Dialogue moves for DIALOG

66
Dialogue moves for DIALOG Elena Karagjosova, Dimitra Tsovaltzi [email protected],[email protected] Reviewer: Armin Fiedler February 28, 2005

Transcript of Dialogue moves for DIALOG

Dialogue moves for DIALOG

Elena Karagjosova, Dimitra Tsovaltzi

[email protected],[email protected]

Reviewer: Armin Fiedler

February 28, 2005

2

Contents

1 Motivation 51.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51.2 Philosophy of Taxonomy . . . . . . . . . . . . . . . . . . . . . . . 51.3 Background and Related Work . . . . . . . . . . . . . . . . . . . 6

1.3.1 DAMSL dimensions . . . . . . . . . . . . . . . . . . . . . 61.3.2 Adaptations . . . . . . . . . . . . . . . . . . . . . . . . . . 7

1.4 Overview of The Taxonomy . . . . . . . . . . . . . . . . . . . . . 81.4.1 The Dimensions in the Taxonomy . . . . . . . . . . . . . 8

1.5 Tutorial Dialogue: Task Dimension Modelling . . . . . . . . . . . 111.6 Conclusion and Future Work . . . . . . . . . . . . . . . . . . . . 14

2 Annotations 172.1 Segmentation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 172.2 Procedure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18

3 DIALOG Annotation Scheme 193.1 Forward-looking Function Dimension . . . . . . . . . . . . . . . . 19

3.1.1 Statement . . . . . . . . . . . . . . . . . . . . . . . . . . . 203.1.2 Influencing-addressee-future-action . . . . . . . . . . . . . 213.1.3 Info request . . . . . . . . . . . . . . . . . . . . . . . . . . 233.1.4 Diagnostic query . . . . . . . . . . . . . . . . . . . . . . . 243.1.5 Understanding query . . . . . . . . . . . . . . . . . . . . . 253.1.6 Committing-speaker-future-action . . . . . . . . . . . . . 253.1.7 Conventional . . . . . . . . . . . . . . . . . . . . . . . . . 263.1.8 Apologise . . . . . . . . . . . . . . . . . . . . . . . . . . . 273.1.9 Gratitude . . . . . . . . . . . . . . . . . . . . . . . . . . . 273.1.10 Signaling emotion . . . . . . . . . . . . . . . . . . . . . . 283.1.11 Other forward function . . . . . . . . . . . . . . . . . . . 283.1.12 Decision tree . . . . . . . . . . . . . . . . . . . . . . . . . 29

3.2 Backward-looking Function Dimension . . . . . . . . . . . . . . . 293.2.1 Agreement . . . . . . . . . . . . . . . . . . . . . . . . . . 303.2.2 Understanding . . . . . . . . . . . . . . . . . . . . . . . . 323.2.3 Answer . . . . . . . . . . . . . . . . . . . . . . . . . . . . 363.2.4 Information-relation . . . . . . . . . . . . . . . . . . . . . 37

3

4 CONTENTS

3.2.5 Decision tree . . . . . . . . . . . . . . . . . . . . . . . . . 393.3 Task Dimension . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39

3.3.1 Domain contribution . . . . . . . . . . . . . . . . . . . . . 403.3.2 Hint . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 413.3.3 Domain contribution evaluation . . . . . . . . . . . . . . . 433.3.4 Check origin of problem . . . . . . . . . . . . . . . . . . . 463.3.5 Request evaluation . . . . . . . . . . . . . . . . . . . . . . 473.3.6 Prompt . . . . . . . . . . . . . . . . . . . . . . . . . . . . 473.3.7 Encourage . . . . . . . . . . . . . . . . . . . . . . . . . . . 483.3.8 Request assistance . . . . . . . . . . . . . . . . . . . . . . 483.3.9 Resign . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 493.3.10 Align . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 503.3.11 Decision tree . . . . . . . . . . . . . . . . . . . . . . . . . 50

3.4 Conventional Task Management Dimension . . . . . . . . . . . . 503.4.1 Initiate task . . . . . . . . . . . . . . . . . . . . . . . . . . 513.4.2 Close task . . . . . . . . . . . . . . . . . . . . . . . . . . . 513.4.3 Initiate subtask . . . . . . . . . . . . . . . . . . . . . . . . 523.4.4 Close subtask . . . . . . . . . . . . . . . . . . . . . . . . . 52

3.5 Conventional Communication Management Dimension . . . . . . 523.5.1 Initiate dialogue . . . . . . . . . . . . . . . . . . . . . . . 523.5.2 Close dialogue . . . . . . . . . . . . . . . . . . . . . . . . 523.5.3 Initiate subdialogue . . . . . . . . . . . . . . . . . . . . . 533.5.4 Close subdialogue . . . . . . . . . . . . . . . . . . . . . . 533.5.5 Discourse marking . . . . . . . . . . . . . . . . . . . . . . 53

3.6 Communicative Status Dimension . . . . . . . . . . . . . . . . . . 533.6.1 Uninterpretable . . . . . . . . . . . . . . . . . . . . . . . . 533.6.2 Abandoned . . . . . . . . . . . . . . . . . . . . . . . . . . 543.6.3 Self talk . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54

