Cognitive Dialogue Management

Vincenzo Pallotta

Laboratory of Theoretical Computer Science

Faculty of Information and Computer Science

Swiss Federal Institute of Technology Lausanne

Abstract. Cognitive Dialogue Management is a novel approach to Dialogue Man-

agement which proposes a novel architecture for building up advanced interactive

interfaces in natural language to computers, that are both ﬂexible and robust. The

architecture, based on dialogue, results from the integration of the Information

State Dialogue Management model and the ViewFinder framework for the man-

agement of mental representations. This paper describes the principles of Cog-

nitive Dialogue Management and provides hints about how metaphors of human

information processing can be used both for improving the degree of communi-

cation understanding in human-computer interfaces, and for their rational design

and development.

1 Introduction

Dialogue Models have recently gained a great interest in Computational Linguistics

since they offer a natural framework both for the analysis of human dialogues and the

design of man-machine interfaces. A common issue of these tasks is how to model the

various types of interaction that happens among artiﬁcial and human agents at different

levels of abstraction. There are often situations where the literal meaning is not sufﬁ-

cient to understand the role of the utterance in the dialogue, when the corresponding

dialogue act cannot be directly recognized by simply looking at its linguistic content,

but it must be inferred. This is achieved by means of their cognitive skills: their abilities

to perform inferences based on background knowledge and assumptions on the other

participants’ mental states.

1.1 Cognitive Approaches to Dialogue Management

An important issue in the study of the behaviour of an intelligent agent is the logical

link between perception, cognition, and action. Communication among agents using the

language is a natural phenomenon, where the speaker acts by affecting the environment

in which both the speaker and the hearer are situated. The hearer exploits its percep-

tive capabilities in order to detect which changes have been produced by the speaker’s

communicative actions. From the hearer’s point of view, the environment is made of

everything that he/she/(it) is able to perceive, including the speaker agent. This means

that the hearer (and symmetrically the speaker) should be able to access the changing

Pallotta V. (2004).

Cognitive Dialogue Management.

In Proceedings of the 1st International Workshop on Natural Language Understanding and Cognitive Science, pages 37-50

DOI: 10.5220/0002687100370050

 SciTePress

features of the environment, including the features of the speaker’s act of speaking (i.e.

the utterance). Not all the environment’s features are directly or equally accessible by

perception. Moreover, the perception system may be insufﬁcient in extracting all the

necessary information about the communicative event or worse, it is not capable of fo-

cusing enough on relevant parts of the phenomenon that enable the understanding of

the speaker’s communication

In this paper we try to provide a more satisfactory answer to this question by pre-

senting a systematic approach to Dialogue Modelling that takes into account the above

issues and provides a cognitive model of Natural Language Communication based on

the notion of Mental State. If our main goal is to design artiﬁcial systems capable of

understanding natural language communication, we should at least imagine how cogni-

tive processes can be mapped into a computational architecture. Note that we are taking

a different perspective than that of the classical Cognitive Science: we do not commit

ourselves with the view that human cognitive processes are computational in nature,

but we rather propose that intelligent artiﬁcial systems should be modeled as cognitive

entities.

1.2 Mental States

The notion of a Mental State is fundamental and it is advocated by all the people who

believe in the potential of computational intelligence as a cognitive process. A mental

state is a container for some mental or propositional attitudes toward some speciﬁc

content in a given propositional form, which is currently stored in the agent’s long term

memory. A propositional attitude has at least three components: the agent who holds

the attitude, the type of attitude and its propositional content. Examples of propositional

attitudes types are belief, desire, intention, commitment, obligation, etc.

Following Allen in [1], we consider the Belief-Desire-Intention (BDI) agent model

as the underlying model for a conversational (cognitive) agent. The notion of mutual

belief is of central importance mutual understanding in communication. A proposition

P is mutual believed (i.e. MB(A,B,P)) by two agents A and B if A believes P, B believes

P, A believes B believes P, B believes A believes P, etc. This deﬁnition doesn’t allow

any axiomatic characterization nor computational implementation. Mutual belief can be

postulated as primitive operator without reference to simple beliefs as in [9], avoiding

inﬁnite recursion. Other problems with mutual beliefs arise when trying to distinguish

the pragmatic effect in dialogue: the agent A states that some proposition P is true, but it

is not always the case that the proposition is believed by both agents (e.g. the hearer may

have more reliable prior knowledge, for instance during an argument where someone

says he/she agrees only to stop arguing).

