Reasoning and communicative strategies in a model of argument-based negotiation

We are developing a model of negotiation in a natural language where the communicative goal of the initiator is to achieve the decision of the communication partner about doing a certain action. In order to make a decision, the partner starts a reasoning process, checking the existence of the needed resources as well as the positive and negative aspects of doing the proposed action. If the partner does not make the expected decision then arguments for and against of doing the action will be presented by the participants in the following communication. We discuss involvement of a reasoning model as well as communicative strategies and communicative tactics in the negotiation model. We consider how the partners can influence each other when reasoning in order to achieve their communicative goals. A limited version of the model of negotiation is implemented on the computer.


Introduction
Negotiation is a strategic discussion that resolves an issue in a way that both parties find acceptable (Negotiation, 2018). Each party tries to persuade the other to agree with his or her point of view. When making a proposal or assertion, the speaker has to be prepared to receive critiques or counterproposals and react to them (Rahwan et al., 2004). Both parties try to maximize some personal utility in the face of partially conflicting interests, while striving to reach an agreement (Rosenfeld & Kraus, 2016a). Argumentation-based negotiation is the process of decision-making through the exchange of arguments (Thimm, 2014).
Many researchers have been modelling negotiation and persuasion on the computer. Yuan, Svansson, Moore, and Grierson (2007) present a computer game for abstract argumentation and strategies for a software agent to act as a game player. The game enables two human and/or computational agents to exchange arguments, and this provides a basis for extending the game for use in argumentative agent systems. Hadjinikolis, Siantos, Modgil, Black, and McBurney (2013) consider opponent modelling in persuasion dialogs based on an agent's experience obtained through dialogs. Thimm (2014) provides a review of strategies in multi-agent argumentation. Hunter (2015) investigates a probabilistic user model, including how the system updates the model at each step of the dialog, how it uses the model to choose moves, and how it can query the user to improve the model. He claims that strategies for persuasion, in particular taking into account beliefs of the opponent are underdeveloped. Kang, Tan, and Miao (2015) introduce a framework for different persuasion strategies and present a model for adaptive persuasion for virtual agents. The agent is able to change the others' attitudes and behaviours intentionally, interpret individual differences between users, and adapt to user's behaviour for effective persuasion. Rosenfeld and Kraus (2016a) present a methodology for persuading people through argumentative dialogs by combining theoretical argumentation modelling, machine learning, and Markovian optimization techniques.
We study argument-based negotiations in a natural language between two participants where the initiator A makes a proposal to the partner B to do (or prevent doing) an action D and argues for positive outcomes of doing (respectively, not doing) D by B. The communicative goal of the partner B can be the same or opposite. In the last case, B has to rebut A's arguments and to present her own counterarguments. Even if having the same goal with A, there can be obstacles which prohibit or constrain B to do D. In both cases, A has to response to B's counterarguments.
If B refuses to accept A's communicative goal, then the participants are involved in dispute. Both parties present their arguments and finally, whether A wins dispute, i.e. achieves B's decision that corresponds to A's communicative goal, or A loses, i.e. has to withdraw.
We have worked out and implemented on the computer a formal model of argumentation dialog (Koit, 2015(Koit, , 2016Koit & Õim, 2014). In the current paper, we will further develop the model and concentrate on the reasoning of the participants who apply communicative strategies and tactics for achieving their communicative goals. The paper is an extended version of Koit (2017) presented at the 2017 IEEE International Conference on INnovations in Intelligent SysTems and Applications (INISTA 2017). In addition to the conference paper, implementation of the model is considered as moving from one information state to another. The structure of information states is explained in more detail.
The paper has the following structure. In Section 2, we introduce our model of negotiation which includes reasoning about doing an action. We introduce communicative strategies and tactics of participants and consider how the beliefs of the participants about positive and negative aspects of the action are changing in interaction. Section 3 considers the implementation of the model on the computer and Section 4 discusses how human-like negotiations will be produced with the model. In Section 5, we draw conclusions.