4 Annotation tool 59

A Font Coded Taxonomy 61A.1 Forward-looking . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61A.2 Backward-looking . . . . . . . . . . . . . . . . . . . . . . . . . . . 62A.3 Task . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63

A.3.1 Proof task . . . . . . . . . . . . . . . . . . . . . . . . . . . 63A.3.2 Tutoring task . . . . . . . . . . . . . . . . . . . . . . . . . 63

A.4 conventional task management . . . . . . . . . . . . . . . . . . . . 63A.5 conventional communication management . . . . . . . . . . . . . . 64A.6 communicative status . . . . . . . . . . . . . . . . . . . . . . . . . 64

Chapter 1

Motivation

1.1 Introduction

Empirical evidence shows that natural language dialogue capabilities are nec-essary for the success of tutorial sessions (Moore, 1993). Moreover, the use ofteaching strategies, which can be well modelled in dialogue systems, is of par-ticular interest and importance for intelligent tutoring systems, as the aim is toteach in an effective way. The modelling of teaching strategies can manipulateaspects of learning, such as help the student build a deeper understanding of thedomain, eliminate cognitive load and promote schema acquisition (Wilson andCole, 1996; Lim and Moore, 2002). Despite these insights, only few state-of-the-art tutoring systems use natural-language interaction and teaching strategies,which constitute genre specific modelling, in an elaborate and structured way.

We suspect the reason for the limited use of these ingredients lies in thesystem architectures, in which domain knowledge, tutoring and pedagogicalknowledge, and dialogue management are tightly intertwined. Therefore, inthe DIALOG project (Benzmuller et al., 2003), we aim at a mathematical tu-toring system, in which domain knowledge, dialogue capabilities, and tutorialphenomena can clearly be identified. The sub-domain chosen is elementary settheory.

1.2 Philosophy of Taxonomy

In this taxonomy, we focus on the separation of generic dialogue management,on the one hand, and the manipulation of genre and domain specific phenomena,involved in modelling different teaching models and domains, on the other hand.The latter we view as only a sub-part of the dialogue manager. In order toachieve that we propose a dialogue move taxonomy in the tradition of the multi-dimensional structure of DAMSL (Allen and Core, 1997) to capture the multi-functionality of utterances. We define dialogue moves in it in a top-down mannerusing empirical data that we have collected (cf. (Wolska et al., 2004)) and guided

5

6 CHAPTER 1. MOTIVATION

by previous taxonomies (Allen and Core, 1997; Core et al., 2002). We seekinter-relations among the moves in the dimensions and carry out an annotationprogramme in order to test our hypotheses.

We mainly divert from other approaches in that we suggest an expandedtask dimension for clearly separating out and defining the genre and domainspecific characteristics of dialogue. This further facilitates the original idea ofthe DAMSL attempt for reusability and reconfigurability, as it provides a betterframework for capturing what is generic in dialogue management and reusablebetween genres, and, contrary, what is specific to the genre or even the differentdomains. The latter also makes modelling the genre, and the different domainsin it, more straight forward.

In the rest of this chapter, we first discuss related work in Section 1.3. InSection 1.4 we look into the different dimensions in our taxonomy, and in Section1.5 we explain in more detail our approach to capturing specific tutorial dialoguephenomena via the closer investigation of the task dimension. Section 2 outlinesour annotation programme before the draw some conclusions.

1.3 Background and Related Work

We developed a taxonomy of dialogue moves based on a preliminary analysisof the DIALOG corpus1. The taxonomy is based on DAMSL (Allen and Core,1997), which is an attempt to provide a standard top-level structure for annotat-ing dialogues aiming at enhancing reusability of dialogue annotation schemes.More specifically, we make use of the multiple level structure of the DAMSLscheme, which allows catering for the various functions an utterance may playin dialogue. This is an advantage of DAMSL compared to other major pro-posals, such as the Verbmobil dialogue act annotation scheme (Alexanderson etal, 1997). The Verbmobil scheme represents a flat hierarchical structure withgrowing specificity towards the leaves. The core of the dialogue moves definedfor Verbmobil is moreover specific to the task, namely appointment scheduling,which does not allow the reusability of the annotation scheme.

Let us briefly describe the DAMSL scheme, as it is the basis of our work, inorder to look into our particular adaptations.

1.3.1 DAMSL dimensions

We take as a basis the DAMSL annotation scheme (Allen and Core, 1997).DAMSL allows multi-level annotation. We modify it and extend it with movesfrom (Core et al., 2002) and newly defined ones.