1.3 Goals and outline of the paper

We would like to push a little bit further the idea of simulating cognitive processes for

language understanding in communication by designing an computational architecture

This standpoint has also driven one of the most successful theory of communication and lan-

guage understanding, Relevance Theory [25], but has not been yet incorporated in any real

dialogue system or natural language interface to computers.

for Cognitive Dialogue Management that allows us to easily deal with mental repre-

sentations and their operations. Within this perspective, we propose an Architecture for

Dialogue Systems which is modular and is based on a speciﬁc decomposition where we

have mainly two macro components: the Robust Interpretation (RI) and the Knowl-

edge Assimilation (KA), also paraphrased as a perception/action (PA) component and

a mental state management (MSM) component. As a metaphor of this dichotomy we

illustrate this architecture by two (brain) hemispheres as shown in ﬁgure 1. The two

hemispheres are loosely coupled and may share some (neural) sub-structures. We de-

cided to model the Knowledge Assimilation component as a Dialogue Management

System. For this goal, we consider an efﬁcient dialogue model based on the notion of

Information State (IS) and we enhance it by adding a notion of Mental Structure (MS).

In this paper we describe a novel architecture for Dialogue Management by extend-

ing and implementing the ViewFinder framework [5] and by integrating it within the

TRINDI Dialogue Management system [19]. We will show that this solution can be

viewed as an efﬁcient and viable alternative to the plan-based approaches to dialogue

management [10, 17, 2]. We do not discuss the Robust Interpretation/Perception com-

ponent here. The interested reader may refer to [3, 23].

Fig. 1. Cognitive Language Architecture for Dialogue Systems

2 Dialogue Management in TRINDI

The Dialogue Manager (DM) is the program which coordinates the activity of several

subcomponents in a dialogue system and, traditionally, it has as its main goal that of

maintaining a representation of the current state of the ongoing dialogue. Typically,

a DM receives as input a dialogue act which contains a semantic representation of a

user’s utterance (or a set of utterances) which has been interpreted by an interpretation

module. The term dialogue act goes back originally to Bunt [7] and can be under-

stood both loosely in the sense of “speech act used in a dialogue” and, in a more spe-

cialised sense, as functions which updates the dialogue context. Bunt also claims that

the context-change approach to dialogue acts can solve difﬁculties arisen from pure

speech-act theory, which concerns only with the assignment of the illocutionary force

and the propositional content of utterances and their further classiﬁcation into a tax-

onomy. The Bunt standpoint of dialogue acts as context updates is best represented by

the Information State approach to Dialogue Management proposed in [11] and imple-

mented in the TRINDI toolkit [19]. The Information State theory of dialogue modelling

consists of:

– a description of the informational components of the theory (e.g. participants, com-

mon ground, linguistic and intentional structure, obligations and commitments, be-

liefs, intentions, user models, etc.);

– a formal representation of the above components (e.g. as a typed feature structure,

modal logic, abstract data types);

– a set of dialogue moves that will trigger the update of the information state;

– a set of update rules, that govern the updating of the information state;

– an update strategy for deciding which rule(s) to select at a given point, from the set

of applicable ones.

The information state (IS) is stored internally by an agent (e.g. dialogue system). The

information state and all resources are seen as abstract datatypes (i.e. sets, stacks etc.)

with related conditions and operations. The IS is composed of a static part (SIS), which

remains constant during a dialogue. It includes rules for interpreting utterances, up-

dating the dynamic part of the information state, and selecting further moves; option-

ally, move deﬁnitions, plan libraries, static databases etc. The dynamic part of IS (DIS)

changes over time depending on the occurring events and how these events are treated

by the dialogue move engine (DME). A typical (minimal) information state structure

is showed in ﬁgure 2. The main division in the information state is between informa-

Fig. 2. Information State (Cooper & Larson)

tion which is private to the agent and that which is (assumed to be) shared between

the dialogue participants. Shared information is considered here what has been estab-

lished (or grounded) during the conversation. Hypotheses about the ungrounded shared

IS are kept in a temporary slot in the private IS, until an update rule for grounding is

enabled and executed. The “QUD” slot (i.e. the question under discussion) provides the

underlying mechanism for dealing with sub-dialogues.

