel agency dialogue (Sidner, 1994b), the travel agent's successive proposals of ights are seen as counterproposals to his own previous proposals, each modelled as a proposition. The difierence between proposals and counterproposals is that the latter not only make a new proposal but also proposes the proposition that the new proposal con icts with the previous proposal (by supporting the negation of the previous proposal). This can be seen as an attempt by Sidner to establish the connection between the two proposals as somehow concerning the same issue.</Paragraph> <Paragraph position="1"> This analysis is problematic in that it excludes cases where alternatives are not mutually exclusive, which is natural when e.g. booking a ight (since the user presumably only want one ight) but not e.g. when buying a CD (since the user may want to buy more than one). Also, it seems odd to make counterproposals to your own previous proposals, especially since making a proposal commits you to intending the addressee to accept that proposal rather than your previous ones. In many cases (including travel agencies) it seems that the agent may often be quite indifierent to which ight the user selects. Travel agents may often make several proposals in one utterance, e.g. \There is one ight at 7:45 and one at 12:00&quot;, in which case it does not make sense to see \one at 12:00&quot; as a counterproposal as Sidner deflnes them.</Paragraph> <Paragraph position="2"> Of course, one would not want to use the term \counterproposal&quot; in these cases; what we need is some way of proposing alternatives without seeing them as counterproposals. The basic problem seems to be that when several proposals are \on the table&quot; at once, one needs some way of representing the fact that they are not independent of each other. Sidner does this by adding propositions of the form (Supports (Not belief1) belief2) to show that belief1 and belief2 are not independent; however, this proposition not only claims that the propositions are somehow dependent, but also that they are (logically or rhetorically) mutually exclusive. In our view, this indicates a need for a theory of negotiation which makes it possible to represent several alternatives as somehow concerning the same issue, independently of rhetorical or logical relations between the alternatives. Negotiation, in our view, should not in general be seen in terms of proposals and counterproposals, but in terms of proposing and choosing between several alternatives.</Paragraph> <Paragraph position="3"> 5 Negotiation as discussing alternatives In this section, we will attempt to provide a more detailed description of negotiative dialogue. Clearly, negotiation is a type of problem-solving (Di Eugenio et al., 1998). We deflne negotiative dialogue more speciflcally to be dialogue where DPs3 discuss several alternative solutions to a problem (issue) before choosing one of them. In line with our issue-based approach to dialogue management, we propose to model negotiable problems (issues) semantically as questions and alternative solutions as alternative answers to a question.</Paragraph> <Paragraph position="4"> We also propose to keep track of these issues under negotiation and the answers being considered as potential solutions to each issue in the form of a stack (or ordered set) of pairs of issues S> where do you want to go? ask question U> flights to paris on june 13 please answer question S> there is one flight at 07:45 and one at 12:00 propose alternatives, give information about alternatives U> what airline is the 12:00 one ask question S> the 12:00 flight is an SAS flight answer question U> I'll take the 7:45 flight please accept alternative, answer question \which ight?&quot;</Paragraph> <Section position="1" start_page="0" end_page="0" type="sub_section"> <SectionTitle> 5.1 Degrees of negotiativity </SectionTitle> <Paragraph position="0"> Starting from this deflnition, we can distinguish between fully negotiative dialogue and semi-negotiative dialogue. In non-negotiative dialogue, only one alternative can be discussed. In semi-negotiative dialogue, a new alternative can be introduced by revising parameters of the previous alternative; however, previous alternatives are not retained. Finally, in negotiative dialogue: several alternatives can be introduced, and old alternatives are retained and can be returned to.</Paragraph> <Paragraph position="1"> Semi-negotiative information-oriented dialogue does not require keeping track of several alternatives. All that is required is that information is revisable, and that new database queries can be formed from old ones by replacing some piece of information. This property is implemented in a limited way for example in the Swedish railway information system (a variant of the Philips system (Aust et al., 1994)), which after providing information about a trip will ask the user \Do you want an earlier or later train?