DAMSL distinguishes four dimensions according to the unit’s purpose androle in dialogue:

• Communicative status: whether utterance is intelligible and whether itwas successfully completed (uninterpretable, abandoned, self-talk)

1Soon available online, cf. also (Wolska et al., 2004)

1.3. BACKGROUND AND RELATED WORK 7

• Information level: provides an abstract characterisation of the semanticcontent of the utterance

– Task: “Doing the task”, i.e., utterances that advance the task

– Task-management: “Talking about the task”, i.e., utterances thatdiscuss the problem solving process or experimental scenario

– Communication management: “Managing the communication”, i.e.,conventional phrases that maintain contact, perception, and under-standing during the communication process: greetings, closings, ac-knowledgments (“Okay”, “uh-huh”), stalling for time (“Okay”, “Letme see”), signals of speech repairs (“oops”) or misunderstandings(“sorry?”, “huh?”)

– Other-level

• Forward-looking function: characterises what effect an utterance has onsubsequent dialogue and interaction

• Backward-looking function: captures the way the current utterance isrelated to the previous dialogue

1.3.2 Adaptations

We now look into our basic adaptations to the DAMSL scheme.We assume six dimensions instead of only four in DAMSL by making the

DAMSL sub-levels task, task-management and communication management,which belong to the top-level information level, separate dimensions of theirown, for reasons that are explained in Section 1.4.1.

We modified and extended the dialogue moves in DAMSL with moves fromthe BE&E annotation scheme (Core et al., 2002), which was developed for thetutorial dialogue genre. The BE&E annotation scheme is based on DAMSL,however without distinguishing all levels that DAMSL provides for. It, fur-thermore, provides additional dialogue moves derived from a tutorial dialoguecorpus. We adopt some of these moves, but in order to cater for more tutorialdialogue genre specific and domain specific phenomena, we define new dialoguemoves and group them in a separate task level, following DAMSL (cf. Section1.4.1.3). The need for such separation in dialogue systems has been also advo-cated by (Allen et al, 2001) who use a modular architecture for dialogue andtask planning, for the domain of route planning.

We modified the definitions of some of the moves we took over from DAMSLand BE&E. For instance in DAMSL, an info request is defined as any utter-ance that creates an obligation for the hearer to provide information. However,investigating tutorial dialogue we recognised the need to distinguish betweenquestions whose answer the speaker knows, and such whose answer they do notknow. We consider only the latter to be an info request, and account for theformer as a diagnostic query (Core et al., 2002).

8 CHAPTER 1. MOTIVATION

1.4 Overview of The Taxonomy

In this section, we describe the dimensions in the taxonomy and state theirpractical use.

1.4.1 The Dimensions in the Taxonomy

The taxonomy features many different dimensions to capture the fact that an ut-terance produced by a speaker may have many functions (serve many purposes)at the same time, much in the DAMSL fashion (Allen and Core, 1997). Everyone of the dimensions represents one of these potential different functions. Theactual detailed description of what the utterance does in the particular context(the dialogue at hand) corresponds to one of the categories of dialogue movesdefined in the dimensions themselves. In every dimension, there are differentlevels of description of the dialogue moves. These are structured in classes andsubclasses, up to three different levels deep. The current function of an utteranceis represented on the deepest level of subclasses, which inherit the properties ofthe respective super-classes. There are in total six dimensions in the taxonomy.

1.4.1.1 Forward-looking Dimension

The forward-looking dimension characterises the effect an utterance has on thesubsequent dialogue (Allen and Core, 1997). In this dimension there are 11classes of dialogue moves, some of them with subclasses. Most of the DMs areadopted from DAMSL with slightly modified definitions, two are taken fromBE&E, and four are newly defined. An instance of a dialogue move in thisdimension is info request. We define it as an utterance that requests informationwhich the speaker does not possess. This definition differs from the one providedin DAMSL, according to which any utterance that creates an obligation for thehearer to provide information is an info request. However, in tutorial dialogues itis necessary to distinguish between questions whose answer the speaker knows,and such whose answer he does not know, especially in the case where thespeaker is the tutor. We consider only the latter to be info request, and caterfor the former by adopting the dialogue move diagnostic query defined in (Coreet al., 2002) as an utterance by which ”the speaker is testing whether a listenerknows a piece of information by asking them to supply the information” (p.4).An example from our corpus for an info request is (1.1) (Dialogue: soc23p) andfor a diagnostic query (1.2) (where T is the tutor; Dialogue: soc12k)2.

(1.1) T: Glauben Sie, dass Sie das nun beweisen mussen oder ist das eine Folgerungaus dem Obigen?

(1.2) T: Wissen Sie, wie Sie diese Beziehung benutzen koennen?

2The examples are from the DIALOG corpus. The reference name of the particular dialoguesignifies, from right to left, the teaching method, the number of the subject and the proofthat was the task.

1.4. OVERVIEW OF THE TAXONOMY 9

1.4.1.2 Backward-looking Dimension

The backward-looking dimension captures how the current utterance relatesto the previous discourse (Allen and Core, 1997). Every move here has anantecedent, i.e., a dialogue move in the preceding discourse, which is affectedby the backward move currently performed. Note that the antecedent need notbe necessarily in the immediately preceding discourse but can be more distant.It can also be the case that a move has multiple antecedents.