A model of negotiation
Let us consider a dialog between two participants (human or artificial agents) in a natural language about doing an action. Let the communicative goal of the initiator A be to convince his partner B to decide to do (or, alternatively, not to do) an action D. In communication, A uses a partner model which evaluates B's resources, positive and negative aspects of doing D and motivates A to believe that B will accept A's communicative goal. A starts the dialog by making a proposal to B. The partner B has her own modelevaluations of the aspects of D which can be different as compared with A's partner model. After the proposal is made, B starts a reasoning procedure in her mind taking into account her evaluations of D and finally, she comes to a decision. If B accepts A's goal then the dialog can finish. If B does not accept A's goal then A must correct his partner modelbecause it did not correspond to the reality. Depending on B's counterarguments, A has to find out new arguments in order to bring the negotiation to the desired end. Therefore, reasoning plays an important role in interaction and it should be taken into account in a dialog model.

Reasoning model
2.1.1. Determinants of human reasoning When aiming at a certain goal in communication, the subject must know how to direct the functioning of the partner's psychological mechanisms in order to bring about the intended result in the partner. When one attempts to change a person's attitude through communication then he/she might use, e.g. the Elaboration Likelihood Model (ELM) (Cacioppo, Petty, Kao, & Rodriguez, 1986)a theory of thinking process. Alternative models are, for example, Social Judgment Theory which emphasizes the distance in opinions and Social Impact Theory which emphasizes the number, strength, and immediacy of the people trying to influence a person to change its mind, etc.
We are using a naïve, 'folk' theory (D'Andrade, 1987;Davies & Stone, 1995). According to the naïve theory of reasoning, there are three kinds of determinants which can cause humans to reason about doing an action (Õim, 1996). The determinants can be either internal or external. The internal determinants are the wishes of the subject related to the action D (we call them WISH-determinants) and his/her considerations that it would be needed, reasonable, necessary to do D in the given situation (we call them NEEDEDdeterminants). WISH-determinants get activated when the subject finds that the action itself or some of its consequences would be pleasurable to him/her; D is evaluated on the scale 'pleasant-unpleasant'. The corresponding scale of NEEDED-determinants is 'useful-harmful'.
The third class of determinants -MUST-determinantsincludes those which originate from outside the subject (thus, the external determinants) and which force him/her to do (or withhold from doing) D independent on his/her own likings: obligations, prohibitions, norms, orders, requests, etc. They operate through the idea of punishmentan action as a reaction to subject's not fulfilling orders or prohibitions.
When reasoning about doing D, the 'weights' of its pleasant and unpleasant, useful and harmful aspects, and also the 'weight' of the possibility to get punished, should be put together and the general balance of the 'weights' of negative and positive aspects should be computed. This suggests that the corresponding scales should be represented in some form which makes the cross-scale comparison possible (e.g. in a numeric form where the use of certain numeric values should be empirically grounded).
The process of reasoning which leads to a decision can be described as consisting of certain steps where the contents of the steps constitute weighting of different aspects of the action D. Depending on the result of a certain step, the reasoner moves further to other aspects of D.
The reasoning itself depends on the determinant which triggers it (respectively, WISH, NEEDED, or MUST) and it ends with a decision to do D or not. Thus, we can describe three different prototypical 'reasoning procedures' with WISH-, NEEDED-, or MUST-determinants as their inputs.