&quot;. This allows the user to modify the previous query (although in a very limited way) and get information about further alternatives.</Paragraph> <Paragraph position="2"> However, it is not possible to compare the alternatives by asking questions about them; indeed, there is no sign that information about previous alternatives is retained in the system.</Paragraph> </Section> <Section position="2" start_page="0" end_page="0" type="sub_section"> <SectionTitle> 5.2 Factors in uencing negotiation </SectionTitle> <Paragraph position="0"> There are a number of aspects of the dialogue situation which afiect the complexity of negotiative dialogues, and allows further sub-classiflcation of them. This sub-classiflcation allows us to pick out a subspecies of negotiative dialogue to implement.</Paragraph> <Paragraph position="1"> On our deflnition, negotiation does not require con icting goals or interests, and for this reason it may not correspond perfectly to the everyday use of the word \negotiation&quot;. However, we feel it is useful to keep collaborativity (i.e. lack of con icting goals) as a separate dimension from negotiation. Also, it is common practice in other flelds dealing with negotiation (e.g. game theory, economy) to include collaborative negotiation (cf. (Lewin et al., 2000)). A second factor in uencing negotiation is the distribution of information between DPs. In some activities, information may be symmetrically distributed, i.e. DPs have roughly the same kind of information, and also the same kind of information needs (questions they want answered). This is the case e.g. in the Coconut (Di Eugenio et al., 1998) dialogues where DPs each have an amount of money and they have to decide jointly on a number of furniture items to purchase. In other activities, such as a travel agency, the information and information needs of the DPs is asymmetrically distributed.</Paragraph> <Paragraph position="2"> The customer has access to information about her destination, approximate time of travel etc., and wants to know e.g. exact ight times and prices. The travel agent has access to a database of ight information, but needs to know when the customer wants to leave, where she wants to travel, etc.</Paragraph> <Paragraph position="3"> A third variable is whether DPs must commit jointly (as in e.g. the Coconut dialogues) or one DP can make the commitment by herself (as e.g.</Paragraph> <Paragraph position="4"> in ight booking). In the latter case, the acceptance of one of the alternatives can be modelled as an answer to an issue on IUN by the DP responsible for the commitment, without the need for an explicit agreement from the other DP. In the former case, a similar analysis is possible, but here it is more likely that an explicit expression of agreement is needed from both DPs.</Paragraph> <Paragraph position="5"> This variable may perhaps be referred to as \distribution of decision rights&quot;.</Paragraph> <Paragraph position="6"> Travel agency dialogue, and dialogue in other domains with clear difierences in information and decision-right distribution between roles, has the advantage of making dialogue move interpretation easier since the presence of a certain bits of information in an utterance together with knowledge about the role of the speaker and the role-related information distribution often can be used to determine dialogue move type. For example, an utterance containing the phrase \to Paris&quot; spoken by a customer in a travel agency is likely to be intended to provide information about the customer's desired destination.</Paragraph> </Section> </Section> <Section position="6" start_page="0" end_page="0" type="metho"> <SectionTitle> 6 Issues Under Negotiation (IUN) </SectionTitle> <Paragraph position="0"> In this section we discuss the notion of Issues Under Negotiation represented by questions, and how proposals relate to issues on IUN. We also discuss how this approach difiers from Sidner's.</Paragraph> <Section position="1" start_page="0" end_page="0" type="sub_section"> <SectionTitle> 6.1 Negotiable issues and activity </SectionTitle> <Paragraph position="0"> Which issues are negotiable depends on the activity. For example, it is usually not the case that the name of a DP is a negotiable issue; this is why it would perhaps seem counterintuitive to view an introduction (\Hi, my name is NN&quot;) as a proposal. However, it cannot be ruled out that there is some activity where even this may become a matter of negotiation. Also, it is usually possible in principle to make any issue into a negotiable issue, e.g. by raising doubts about a previous answer.</Paragraph> <Paragraph position="1"> However, for our current purposes we may make a distinction between negotiable and nonnegotiable issues in an activity. The advantage of this is that the more complex processing and domain-speciflc knowledge required for negotiable issues are only required for issues which the system needs to be able to negotiate. The drawback, of course, is that the system becomes less exible; however, there is always the possibility of deflning all issues as negotiative if one so desires.