Most of the dialogue moves in this dimension are also adopted from DAMSL.An exception is the category Information-relation which DAMSL suggests as apossible aspect of the backward-looking function but does not elaborate on it.We define five address-moves which apply when neither of the more specific di-alogue moves in this dimension accept, reject and answer apply. The main mo-tivation for having address-moves is a distinction we want to make between theaspects resolvedness vs aboutness of answers to questions, following (Ginzburg,1996). Resolvedness and aboutness are relations between questions and answers.An answer resolves a question when the answer provides information that posi-tively or negatively resolves the question (Larsson, 2002). ‘About’ is a relationthat accounts for the range of information associated with a question (Ginzburg,1994). An answer move always resolves the question. If the question is followedby an utterance which does not resolve it, but is about it, we consider it as anaddress info request. For instance, the tutor T’s utterance in (1.3) (Dialogue:min11d) is not an answer according to the distinction resolve vs. about: it doesnot answer the student S’s question but rather refuses to answer it, but it doesaddress the question and realises, thus, a address info request move:

(1.3) S: was ist K ( a )?

T: Das kann ich nicht beantworten

We modify the definition of an answer move as an utterance that complieswith an info-request action in the antecedent and resolves it. An example foran answer move from our corpus is (1.4) (Dialogue: did10p) below, where thestudent S gives an answer to the tutor T’s info request that resolves it:

(1.4) T: Ist das noch derselbe Losungsweg wie in der vorigen Antwort?

S: Nein

1.4.1.3 Task Dimension

The task dimension captures functions that are particular to the task at handand its manipulation, and hence to the genre. It comprises mostly newly definedmoves. These moves require special treatment in the tutorial genre or acquiretheir specific function only due to the genre. They can also only be planned byuse of special modules that deal with domain knowledge (cf. (Benzmuller et al.,2003)) or after consulting information on the current student progress. Movesin this dimension can also be indirect ones. The particular separation of genrespecific moves from genre independent ones, attempted by the expansion of this

10 CHAPTER 1. MOTIVATION

dimension, captures our overall philosophy of an easily reconfigurable dialoguemanager to both other tutoring domains, as well as other dialogue genres.

The task dimension is divided into two sub-dimensions in order to distin-guish between the two parallel tasks that are performed while tutoring, namelythe proof task and tutoring task. In the proof task we define the dialogue movedomain contribution, which is concerned with resolving the domain task for thesession. In our domain the task is finding a mathematical proof to a problem.Example (1.5) (Dialogue: did10d) is a domain contribution, where the studenttries to perform a step in the proof that is expected of him to complete. Con-trast this with Example (1.3), which does not constitute a domain contribution,although they both involve reference to domain knowledge.

(1.5) S: K ( ( a ∪ B ) ∩ ( C ∪ D ) ) = K ( a ∪ B ) ∪ K ( C ∪ D )

The 9 dialogue moves in the tutoring task sub-dimension do not try them-selves to resolve the task. They rather aim at helping the student resolve it and,hence, pertain to the tutoring. This sub-dimension constitutes the frameworkfor the manipulations of teaching strategies.

In Section 1.5, we will come back to a more detailed analysis of and examplesfor the task dimension, after the brief description of all the dimensions, whichwill facilitate the understanding of the concept of the task dimension.

1.4.1.4 Conventional Task Management Dimension

Conventional task-management captures utterances that explicitly address themanagement of the structure of the task. In DAMSL, our tutoring task sub-dimension is grouped in one dimension with conventional task-management,namely task management. As we will explain in more detail later, the tutoringtask is way too important in the genre and it deserves a separate dimensionthat makes the manipulations necessary in it clear. Moreover, although thosemanipulations might be contentious and subject to many changes to cater fordifferent teaching models, conventional task management is much more straightforward and need not be affected by such changes. This makes reconfigurabilityeasier.

This dimension captures the following newly defined dialogue moves: Ini-

tiate task, close task, initiate subtask and close subtask. An example of aclose task is in (1.6) (Dialogue: did15d), which signals the end of the provingtask.

(1.6) T: Damit ist die Aussage Bewiesen

We have also included the moves initiate subtask and close subtask althoughthey did not appear in our corpus, as are already investigating structuring ourtasks, and especially the tutoring task (Tsovaltzi et al., 2004). We would thenneed these moves in order to add the corresponding discourse structure, which isstrictly necessary in order to make it easy for the student to follow the purposeof the subtask better, and with it the task.

1.5. TUTORIAL DIALOGUE: TASK DIMENSION MODELLING 11

1.4.1.5 Conventional Communication Management Dimension

Conventional communication management concerns utterances that explicitlymanage the structure of the dialogue. It comprises the following moves: Ini-

tiate dialogue, Close dialogue, Initiate subdialogue, Close subdialogue, Dis-

course marking. The overall difference between these moves and the respectivemoves in the conventional task management dimension, lies in the distinctionbetween the task management and dialogue management itself. We distinguish,for instance, between sub-dialogues, which may be clarification dialogues aboutthe degree of understanding of the utterance, and tutoring sub-tasks, which areseen more as a deviation from the main line of tutoring, in order to deal witha different aspect before being able to return to the central line again. Such asub-task could be dealing with a student’s misconception, which is not part ofthe tutoring session at hand, but cannot be left uncorrected.