Observe that there is no assumption on how the mental state of the agent should look

like, except from the fact that it is a collection of (believed) proposition. The extension

of IS we propose is one where the “BEL” slot is replaced by a complex dynamic struc-

ture which represents the evolution of the agent mental state in terms of nested mental

attitudes as the one that will be discussed in the next section.

The Dialogue Move Engine updates the IS on the basis of the observed dialogue

moves and selects the appropriate moves to be performed. Dialogue Moves (DM) are

theoretically deﬁned using preconditions, effects & decomposition, but in practice DMs

are not directly associated with preconditions and effects (i.e., output of the interpreta-

tion module and input to the generation module).

The main difference between information state approaches and other structural, di-

alogue state approaches lies on the fact that the latter are based on the notion of “legal”

or “well-formed” dialogue, described by some generative formalism (e.g., grammars,

ﬁnite state automata). Information states can be partially described. Moreover, the mo-

tivation for update the information state and selecting a next dialogue move may rely

only on part of the information available. The model is simpler than plan-based models,

but extensible enough to cope with representations of mental states, as we will discuss

later.

3 The ViewFinder framework

In this section we describe a framework for the representation of knowledge assimilated

through perception (including communication) and cognitive processes into the men-

tal state of a cognitive agent. In our proposal for Cognitive Dialogue Management we

adopt and extend the ViewFinder framework, early proposed by Ballim and Wilks in [4]

and further formalized by Ballim in [5]. The main motivation is that language under-

standing cannot be obtained without a representation of intentionality. Moreover, this

representation must be computational if our goal is to build artiﬁcial systems capable of

natural language processing beyond pure structural linguistic analysis (e.g., phonology,

morphology, syntax).

In order to build mental representations of the perceived information, one can as-

sume that there exist some general principles and general representations that are com-

mon to all the cognitive agents. It is of fundamental importance the way information is

organized in cognitive agent’s mental structures. Nonetheless, we believe that feeding

mental structures with representations of the acquired information can only be achieved

if coupled with some form of linguistic analysis. We do not address this problem here,

but the interested reader can look at [24] for the development of this topic.

Partitioned representations have been advocated as a main conceptual tool for struc-

turing information without adding any spurious semantics to its content . The main

principle relies on the assumption that information typically has local coherence. In Ar-

tiﬁcial Intelligence and in particular in Knowledge Representation and Reasoning, the

usefulness of partitioned representations have been early recognized by John McCarthy

[22]. A survey on this topic [6] classiﬁes frameworks for partitioned representations (or

contexts) in two main categories: the divide-and-conquer approach and the compose-

and-conquer approach. The former sees partitioned representations as a way of parti-

tioning a global model of the world into smaller and simpler pieces (examples are the

work of Dinsmore [12] and that of McCarthy, formalized in [8]). The latter considers

partitioned representations as local theories of the world interconnected by a network

of relations (as in both Local Model Semantics (LMS) [15] and ViewFinder [5]).

ViewFinder is a formal framework for representing, ascribing and maintaining nested

attitudes of interacting agents. Viewpoints on mental attitudes of communicating agents

are represented by means of nested typed environments. Operations over typed environ-

ment are deﬁned and used to simulate several forms of reasoning, which are necessary

in order to assimilate the information into knowledge structures from communication.

The early implementation of ViewFinder is the ViewGen system [4] and it is intended

for use in modelling of nested beliefs in autonomous interacting agents. The ViewFinder

framework provides the theoretical foundations for ViewGen in particular with respect

to the following issues related to the manipulation of environments:

– Correspondence of concepts across environments (i.e. intensional objects);

– Operations performed on environments (e.g. ascription, adoption);

– Maintenance of environments.

Relationships between environments can be speciﬁed hierarchically or by the explicit

mapping of entities. Each environment has associated an axiomatization and a reasoning

system.

ViewFinder has been recently implemented as a computational framework [24], and

successfully tested on classical knowledge representation problems in dialogue such as

the “Three Wise Men” problem [21, 22]. Moreover, plan recognition in ViewFinder has

been addressed by Lee in [20]. In the rest of the section we rapidly review the basic

notions of ViewFinder. Details are available in [5].