Beliefs and reasoning procedures
Our reasoning model consists of two parts: (1) a model of human motivational sphere which includes beliefs of a reasoning subject about different aspects of doing the action under consideration and (2) reasoning procedures (Koit & Õim, 2014).
Let us represent the model of motivational sphere of a subject by a vector of weights of different aspects of the action (with numerical values of its components): w = (w(resources), w(pleasant), w(unpleasant), w(useful), w(harmful), w(obligatory), w(prohibited), w(punishment-do), w(punishment-not)).
In the description, w(resources) = 1 if the subject has all the resources necessary to do D (otherwise 0); w(obligatory) = 1 if D is obligatory for the reasoning subject (otherwise 0); w (prohibited) = 1 if D is prohibited (otherwise 0). The values of other weights can be nonnegative natural numbers. Here, w(pleasant), w(unpleasant), etc. mean the weight of pleasant, unpleasant, etc. aspects of D; w(punishment-do)the weight of punishment for doing D if it is prohibited, and w(punishment-not)the weight of punishment for not doing D if it is obligatory. Still, when reasoning people do not operate with numbers but they rather use words in a natural language for characterizing the aspects of an action (e.g. useful, reasonable, essential, vital, wise, unwise, unreasonable, doubtful, thoughtless, harmful on the usefulness-harmfulness scale). Instead, we use numbers in our model in order to make summarization possible.
The second part of the reasoning model consists of reasoning procedures that supposedly regulate human action-oriented reasoning. A procedure uses a model w of motivational sphere of the reasoning subject, i.e. his/her evaluations of the aspects of the action D. As said before, there are three different reasoning procedures. Accordingly, a reasoning procedure can be triggered by the WISH-determinant if the pleasant aspects of D overweight the unpleasant ones (i.e. w(pleasant) > w(unpleasant)); by the NEEDED-determinant if the useful aspects of D overweight the harmful ones (i.e. w(useful) > w(harmful)), and by the MUST-determinant if doing D is obligatory for the subject and not doing implies some punishment (i.e. w(obligatory) = 1, and w(punishment-not) > 0). Every reasoning procedure consists of steps that the subject goes through in his/her reasoning process; in every step weights of different aspects of D are summarized and compared, and the output is the decision: to do D or not.
As an example, let us present a reasoning procedure triggered by the NEEDED-determinant as the following step-form algorithm (where the subject believes that doing D is more useful than harmful).

Prerequisite: w(useful) > w(harmful)
(1) Are there enough resources for doing D? If not then do not do D.
(3) Is D prohibited? If not then do D.
(4) Is w(pleasant) + w(useful) > w(unpleasant) + w(harmful) + w(punishment)? If yes then do D. Otherwise do not do D. (5) Is D obligatory? If not then do not do D.
We use two different vectors of weights in our dialog model: w B is the model of B herself which includes B's (actual) evaluations of D's aspects and is used by B, and w AB is the partner model which includes A's beliefs concerning B's evaluations and is used by A. Let us suppose that A has some preliminary information about B in order to compose the initial partner model. Still, both the models w AB and w B will change after A and B entered into negotiation.