Discourse marking refers to utterances which mark the discourse structure,i.e. indicate how the current utterance relates to the preceding or succeedingcontext. Example (1.7) (Dialogue: did16k) signals the beginning of a previouslyannounced action, namely a recapitulation of the proof.

(1.7) T: Dies mache ich nun.

1.4.1.6 Communicative Status Dimension

Communicative status refers to features of an utterance unit such as whetherit was intepretable. These features mark exceptional cases. Therefore, mostutterance units do not have a function in this dimension (Allen and Core, 1997).It is rather included in the taxonomy for the sake of completeness.

1.5 Tutorial Dialogue: Task Dimension Mod-

elling

The careful expansion of the task dimension can prove extremely beneficial ina difficult genre as tutorial dialogues. The task level moves can be more clearlydefined for use in tutorial dialogues. Such definitions facilitate our understand-ing of the phenomena in tutorial dialogues and their implementation separatelyfrom the general dialogue management moves, and in parallel with them.

We will now have a closer look into how the clear separation of the task levelhas allowed us the formalisation of specific tutorial phenomena.

Before any reasonable feedback can be given to the student, the student’sinput need to be evaluated. There are however two categories of student inputthat are relevant to tutoring. There is input that does not contribute to thetask but carries tutoring relevant information, such as motivation levels. Thisaspect of the input is useful for the genre irrespective of the teaching modelimplemented. For example, it is useful to represent the move resign, i.e., thestudent stating that they do not wish to continue with the task. The way thismove will be treated, can be decided according to the pedagogical model of

12 CHAPTER 1. MOTIVATION

choice. We represent the resign dialogue move in the task dimension, which isrealised, for instance, by the utterance in Example (1.8) (Dialogue: soc5p).

(1.8) S3: Gebe auf.

On the other hand, there is task related input with which the student triesspecifically to handle the task, one way or another. It is characteristic of thiskind of input that it is, first, domain specific and, second, that it needs to beevaluated with respect to its success in bringing the task forward. Again thespecific way in which the evaluation will take place is up to the implementorand dependent on the teaching model assumed and the domain. Nonetheless,the task dimension expansion which we are proposing and the dialogue moveswhich we define in it provide a framework and point to a possible way for thefurther formalisation. More specifically, the dialogue move domain contribution

must be available in order to represent an attempt to bring the task forward.Once this is represented, it is up to the implementor to decide on categories forthe evaluation of this attempt, that is the domain contribution, which serve thetutoring purposes. This decision is both domain dependent (different domainsmight require diverging evaluation methods) and teaching strategy dependent(what information is useful for tutoring may vary between models). Hence, itlies beyond the scope of dialogue management, and the focus of this document.We are, nonetheless, currently formalising such categories and incorporatingthem in feedback managing algorithms, either as domain specific or not, wherepossible. These algorithms are external resources for the dialogue manager(Tsovaltzi and Fiedler, 2003).

Yet a third dialogue move that we define in the task dimension and whichis relevant to the previous two, is the move domain contribution evaluation.This category comprises dialogue moves which signal the domain evaluationof domain contributions, and correspond to the above mentioned evaluationcategories. These dialogue moves are the task dimension counterpart of thecategories in the super-class Agreement in the backward-looking dimension,which refer back to answers other than domain contributions. For example,the backward-looking move reject is reserved for rejecting things at the dia-logue level, eg. like offers. One of the task dimension equivalents, signal reject,can reject an incorrect domain contribution. Notice that the surface realisationof the two moves might well be the same, but the information they capture is dif-ferent. That is, the one is relevant to tutorial dialogues only, whereas the otheris not. signal reject is a super-category. Its sub-categories are the evaluationcategories, which the implementor chooses for their ITS. An example of sucha subcategory in our taxonomy, which is equivalent to the backward-lookingreject is, signal irrelevant. It lets the student know that their answer, correct asit might be in principle in the domain, it does not bring the task forward. Theutterance in Example (1.9) (Dialogue: soc23k) has a signal irrelevant functionat task level.

(1.9) T: Das ist zwar richtig, aber im Augenblick uninteressant.

1.5. TUTORIAL DIALOGUE: TASK DIMENSION MODELLING 13

We have further been able to better formalise the task dimension move hint,which bares a particular importance for the socratic teaching model that we aremodelling. More specifically, we were able to formalise the very complicatedsocratic tutoring strategy, which is most appropriate for tutoring proving, andwith it the hinting process that implements it and focuses on eliciting infor-mation and active learning (Fiedler and Tsovaltzi, 2003a). Thus, we definedhint categories in a hint taxonomy based on the needs in the domain. Our hinttaxonomy captures the underlying function of hints. The latter is mainly re-sponsible for the educational effect of hints. The structure of the hint taxonomyalso reflects the function of the hints with respect to the information that thehint addresses or is meant to trigger.