</Paragraph> </Section> <Section position="2" start_page="0" end_page="0" type="sub_section"> <SectionTitle> 6.2 Alternatives as answers to issues on IUN </SectionTitle> <Paragraph position="0"> Given that we analyze Issues Under Negotiation as questions, it is a natural move to analyze the alternative solutions to this issue as potential answers. On this view, a proposal has the efiect of adding an alternative answer to the set of alternative answers to an issue on IUN. An answer to the question on IUN is equivalent to accepting one of the potential answers as the actual answer. That is, a question on IUN is resolved when an alternative answer is accepted.</Paragraph> <Paragraph position="1"> Here we see how our concept of acceptance difiers from Sidner. On our view a proposal can be accepted in two difierent ways: as a proposal, or as the answer to an issue on IUN. Accepting a proposal move as adding an alternative corresponds to meta-level acceptance. However, accepting an alternative as the answer to an issue on IUN is difierent from accepting an utterance.</Paragraph> <Paragraph position="2"> Given the optimistic approach to acceptance, all proposals will be assumed to be accepted as proposals; however, it takes an answer-move to get the proposed alternative accepted as the solution to a problem.</Paragraph> </Section> </Section> <Section position="7" start_page="0" end_page="75" type="metho"> <SectionTitle> 7 Adding IUN to the GoDiS </SectionTitle> <Paragraph position="0"> information state The ideas described in this paper are currently being implemented in GoDiS (Bohlin et al., 1999), an experimental dialogue system initially adapted for the travel agency domain but later adapted for several other domains. GoDiS is implemented using the TrindiKit(Larsson and Traum, 2000; Larsson et al., 2000), a toolkit for experimenting with information states and dialogue move engines and for building dialogue systems.</Paragraph> <Paragraph position="1"> The notion of information state used by GoDiS is basically a version of the dialogue game board which has been proposed by (Ginzburg, 1998). We represent information states of a dialogue participant as a record of the type shown in Figure 3.</Paragraph> <Paragraph position="2"> The main division in the information state is between information which is private to the agent and that which is shared between the dialogue participants. The private part of the information state contains a plan fleld holding a dialogue plan, i.e. is a list of dialogue actions that the agent wishes to carry out. The agenda fleld, on the other hand, contains the short term goals or obligations that the agent has, i.e. what the agent is going to do next.</Paragraph> <Paragraph position="3"> We have included a fleld tmp that mirrors the shared flelds. This fleld keeps track of shared information that has not yet been conflrmed as grounded, i.e. as having been understood by the other dialogue participant. The shared fleld is divided into a local and a global part. The local part contains information about the latest utterance, and information which may be relevant for interpreting it. The flrst subfleld is for a stack of questions under discussion (qud). These are questions that can currently be answered elliptically. The lu fleld contains information about the speaker of, and the moves performed in, latest utterance.</Paragraph> <Paragraph position="4"> The global part contains shared information which re ects the global state of the dialogue. It contains a set of propositions (commitments) which the agent assumes have been jointly committed to in the dialogue (com).</Paragraph> <Paragraph position="5"> In order to include Issues Under Negotiation and alternative answers to issues on IUN in the information state, we have also added a new information state fleld of type OpenStack(Pair(Question,Set(Answer)))4. null We deflne update rules for updating the information state based on the recognized move(s). The rules are deflned in terms of preconditions and efiects on the information state; the efiects are a list of operations to be executed if the pre-conditions are true.</Paragraph> <Paragraph position="6"> Regarding the semantics of questions, propositions, and short answers, we use a simplifled version of flrst order logic with the addition of lambda abstracts for representing questions. Questions and answers can be combined to form propositions. For example, the content of \when do you want to leave?&quot; can be represented as ?x.desired dept time(x), the answer \twelve thirty&quot; as 12:30, and the proposition resulting from combining the two desired dept time(12:30). As a further example, the proposition that a certain ight (denoted f1) departs at 7:45 is represented as dept time(f1,0745). For a more comprehensive description of the semantics used, see (Larsson, 2002).