Communicative strategies and tactics
Let us define a communicative strategy as an algorithm used by a participant for achieving his/her goal in the interaction (Koit, 2015). Input of such an algorithm is a communicative goal of the participant; and output is the goal achieved, postponed, or not achieved.
The initiator A can realize his communicative strategy in different ways: he can entice, persuade, or threaten the partner B to do (respectively, not to do) D. We call these ways of realization of a communicative strategy communicative tactics. If A's communicative goal is 'B will do D' then by enticing, A tries to trigger B's reasoning by the WISH-determinant (i.e. he tries to increase the pleasantness of D for B as compared with its unpleasantness). Respectively, when persuading, A tries to trigger B's reasoning by the NEEDED-determinant (to increase the usefulness of D as compared with its harmfulness) and when threateningby the MUST-determinant (to increase the punishment for not doing an obligatory action D). We call the affected aspect (respectively, pleasantness, usefulness, or punishment) the title aspect of the communicative tactics. By choosing the tactics, A believes that B's reasoning triggered by this determinant will give a positive decision in his partner model. Still, the participants can change their communicative tactics during negotiation.
A strategy is the overall plan to achieve one's goals in a negotiation based on situation and resources available; tactics are short-term, actions designed to enact or pursue broad strategies (Lewicki, Barry, & Saunders, 2010).
As an example, let us present a communicative strategy of A for achieving the decision of the partner B to do the action D as the following algorithm, i.e. G A = 'B decides to do D' (here '--' starts a comment).
Choose communicative tactics with title aspect t Implement the tactics to generate a proposal to B REPEAT Analyze B's utterance IF B indicated the missing resources THEN choose and present a (counter) argument to demonstrate that the resources exist ELSE --(1) choose a (counter) argument if B indicated some other aspect of D CASE B's utterance OF pleasantness: increase pleasantness unpleasantness: decrease unpleasantness usefulness: increase usefulness harmfulness: decrease harmfulness punishment for not doing an obligatory D: increase punishment END CASE -- (2) Choose an argument to support t Present the chosen argument(s) to B (A can optionally present both (1) and (2), OR only (2) A is supposed to have a set of different utterances (arguments) in order to increase or decrease the values of the weights in the vector of motivational sphere (i.e. in his partner model w AB ). For example, if A's communicative goal is 'B decides to do D' then A can use the arguments for increasing the weights of resources, of the pleasantness, usefulness and punishment for not doing D (if D is obligatory). Similarly, A can use the arguments of decreasing the weights of the unpleasantness, harmfulness, and punishment for doing a prohibited D. We suppose (in our implementation) that all the arguments are 'equal'that they have the same numerical value ('utility') 1, i.e. every argument increases or, respectively, decreases the corresponding weight exactly by 1. In addition, we suppose that every argument can be used only once by the participant and after the communicative tactics have been changed there is not allowed to return to the previous tactics. Still, in reality, arguments can have different utilities, i.e. one argument 'weighs' more (is more influential) than another and using one argument or communicative tactics repeatedly is not excluded.
When following the communicative strategy, A rebuts B's counterargument (if it was presented) and adds another argument to support the title aspect of the current communicative tactics. For example, if B pointed to a small pleasantness of D then A presents an argument for increasing the pleasantness (case (1) in the algorithm) and in addition, when following the communicative tactics of persuasion, A presents another argument to increase the value of the title aspect of the tacticsthe usefulness (case (2) in the algorithm). At the same time, he introduces the needed changes into the partner model w AB , increasing the value of the usefulness (by 1). The value of the pleasantness remains the same: it has been decreased as a result of B's counterargument and then increased as a result of A's own argument (in both cases by 1). Still, A can also avoid rebutting of B's counterargument (case (2) in the algorithm). Then he by default agrees with the counterargument.
The partner B uses a similar communicative strategy which can be presented as the following algorithm (in the case if B's communicative goal is G B = 'do not do D' and A's goal is G A = 'B decides to do D').
Choose an input determinant (WISH, NEEDED, or MUST) which determines a reasoning procedure REPEAT Analyze A's utterance --update w B depending on A's utterance CASE A's utterance OF resources: increase resources pleasantness: increase pleasantness unpleasantness: decrease unpleasantness usefulness: increase usefulness harmfulness: decrease harmfulness punishment for not doing an obligatory D: increase punishment END CASE Change the reasoning procedure? If yes then choose a new procedure Run the reasoning procedure in w B Choose and present a new (counter) argument depending on the result of the reasoning procedure, OR refuse to do D (without presenting any argument) UNTIL B has used all the reasoning procedures AND B has used all the arguments AND B's current reasoning gives decision which corresponds to G A in the current w B (B did not achieve G B and A achieved G A ), OR A abandoned G A (B achieved G B and A did not achieve G A ), OR B decides to postpone the decision (G B not yet achieved by B and G A not yet achieved by A).
A difference as compared with A's strategy is that B does not have initiative. Similarly with A, the partner B is also supposed to have several utterances (arguments) for affecting the values of the different aspects of D which she can use in order to rebut (or support) A's arguments. Here we assume as before that every argument will increase or, respectively, decrease a value of the weight of an aspect in w B by 1 and every argument and every reasoning procedure can be used only once.
For example, if A presented an argument for increasing the pleasantness of D, then B increases the corresponding value in her model w B (by 1), and applies her current reasoning procedure for making a decision. If the decision does not correspond to A's goal (do D) then B can present her counterargument (or she can choose a refusing utterance without any argument).