In order to capture the different underlying functions of a hint in our hinttaxonomy we have defined hint categories across two dimensions. The firstdimension distinguishes between the active and passive function of hints. Thesecond dimension distinguishes between different classes of hints. Each of theseclasses consists of single hint categories that elaborate on one of the attributesof the proof step under consideration.

The same underlying function can be common for different surface realisa-tions. Part of the surface realisation is captured by the forward and backward-looking dimensions and the direct dialogue moves represented there. In otherwords, the hint categories are defined in a taxonomy within the top level di-alogue move taxonomy. This makes it possible to investigate the underlyingfunction of hints, which is itself a very complicated issue, separately from theirsurface realisation, whose function can be captured in the spirit of the morestandard dialogue moves. Example (1.10) (Dialogue: soc21k) is a hint in thetask dimension and it realises a passive relevant-concept hint function, based onour hint taxonomy and according to the needs as they arise from the the currentstudent level. The particular realisation is an open option in the forward-lookingdimension. Given a different dialogue context, however, the same underlyinghint function could just as well be an action directive without altering the hintfunction at the cognitive level, but serving the dialogue context better. This ap-proach is contrary to what has been done before in modelling tutorial dialogues,which do not distinguish between the cognitive and dialogue function of hintsand have, thus, provided no clear account of the status of hints in the proposeddialogue move taxonomies (Person and Graesser, 2003; Core et al., 2002). Thehint categories in the hint taxonomy constitute output of a hinting algorithmwhich is called upon by the dialogue manager (Fiedler and Tsovaltzi, 2003b).

(1.10) T: Sie mussen als erstes die wenn-dann-Beziehung betrachten.

A further advantage of the expansion of the task dimension which we areproposing is that it caters for reusability and easy reconfigurability. The latteris a counter effect of the robust definitions of the functions at that level. Thismakes it possible to define the manipulation of the information captured at thetask level for serving the purposes of any teaching model that might be pre-ferred. At the same time the definition of the task level function as separate

14 CHAPTER 1. MOTIVATION

from the function at the most standard forward and backward-looking levelsclarifies exactly what needs to be manipulated for the tutoring and task (do-main) purposes, as opposed to what can be reused intact since it pertains tothe general dialogue management attributes. The required manipulation of thelatter is not expected to differ for different genres.

Clear distinctions with regard to dialogue move functions, as we have de-liberated, offers itself as a valuable basis for specifying interfaces between thedialogue manager and other modules in a system. Examples of such modulesin our genre may be a task manager, the pedagogical knowledge, the teachingmodel, a user model (cf. (Benzmuller et al., 2003)). For example, the recog-nition of the student input that constitutes a domain contribution at the tasklevel, that is that it tries to advance the task, also means that it must be evalu-ated for the domain knowledge that it demonstrates. Thus, it has to be sent tothe task manager and the evaluation module. On the contrary, other kinds ofstudent input do not need to involve the task manager for their manipulation.Early realisation of this fact allows valuable speed.

In terms of the inter-relations that this dimension has with moves in otherdimensions, it allows us to capture the special status that moves have for thetutorial dialogue genre. The same moves especially in the forward and backward-looking dimensions are highly unlikely to serve the same purpose in other genres.For instance, an understanding query, which usually is read as a simple clari-fication question, often causes the student to reflect upon whether indeed theyunderstand, and functions as a prompt (cf. Example (1.11), did15k) at the tasklevel, potentially motivating the student to input a further contribution. Notethat an utterance is an understanding query only when the speaker does notknow the answer to the query, and can, nevertheless, function as a prompt.

(1.11) T: Wie konnte es weitergehen?

On the whole, capturing genre and domain specific characteristics in one di-mension and separating them from the genre independent characteristics, makesreusability and reconfigurability easier, which is in line with and adds to theoriginal philosophy of the DAMSL approach, which we are following.

1.6 Conclusion and Future Work

We are presenting here a dialogue move taxonomy, which attempts to clearlyseparate the general dialogue management attributes of utterances from themore genre and domain specific ones. We claim that such separation facilitatesa deeper understanding for the appropriate manipulation of the genre specificphenomena and allows their formalisation. That in turn, makes a dialoguemanager built on such principles reusable and easily reconfigurable.

We are currently testing both the taxonomy and our hypotheses about valu-able inter-relations of dialogue moves through an annotation programme. Theresults of the annotations will give us better grounds for asserting the hypotheses

1.6. CONCLUSION AND FUTURE WORK 15

or altering and amending the taxonomy. We can then extract an advanced dia-logue model and specify it for the dialogue management and the manipulationof the complex tutorial dialogue phenomena, as well as for the genre indepen-dent dialogue phenomena. The inter-relations of functions at different levels willalso help us specify the communication protocols between the dialogue managerand other modules, such as the proof manager, the pedagogical knowledge, thehinting algorithm, a user model.

Another thing that we want to test is whether our expectations hold forhuman-computer interaction and especially as far as the student behaviour isconcerned (Shechtman and Horowitz, 2003). For example, diagnostic query,which checks the understanding of the student on a particular piece of informa-tion, is expected to pose the obligation to the student to answer it. However,it might well be the case that this obligation is not effected in the student-computer interaction, due to the change of the social role and the authority ofthe tutor (computer) that it fails to bring about.