3.1 Backgrounds of ViewFinder

The fundamental notion in ViewFinder is that of environment (or spaces). Environments

are containers for information and can be nested. This means that an environment can

contain other environments which can in turn contain other environments and so on. An

environment contains some propositional content which is assumed to be consistent.

The propositional content can be organized with respect to topics. A nested environment

is thus a point-of-view with respect to a particular topic, that is, an attitude toward a

particular content. A topic itself is a special type of environment.

Different attitudes (e.g. beliefs, goals, intentions) correspond to different environ-

ment types. Like other environment types, belief spaces can be nested. We can thus

represent a point-of-view on a given topic with respect to another nested environments

as, for instance, in the expression:

Bel(John, wants(F red, Bel(Anne, P ))).

Nested environments can be created extensionally or intensionally. Environment oper-

ators project the content of an environment onto another environment. Two operators

are deﬁned: the ascription and the adoption. The ascription takes the propositional con-

tent of an environment and projects it onto an inner environment. It can be viewed

as a rule that dynamically generates or updates the content of nested environments by

projecting the content of the outer environment onto the inner environments, provided

that none of the inner spaces contain propositional content which is inconsistent with

the projected information. Figure 3 shows the prototypical situation with the belief en-

vironment type. The second operation on environments is adoption, referred in [4] as

Fig. 3. Default ascription

percolation. Adoption is the ascription’s dual operation. Adoption projects information

from an inner environment to an outer environment. While the purpose of the ascription

algorithm is to allow the dynamic generation of viewpoints, the purpose of the adoption

algorithm is to accommodate information in the system’s knowledge base (i.e. the men-

tal state) from viewpoints on other agents attitudes. This situation is sketched in ﬁgure

4. The adoption operator allows an agent to extend its mental state with information

Fig. 4. Default adoption

he gathers from other agents through communication and collaboration, and not only

through his own observations of the real world. The adoption operator is triggered when

new information becomes available in an environment, for instance after the processing

of a dialogue act. The set of new information is propagated to other environments using

the adoption.

The whole theory of ViewFinder is based on the notion of default reasoning. An

agent can always assume that another agent mirrors its mental space except when con-

trary evidence exists. ViewFinder adopts a different approach to common, shared, or

mutual knowledge as found in other theories of mental attitudes [9, 14, 18], since it as-

sumes that knowledge is always shared unless a contrary evidence arises.

Feature-based Topic Language In contrast with ViewGen and the formal model of

ViewFinder, we introduce a novel propositional content language for ViewFinder. A

topic T is associated a set of features identiﬁed by a common name, that is F

, . . . , f

}. Each feature f can take values in a ﬁnite domain set (enumeration type)

denoted by D

. A topic can be part of a hierarchy and it inherits features from its parent

topics. A feature can be redeﬁned in the child topic. In this sense a topic is a container

for structured information.

For example, the topic WeatherBroadcast:

topic GeneralWeatherBroadcast

date: [tomorrow, in 2 days, in one week]

weather: [sunny, cloudy, rainy]

temperature: [cold, warm, hot]

can be specialized to:

topic DetailedWeatherBroadcast

inherits from GeneralWeatherBroadcast

temperature: [<5C, 5-25C, >25C]

wind: [N, S, W, E]

velocity: [0n, 5n, 10n, 15n, 20n, 30n]

The notion of stereotypes of ViewGen is also present in ViewFinder. This is a powerful

concept for classifying agents and ascribe them some default behavior, knowledge or

reasoning capability. We have decided to represent this notion differently by grouping

agents into classes. Each agent belongs to one or more classes. The appropriate topic

deﬁnition is selected by an agent A belonging to the class A and it is denoted by T (A).

This parameter allows us to have different deﬁnitions of the same topic for different

classes of agent. A topic can have more or less features with respect to an agent class,

modeling the notion of agent competency. We say that the knowledge about a topic T is

complete if all the features f

have a unique value, or equivalently if the topic is repre-

sented by only one possible combination of feature-values pairs (a tuple)

. Conversely,

we say that an agent has no opinion about a topic, if he admits that any tuple of the topic

space is valid. We call opinion (of an agent about a topic) a subspace of the whole topic

space.