Implementation
A limited version of our dialog model is implemented as a simple dialog system (DS) which interacts with a user in written Estonian. Some limitations have already been introduced in Section 2.2. The computer can optionally play A's or B's role and the user, respectively, will play B's or A's role. A's communicative goal is 'B will do D', and B's goal is opposite -'do not do D'. Therefore, A and B are involved in dispute. The participants only present their arguments and counterarguments, asking and answering questions is not allowed. In the implementation, the computer has ready-made sentences (assertions) for expressing arguments, that is, for stressing or downgrading the values of different aspects of the proposed action, which depend on its user model and on the action under consideration. The user when playing B's role can optionally put in free texts or choose ready-made sentences but she can only use ready-made sentences when playing A's role.
The initiator of dialog is A. For example, when the computer is playing A's role, then it determines a partner model w AB , fixes its communicative strategy and the communicative tactics which it will follow, that is, the computer respectively determines a reasoning procedure which it will try to trigger in B's (user's) mind. Then, A applies the reasoning procedure in its partner model, in order to 'put itself' into B's role and to choose suitable arguments when convincing B to make a decision to do D. Supposedly, the models w B and w AB are different when a dialog starts but they are approaching each to another during negotiation, as influenced by the presented arguments and counterarguments. Still, the user B is not obliged (but can) to follow neither certain communicative tactics nor reasoning procedures. She is also not obliged to fix her beliefs in relation to D by composing a model w B . Instead, the user may reason and act as 'a normal human'. However, A does not exactly 'know' B's evaluations (the values of the coordinates of the vector w B ), it can only choose its arguments on the basis of B's counterarguments which are the single signals about B's reasoning. Respectively, A is making changes in its partner model w AB during a dialog.
When playing B's role, the computer determines the model w B of itself which includes its beliefs about the different aspects of the action under consideration and chooses the communicative tactics it will follow in order to achieve its communicative goal ('do not do D' in our implementation). The user when playing A's role is not obliged to compile neither the partner model nor the model of herself, but she behaves as 'a normal human' when interacting with the computer.
Information-state dialog manager is used in the implementation (Traum & Larsson, 2003). The programming language is Java.

Representation of information states
There are two parts of an information stateprivate (information accessible only for one participant) and shared (information accessible for both participants). For example, when the computer is playing the role of B (having the communicative goal 'do not do D'), then the private part of an information state includes the following information:

Update rules
The functions of the dialog manager can be formalized in terms of information state update which is changing during the interaction. There are different categories of update rules which will be used by a participant for moving from the current information state to the next one (Traum & Larsson, 2003). For example, there are the following rules for the computer when it plays B's role. Category I. Rules used by B in order to generate its turns in the following cases: (1) if the current reasoning procedure gives the result 'do not do D' in the model w B and B will present its counterargument (2) if the current reasoning procedure gives the result 'do not do D' in the model w B but there are no unused utterances to present a counterargument (3) if the current reasoning procedure gives the result 'do D' in the model w B and a new procedure can be chosen (4) if the current reasoning procedure gives the result 'do D' in the model w B and there are no unused reasoning procedures remained (5) if A abandoned its communicative goal (and therefore B achieved its goal).
Category II includes the rules for B to analyse A's turns.