</Paragraph> <Paragraph position="7"> 4The choice of a stack is motivated by the fact that several issues may, in principle, be under negotiation at once, and that some issues may be subordinate to others. An open stack is a stack where non-topmost elements are accessible for inspection and deletion.</Paragraph> </Section> <Section position="8" start_page="75" end_page="75" type="metho"> <SectionTitle> 8 An example </SectionTitle> <Paragraph position="0"> In the (invented) example in 4, the question on IUN is ?x:desired ight(x), i.e. \Which ight does the user want?&quot;. The user supplies information about her desired destination and departure date; this utterance is interpreted as a set of answer-moves by the system since it provides answers to questions that the system has asked or was going to ask. As a response to this, GoDiS performs a database search which returns two ights f1 and f2 matching the speciflcation, and stores the database results in private.bel. The system then proposes these ights as answers to the IUN. The system also supplies some information about them. As a result, the IUN is now associated with two alternative answers, f1 and f2. Finally, the user provides an answer to the question on IUN, thereby accepting one of these alternatives as the ight she wants to take.</Paragraph> <Paragraph position="1"> Because of space limitations, this dialogue does not include any discussion or comparison of alternatives, but it could easily be extended to cover e.g. the dialogue in 2.</Paragraph> <Paragraph position="2"> In a travel agency setting, it can be argued that the informational distribution and decision rights associated with the roles of customer and clerk in a travel agency are su-cient to distinguish proposals for acceptances, but in a more complex setting the move recognition will require more information about surface form and dialogue context.</Paragraph> <Paragraph position="3"> 9 Relation to previous work Much work on negotiative dialogue, including that of Sidner, is based on variations on the BDI model of classic AI and uses generalised planning and plan recognition as an integral part of dialogue processing (e.g. (Cohen and Levesque, 1991), (Grosz and Kraus, 1993) (Chu-Carroll and Carberry, 1994)). The risk with this kind of very general reasoning is that it may become computationally complex and expensive. Therefore, we believe it is useful to start with a simple theory involving reasoning speciflc to a certain kind of dialogue and see how far that takes us, and at what point general planning and reasoning is really necessary.</Paragraph> <Paragraph position="4"> In working on GoDiS our strategy has been to start from a basic issue-based approach and gradually extending it, while trying to keep things as simple as possible. We believe that the issue-based approach can be extended with, and is compatible with, planning and general reasoning mechanisms. This is also in line with the idea behind the information state update approach as implemented in TrindiKit, i.e. to allow for experimentation with difierent kinds of information states and information state update strategies at various levels of complexity (rather than being conflned to choosing between oversimplifled flnite state / frame-based approaches on the one hand, or complex plan-based approaches on the other).</Paragraph> <Paragraph position="5"> Also, most of the plan-based dialogue research is focused on relatively complex dialogues involving collaborative planning. The model presented here is not committed to the view that negotiation only takes place in the context of collaborative planning. In the sense of negotiative dialogue used here, i.e. dialogue involving sev- null eral alternative solutions to some problem, negotiation may also concern matters of fact. This can be useful e.g. in tutorial dialogue where a tutor asks a question, gives some alternative answers, and the student's task is to reason about the difierent alternatives and decide on one of them. In the travel agency domain, it is often not necessary to explicitly represent e.g. the fact that the deciding on a ight is a precondition of a general plan for traveling; instead, we can represent it simply as a fact concerning which ight the user wants to take.</Paragraph> <Paragraph position="6"> (RosPe et al., 1995) provide an account of discourse structure for dialogues involving several alternatives (or \threads&quot;), which appears to be compatible with the view presented here. However, the focus on discourse structures rather than information states and dialogue processing (in terms of information state updates) makes this work less relevant to dialogue systems design, at least from the perspective of an information state update approach.</Paragraph> </Section> class="xml-element"></Paper>