We have decided to model knowledge about topics as integrity constraints over

the topic space. An agent builds its opinion about a topic by an elimination process. It

interprets the constraints and eliminates all the incompatible tuples from the topic space.

Complete knowledge is achieved when the constraints have reduced the topic space to

The term ’complete’ has been chosen in analogy to the complete meta-predicate of the Kim

and Kowalski’s work [18].

only one possible combination of features, whereas inconsistency appears when the

constrained topic space is empty. This method allows us to deal with open theories and

disjunctive information. From a logic programming point of view, topic features can be

viewed as abducible predicates with integrity constraints [16].

Representing opinions The topic language may be sufﬁcient when representing one

agent’s knowledge about some aspects of the world, but is not expressive enough to

capture all the information needed to model other agent’s point of view. A topic space

allows us to model in a natural way a disjunction of possible topic instances, that we

shrink each time we acquire further information, but only in case we have complete

knowledge about another agent’s opinions. We say that feature f is undeﬁned and we

write f = λ(A) where the parameter A is the only agent capable of providing a value

for the feature f.

Ballim and Wilks in [4] have proposed a taxonomy of competency in believing

showed in ﬁgure 5. They consider differently an agent who may be competent in the

evaluation of a feature and an agent who actually is an evaluator for that feature. An

agent can have competency to evaluate a feature (i.e. belong the appropriate class), and

then if he is competent, he can possibly be an evaluator for it. In the positive case, an-

other agent might be aware or not of the way it assigns values to the features. As shown

in [24], this topic language allows to cope with all the cases of the belief competency

taxonomy described in ﬁgure 5.

Fig. 5. The Ballim and Wilks’s taxonomy of Competency in Believing

3.2 Assimilating mental attitudes from communication

In this section we tackle the problem of assimilating knowledge from communications

between agents in ViewFinder. We followed the ideas both from the Lee’s [20] and

Dragoni’s works [13]. Assimilating new information requires to solve three problems:

1. decide for a communication language understood by the two agents speaking,

2. represent the knowledge in each mental states

3. transfer the knowledge contained in the message into the mental state.

We already provided a solution for the second issue with the topic language and we have

decided to adopt for the content level the same representation we used for topics, that

is, a set of constraints on the topic space. In other words, topic language can be used

also as the propositional content representation in dialogue acts. Since the moment a

dialogue act is taken in charge by the Dialogue Manager until the whole information

contained is digested, four steps have occurred:

1. the translation of the message content into the mental state representation language.

2. the comparison of the actual mental representation of the sender with the message

sent for consistency check.

3. the derivation of new information (i.e. constraints) using the actual mental state and

the message received;

4. the adoption of the propagated of new information in the whole system (i.e., forward-

chaining).

Fig. 6. The Knowledge Assimilation Process

This process is showed in ﬁgure 6. Once the receiver has decoded the message (i.e.,

extracted the dialogue act), she can use it and update her mental state. Its ability of

managing viewpoints is very interesting at this point. The receiver builds her mental

representation of the sender, then she compares the received message with her point of

view on the sender. If she detects any inconsistency she can take the appropriate action,

for instance to inform the sender of the inconsistency, to ask her for more information

or to reject the message. If the message is compatible, she assimilates it. The integra-

tion process takes place in the point of view of the sender seen by the receiver. The

algorithm, depending on the content language, generates new content. In our solution

the process uses a propagation mechanism to generate new constraints. The receiver

reasons on her viewpoint on the sender and derives constraints that should be consistent

within the sender’s mental state. At the end of the derivation process, the point of view

of the sender seen by the receiver is updated with all this new information. This new in-

formation can then be back-projected (from inner to outer environments) to the receiver

using adoption provided that the receiver trusts the sender, and if the new information

is consistent with the already present propositional content.