Reasoning in interaction: an example
The following example (originally in Estonian)an interaction with our implemented DSdemonstrates in more detail how the beliefs of both participants (i.e. the vectors w AB and w B ) are changing in interaction.
Let us suppose that the computer plays A's and the user B's role. The action D is 'to become a vegetarian'. A's goal G A is to achieve B's decision to do D. Let us suppose that A has generated the following partner model: w AB = (w AB (resources) = 1, w AB (pleasant) = 5, w AB (unpleasant) = 3, w AB (useful) = 5, w AB (harmful) = 2, w AB (obligatory) = 0, w AB (prohibited) = 0, w AB (punishment-do) = 0, w AB (punishment-not) = 0).
The computer A when following its communicative strategy (Section 2.2) has chosen the tactics of persuasion (i.e. it has fixed the input determinant NEEDED which it tries to activate in B in order to direct B's reasoning to a positive decision, by increasing the usefulness of the proposed action).
The reasoning procedure triggered by the NEEDED-determinant yields a positive decision (do D) in the partner model w AB because w AB (useful) > w AB (harmful), w AB (resources) = 1, w AB (pleasant) > w AB (unpleasant), w AB (prohibited) = 0 (cf. Section 2.1.2), therefore, A expects that B's decision will be positive. A makes a proposal to B (turn 1) but B does not agree (turn 2): (1) A: You should agree to become a vegetarian.
(2) B: I cannot abandon meat.
--Refusal: B (the user) refuses to do D because she does not have enough resources (abilities). Therefore, A has to correct the partner model: it decreases the initial weight of the resources. The corrected model is w AB = (0, 5, 3, 5, 2, 0, 0, 0, 0). Now the reasoning procedure triggered by the NEEDED-determinant will give a negative decision like B got. A's next two arguments (turn 3) will increase both the weights of the resources and the usefulness (both by 1) and the current reasoning procedure will again give a positive decision in the updated partner model (1, 5, 3, 6, 2, 0, 0, 0, 0).
(3) A: That's easy, you can. Doctors argue for vegetarian food.
--Refusal: although A indicated that the resources exist and presented an argument for increasing the usefulness, it turns out that the usefulness is ever smaller for B than expected by A (turn 4). Thus, A has to decrease the value of the usefulness in the partner model. Moreover, it turns out that A also has to decrease the value of the pleasantness in such a way that the reasoning procedure NEEDED would give a negative decision like B got (cf. Section 2.1.2). The changed model will be (1, 3, 3, 2, 2, 0, 0, 0, 0). But now, it turns out that A has to change even its communicative tacticsbecause it is impossible to get a positive decision in the current partner model only by increasing the usefulness. A has to go over to enticement by presenting arguments for increasing the pleasantness.
The interaction continues in a similar way until either A achieves the communicative goal (B agrees to become a vegetarian) or abandons it.
At the same time, B's own evaluations of the aspects of D (vector w B ) can be different as compared with the partner model w AB . Let us consider the example again. Let us suppose that the computer plays B's role and has generated the model of itself w B = (0, 3, 3, 2, 2, 0, 0, 0, 0). After A's proposal (turn 1) B when following its communicative strategy (Section 2.2) starts to reason about doing the action but no one reasoning procedure is applicable because the prerequisites are not fulfilled (cf. Section 2.1.2). Therefore, B refuses to do D indicating that there are no resources (turn 2). Missing of resources is a general reason to make a negative decision independent on the used reasoning procedure. After A's arguments for both the resources and the usefulness (turn 3) B corrects its model: w B = (1, 3, 3, 3, 2, 0, 0, 0, 0). Now B is able to trigger a reasoning procedure by the NEEDED-determinant (cf. Section 2.1.2) but the decision again is negative (turn 4) because w B (pleasant) is not bigger than w B (unpleasant). After A's following turn, B will correct its model once more and the interaction continues in a similar way.
We can see that both the models w B and w AB are changing as influenced by the arguments presented by the participants in interaction. Nevertheless, the models are (and, in general, remain) different. Still, the reasoning in both (even if different) models can give the same final decision. Finally, whether G A or G B will be achieved or both will be postponed.
Both A and B are working together on generating the dialog (text) by adding utterances (arguments) in every run of the REPEAT-UNTIL cycles of their communicative strategies (Section 2.2). Argumentation-based framework can be formed by the set of arguments and binary attack relations on the set (Dung, 1995). For example, after A presents an argument for the usefulness of the proposed action (Doctors argue for vegetarian food, turn 3) then B attacks it with the argument for the usefulness of not doing the action (Meat contains many useful components, turn 4).