16 CHAPTER 1. MOTIVATION

Chapter 2

Annotations

To test the adequacy of the dialogue move taxonomy, we designed an anno-tation scheme for annotating the corpus for dialogue moves. The annotationis currently being carried out by two annotators independently, in order to beable to test the reliability of the taxonomy and draw further conclusions fromdivergence points between the annotators.

2.1 Segmentation

The basic units that realise dialogue moves are sentences and coordinated clauses.In some cases, two or more clauses of one sentence that are coordinated witheach other may perform separate moves in one and the same dimension. Forinstance, (2.1) (Dialogue: soc1p) is segmented into two units, where the firstclause is an assert and the second an info request move, both in the forwardlooking dimension.

(2.1) T: Das ist richtig, aber wie bringt das den Beweis weiter?

However, a complex sentence can also realise one single move. Therefore,basic units can also be grouped together and perform a single dialogue move.The dialogue move recapitulation, for instance, is by nature likely to be realisedin more than one sentences, as it represents the recapitulation of the task, or asubtask, which is quite commonly lengthy. The annotation scheme is designedin such a way that a unit should be tagged with at most one dialogue move perdimension. Failure of this aim will give us insight in possible ammendments forthe scheme.

The corpus had been pre-segmented in basic units before the annotatorsstarted their work. We leave it to the annotators to group basic units togetherduring the annotation, if they feel that they together realise one dialogue move.If the annotators feel that a more fine-grained segmentation than the basic unitsabove is needed, they should contact the creators of the manual.

17

18 CHAPTER 2. ANNOTATIONS

We are using the MMAX annotation tool (Muller and Strube, 2003). MMAXallows annotating utterances at multiple levels, which facilitates finding out theabove mentioned correlations. The tool also allows a straight forward imple-mentation of our dialogue move annotation scheme.

2.2 Procedure

The annotators will go through each dialogue six (6) times. that is one time foreach dimension independently. They will only annotate in a run, only, if there isa function of the utterances in the specific dialogue for the dimension currentlyannotated.

The annotations of the dialogues will be regularly submitted for checking.During that time, the annotators will discontinue the annotation until theyreceive feedback on the already submitted annotations.

The annotation scheme is designed in such a way that a unit is tagged withone dialogue move per dimension. However, there might be cases where theannotators consider two (or more) dialogue moves of one dimension to be appli-cable for tagging one unit. In the unlikely event that the annotators encountersuch cases, they are requested to choose a dialogue and make a note of theirsecond choice.

Note that each utterance could have multiple tags in any aspect depending onhow many functions it simultaneously performs. We can eliminate ambiguitiesin implementation, but should keep it in the annotation. However, we don’twant to discourage multiple annotation within one level in order to be able totest the reliability of the scheme.

This is necessary in order to be able to identify such cases, to find out cor-relations, and possibly change the annotation scheme accordingly, i.e., collapsetwo categories/moves, which tend to co-occur. In practice, when following deci-sion trees, the annotators may come to a point where they cannot decide whichway to go.

Chapter 3

DIALOG Annotation

Scheme

Note: This is not the actual annotation manual. It includes notes that theannotators are not allowed to see for purposes of impartiality.

3.1 Forward-looking Function Dimension

It characterises the effect an utterance has on the subsequent dialogue. Theannotators are allowed to look ahead in the dialogue to determine the effectof an utterance, if it is ambiguous or unclear. In this dimension there are 11classes of dialogue moves, some of them with subclasses. Only direct (explicit)speech acts are annotated.

1. Statement

(a) Assert

(b) Reassert

(c) Other statement

2. Influencing-addressee-future-action

(a) Open option

(b) Action directive

3. Info request

4. Diagnostic query

5. Understanding query

6. Committing-speaker-future-action

19

20 CHAPTER 3. DIALOG ANNOTATION SCHEME

(a) Commit

(b) Offer

7. Conventional

(a) Conventional opening

(b) Conventional closing

8. Apologise

9. Gratitude

10. Signalling emotion

(a) Frustration

(b) Satisfaction

11. Other forward function

3.1.1 Statement

Their primary purpose is to make explicit claims about the world. Intuitive test:if the utterance can be followed by “That’s not true”, then it is a statement.Weak forms of statement such as hypothesising or suggesting that somethingmight be true also belong to this class. Statements can be tagged as assert,

reassert, other statement (Allen and Core, 1997).

3.1.1.1 Assert

Definition The speaker is trying to make the addressee adopt a belief bycommunicating a claim about the world (Allen and Core, 1997).

Obligations 1 to address it, i.e., the hearer has no permission to ignore it2.Any utterance addressing an assert will discard this obligation (Kreutel andMatheson, 2001).

Example

(3.1) T2: (...) Gabe es ein Element x ∈ A ∩ B , dann ware x ∈ A und x 6∈ K( B ) , was ein Widerspruch zur Annahme ist .