4 Integration of DME and ViewFinder

We describe, as the last step, the integration of Information State and Mental State in

TRINDI Dialogue Management. We add a new element to the private component of IS,

an instance of the new “Env” datatype, representing the outer mental environment of the

agent. The ViewFinder system’s environment provides more context for the interpreta-

tion and assimilation of dialogue acts in the Information State (including updates of the

ViewFinder environment itself). The following is the description of the Env TRINDI

datatype with its operations and conditions:

type Env

operations:

tell( A1, A2, Topic, Features, Values )

untell( A1, A2, Topic, Features, Values )

ask( A1, A2, Topic, Features, Response )

cleanKB

loadEnv(EnvSignature)

conditions:

instanceOf(TopicName, TopicType);

topicDef(TopicName, AgentClass, TopicDef);

isa(TopicName, ParentTopicName);

agentDef(AgentClass);

adopts(AgentName1, AgentName2, TopicName);

ascribes(AgentName1, EnvType1, AgentName2, EnvType2).

The operations and conditions mirror the ViewFinder operations on the content of

environments expressed in the topic language. We added the following Dialogue

Moves to the basic set of TRINDI moves in order to model dialogue acts having

content expressed in the ViewFinder topic language:

ask(Agent, Topic, Features),

askif(Agent, Topic, Features, Values),

answer(Agent1, Agent2 Topic, Features, Values),

inform(Agent, Topic, Features,Values).

The update rules for grounding and planning are not discussed here. We only show the

rule for answer integration, which is triggered when the dialogue manager decides to

assimilate the information obtained from the successful processing of the answer

dialogue move in the Env componente of the private information state:

RULE : integrateOwnAnswer

CLASS : integrate

PRE:

valRec( SHARED.LU.SPEAKER, PRIVATE.NAME )

assocRec( SHARED.LU.MOVES,

answer(PRIVATE.NAME, A, Topic, Features, Values), true)

EFF:

popRec( SHARED.QUD )

tellRec( PRIVATE.ENV, PRIVATE.NAME , A, Topic, Features, Values)

setAssocRec(SHARED.LU.MOVES, answer(P), true)

5 Conclusions and future work

In this paper we described an architecture for Cognitive Dialogue Management based on

the integration of TRINDI Dialogue Management toolkit with the ViewFinder frame-

work for Mental Spaces Management. The Information State approach to Dialogue

Management has been adopted by integrating the Information State in the the TRINDI

Toolkit. Information State, as initially proposed in the TRINDI framework, does not

sufﬁce for taking into account different dimensions of a dialogue. We also need to con-

sider the evolution of the system’s view about the speaker’s mental state, which can be

reported and simulated within an appropriate data structure. The structure we proposed

to integrate into the Information State is that of Nested Typed Environments provided by

ViewFinder. In order to cope with frequent cases of partial knowledge about the world

and to avoid problems related to the construction of mutual knowledge, we proposed an

extension of ViewFinder by a systematic account of operations between mental spaces

and by introducing a new content representation language based on integrity constraints

over set of features. The choice of the operations and the content language allows us to

robustly and efﬁciently deal with complex epistemic problems in communication, such

as the “three wise men” problem.

Summing up, in the notion of Cognitive Dialogue Modelling we incorporate a set of

tools helpful to achieve robust human-computer interfaces where understanding from

communication is an essential factor. Understanding can be partial in some circum-

stances, but the system should be able to robustly react and complete the missing infor-

mation by contextual knowledge. Context must include also a model of the users, and

in particular a model of their mental state. Moreover, the design of artiﬁcial dialogue

system should be based on a model of cognitive processing rather than on the simple

extraction of information from sophisticated input devices. We proposed to extensively

exploit Computational Logic to conceive and soundly implement inference systems as

the basis of a computational cognition. Cognition in artiﬁcial systems needs not to mir-

ror human cognition, but human cognition metaphors can still be used and adapted to

computer systems.

References

1. J. Allen. Natural Language Understanding. Benjamin/Cummings, Menlo Park, 1987.

2. J. Allen and C.R. Perrault. Analyzing intention in dialogues. Artiﬁcial Intellicence,

15(3):143–178, 1980.

3. A. Ballim and V. Pallotta. Robust methods in analysis of natural language data. Natural

Language Engineering, 8(2), 2002.

4. A. Ballim and Y. Wilks. Artiﬁcial Believers. Lawrence Erlbaum Associates, Hillsdale, New

Jersey, 1991.