Discussion
We suppose in our model of negotiation that both the participants A and B can use a common set of reasoning procedures in order to reason or to put themselves into the role of the partner when arguing. Starting a dialogue, A determines a partner model w AB and following his communicative strategy, chooses the communicative tactics which he will use, therefore, he determines a reasoning procedure which he will call in B's mind. B has her own model w B consisting of her actual beliefs about the aspects of doing the action D. Similarly like A chooses the communicative tactics for influencing B's reasoning, the partner B following her communicative strategy determines a reasoning procedure which she will use in order to make a decision about doing D.
We have implemented the model as a computer program which, as we believe, after the necessary further developments, can be used as a simple tool for training argumentation skills, interacting with a user in written Estonian. The computer, when acting as an opponent of the user, implements a consistent communicative strategy and in this way, it enforces the user systematically to choose certain counterarguments. In the current implementation, the computer has ready-made written sentences for stressing or downgrading the values of different aspects of the proposed action. The sentences are semantically classifiedthere are classes of sentences for increasing/decreasing of the pleasantness, the usefulness, etc. depending on the certain action. The computer randomly chooses a sentence from a suitable class. However, such an approach brings along that the generated dialogs are not quite coherent.
The user when playing A's role, similarly can use ready-made sentences (which are classified semantically) but when playing B's role she can also put in free Estonian texts. In the last case, a database is used for identifying different key words and key phrases in the user input (the input is checked against regular expressions). The database also includes an index of answer files and links to suitable answers to present to the user. Although we already have software for morphological and syntactic analysis of Estonian, there is no software for the semantic analysis. Therefore, the keywordbased topic recognition seems to be a satisfactory solution (as concluded also by the students who worked on testing our DS). Moreover, the available Estonian dialog corpus (Hennoste et al., 2008) is insufficient for implementing statistical or machine learning methods.
In order to evaluate the implementation, we have analysed human-human negotiations in the Estonian dialog corpus and compared their structure with the interactions with our DS (Koit, 2016). The communicative goals of the participants of the negotiations either coincide or are different. The participants are presenting arguments and counterarguments for and against of doing an action during a dialog. They also ask and answer questions in order to make choices among the arguments for averting the partner's counterarguments. A simplified structure of human-human negotiations as a sequence of dialog acts is as follows (here the winding brackets '{' and '}' connect a part that can be repeated; round brackets connect a part that can be missed; '/' separates alternatives; '--' starts a comment).
A: proposal {--information sharing B: question A: giving information } {--argument B: assertion/ justification/ giving information A: accept/ reject/ justification/ giving information (--argument assertion) } B: accept/ deferral/ reject As we can see (already based on the Example in Section 3.3), the actual humanhuman negotiations have more complicated structure than the interactions with our implemented DS. One limitation of the interactions is that no question-answering sub-dialog follows to the initial proposal. The aim of such an information-sharing sequence is to adjust the proposal before the arguments and counterarguments will be presented.

Conclusion
We are modelling dialogs in a natural language where one participant (initiator A) has a communicative goal that the partner (B) will make a decision about doing an action (D). When reasoning, B considers her resources as well as different positive and negative aspects of doing D. If the positive aspects weigh more than negative then the decision will be 'do D' otherwise 'do not do D'. The initiator A chooses a suitable communicative strategy and the communicative tactics in order to direct B's reasoning to the desirable decision. When trying to influence B to make a decision (e.g. to do an action), A uses a partner model and stresses the positive and downgrades the negative aspects of the action. Different arguments for doing D are presented in a systematic way, e.g. A stresses time and again the usefulness of D when persuading B to do the action. Partner B can similarly stress or downgrade a certain aspect of the action, i.e. she can continuously use the chosen reasoning procedure (which can be different as compared with the procedure called by A). Even if having a different communicative goal, B can simply reject A's goal without presenting any counterargument.
We have implemented the model of negotiation as a simple DS. The participants are interacting in written Estonian. The computer uses ready-made sentences which are classified semantically but the user as a communication partner can either choose ready-made sentences (from another set) or put in free texts. So far, the implementation is limited with a dispute about doing an action, i.e. the participants initially have opposite communicative goals. Likewise, the current implementation does not include asking and answering questions in order to adjust both the initial proposal and the arguments like it usually takes place in human-human negotiations. This needs a deeper natural language processing and remains for the further work.

Disclosure statement
No potential conflict of interest was reported by the author.

Funding
This work was supported by the Estonian Ministry of Education and Research , and by the European Union through the European Regional Development Fund [TK145] (Centre of Excellence in Estonian Studies).

Notes on contributor
M. Koit is Professor Emeritus at the University of Tartu, Estonia. She received her Ph.D. in Mathematics at the Computational Centre of the Academy of Sciences of USSR in Moscow. She has been teaching courses in Mathematics, Programming, Artificial Intelligence and Computational Linguistics. Her current research interest covers the dialog modelling. She has published several scientific and professional papers in Language Technology and Artificial Intelligence.