5. A. Ballim. ViewFinder: A Framework for Representing, Ascribing and Maintaining Nested

Beliefs of Interacting Agents. Ph.d., University of Geneva, 1992.

6. P. Bouquet, C. Ghidini, F. Giunchiglia, and E. Blanzieri. Theories and uses of context in

knowledge representation and reasoning. Technical Report 0110-28, IRST, October 2001.

”citeseer.nj.nec.com/bouquet01theories.html”.

7. H. Bunt. Conversationa principles in question-answer dialogus. In D. Krallmann, editor, Zur

Theory der Frage, pages 119–141. Narr Verlag, 1979.

8. S. Buva

c, V. Buva

c, and I.A. Mason. Metamathematics of contexts. Fundamenta Informati-

cae, 23(3), 1995.

9. P. R. Cohen and H. J. Levesque. Intention is choice with commitment. Artiﬁcial Intelligence,

42(2-3):213–261, 1990.

10. P.R. Cohen and C.R. Perrault. Elements of a plan-based theory of speech acts. Cognitive

Science, 3(3):177–212, 1979.

11. R. Cooper. Information states, attituds and dialogue. In Proceedings of the Second Tblisi

Symposium on Language, Logic and Computation, Tblisi, September 1997.

12. J. Dinsmore. Partitioned Representations: A study in mental representation, language under-

standing and linguistic structure, volume 8 of Studies in cognitive systems. Kluwer, 1991.

13. A.F. Dragoni, P. Giorgini, and L. Seraﬁni. Mental states recognition from communication.

Journal of Logic and Computation, 12(1):119–136, February 2002.

14. R. Fagin, J.Y. Halpern, Y. Moses, and M.Y. Vardi. Reasoning about knowledge. MIT Press,

1996.

15. C. Ghidini and F. Giunchiglia. Local model semantics, or contextual reasoning = locality +

compatibility. Artiﬁcial Intelligence, 127(2):221–259, 2001.

16. A. C. Kakas, R. Kowalski, and F. Toni. The role of abduction in logic programming. In D.M.

Gabbay, C.J. Hogger, and J.A. Robinson, editors, Handbook of Logic in Artiﬁcial Intelligence

and Logic Programming, volume 5, pages 235–324. Oxford University Press, 1998.

17. H. Kautz. A formal theory of plan recognition and its implementation. In J.F. Allen, H.A

Kautz, R.N. Pelavin, and J.D. Tennenberg, editors, Reasoning About Plans, chapter chapter

2, pages 69–126. Morgan Kaufmann, 1991.

18. J. S. Kim and R. A. Kowalski. An application of amalgamated logic to multi-agent belief. In

M. Bruynooghe, editor, Second Workshop on Meta-Programming in Logic META90, pages

272–283, Dept. of Computer Science, 1990. Katholieke Univ. Leuven.

19. S. Larsson and D. Traum. Information state and dialogue management in the trindi dialogue

move engine toolkit. Natural Language Engineering, 6(3 & 4):323–340, 2000.

20. M. Lee. Belief, Rationality and Inference: A general theory of Computational Pragmatics.

Ph.d. thesis, University of Shefﬁeld, 1998.

21. J. McCarthy. Formalization of two puzzles involving knowledge. http://www-

formal.stanford.edu/jmc/puzzles/puzzles.html, 1987.

22. J. McCarthy. Notes on formalizing contexts. In Ruzena Bajcsy, editor, Proceedings of the

Thirteenth International Joint Conference on Artiﬁcial Intelligence, pages 555–560, San Ma-

teo, California, 1993. Morgan Kaufmann.

23. V. Pallotta and A. Ballim. Agent-Oriented Language Engineering for robust NLP In

A. Omicini, P. Petta, R. Tolksdorf, Eds., Proceedings of Second International Workshop,

ESAW 2001, Engineering Societies in the Agents World II, pages 86–104. (Springer LNAI

2203). Prague, Czech Republic, July 7, 2001.

24. V. Pallotta. Cognitive Language Engineering: Towards Robust Human-Computer Interac-

tion. Ph.d. thesis, Swiss Federal Institute of Technology Lausanne (EPFL), August 2002.

25. D. Sperber and D. Wilson. Relevance: Communication and Cognition. Language and

Thought Series. Harvard University Press, Cambridge, MA, 1986.