1We want to test, if same the obligations hold in human-computer interaction. We had signsthat students ignored the ”computer” (Tsovaltzi and Karagjosova, 2004). Related researchalso exists (Shechtman and Horowitz, 2003).

2Address sth applies when neither of accept, reject and answer apply. The latter entail animplicit address anyhow.

3.1. FORWARD-LOOKING FUNCTION DIMENSION 21

3.1.1.2 Reassert

Definition The same as assert, but the speaker thinks that the claim hasalready been made, and indicates it.

Notes Only statements that are marked as old should be tagged as reassert,since the annotation takes the speaker’s intentions into account, not the objec-tive familiarity status of the utterances.

Reasserts are probably IRUs.

Obligations Same as for assert, namely a reassert poses the obligation toaddress it. Any utterance addressing a reassert will discard this obligation.

Examples

(3.2) T2: (give-away-relevant-concept) Sie mussen zuerst die wenn-dann-Beziehung auflosen(...)T5: (elaborate-domain-object) (...) Wir mussen ja zuerst die wenn-dann-Beziehung auflosen . (...) (soc5k)

3.1.1.3 Other statement

3.1.2 Influencing-addressee-future-action

The primary purpose of this aspect is to influence the addressee’s future non-communicative actions as in the case of requests and suggestions. A rough testfor this is, if the hearer could coherently respond with “I can’t do that”, thenthe utterance’s purpose is Influencing-addressee-future-action. Questions alsobelong to this class if they suggest a course of action in addition to asking aquestion (Allen and Core, 1997). This class is subdivided into open option andaction directive.

3.1.2.1 Open option

Definition An utterance that suggests a course of action or states a possibility(Allen and Core, 1997). Imperatives are normally not open options.

Obligations It poses an obligation to the student to address it. There are noobligation for the tutor.

Examples

(3.3) T5: Zuerst moechte ich, dass Sie ... (did15p)

(3.4) T: Sie muessen nun versuchen, .... (soc12d)

(3.5) T: Diesen Schritt muessen Sie noch naeher erlaeutern. (soc17k)

22 CHAPTER 3. DIALOG ANNOTATION SCHEME

(3.6) T: Vielleicht sollten Sie noch einmal in Ihren Begleitmaterial nachsehen.(soc20p)

(3.7) S5: Ich moechte die antwort wissen. (soc5k) (Resign at task level)

Counter-example Not an Open option, but a Statement:

(3.8) S3: nehmen wir an, dass ... (soc1k)

Relation to other moves in the same dimension

Relation to moves in other dimensions Open option may be a hint in thetask dimension.

Notes

• The obligation is different from (Allen and Core, 1997) where open option

does not put an obligation on the hearer, i.e., the hearer can ignore itwithout any negative effect.

• We distinguish between Open option and Action directive for the sake ofgeneration/realization.3 In the genre, both pose an obligation to addressthem. Hints can be realizes either way, based on other issues, e.g.,howstrongly the tutor feels about the command being followed.

3.1.2.2 Action directive

Definition An utterance that requests an action to be performed, i.e., com-mands, pleas etc.

Obligations The listener is obliged to either perform the requested actionor respond to the request, e.g.,refuse to perform the action (Allen and Core,1997). In other words, he does not have the permission to ignore it (Traum andAllen, 1994). This means in our terms that the listener is obliged to address it(Address action dir).

The action directive can be accepted (the action is then performed) or re-jected (hearer refuses to perform the action, e.g., the student says ’I want to trysomething else’) or otherwise addressed (e.g., the hearer wants the speaker toclarify request).

Examples

(3.9) T1: Bitte zeigen Sie ...

3Staffan Larsson also follows a similar method. He makes generation distinctions forgrounding moves in the hope that they will be correctly interpreted by the human users(Larssom, 2003).

3.1. FORWARD-LOOKING FUNCTION DIMENSION 23

3.1.3 Info request

Definition An info request is an utterance that requests information whichthe speaker does not possess.

Obligations In tutorial dialogues, the student is obliged to answer every tutorquestion, if the question is domain related, i.e., the student is obliged to resolveit or attempt to resolve it.

See also (Kreutel and Matheson, 1999): replying a question involves an obli-gation to answer, i.e., assert, reject (inability to answer) or ask (info request),which is a request clarification at the Backward-looking function.

This does not hold for the tutor who can refuse to answer questions, orignore them. The tutor will really only answer info requests that are either non-domain related (no correspondence to task level) or based on teaching modelconsiderations.

The following example illustrates the obligation:(Soc23p)

(3.10) T4: Glauben Sie , daß Sie das nun beweisen mussen oder ist das eineFolgerung aus dem Obigen?S3: (unknown) A ∩ B ∈ P ( ( C ∪ A ) ∩ ( C ∪ B ) ) = A ∩ B ∈ P ( C ∪ A) ∩ P ( C ∪ B )T5: Bitte beantworten Sie zuerst meine Frage , bevor Sie mit dem Beweisfortfahren.

In T4 the tutor does an info request. The student in S3 ignores the tutor’squestion, so the tutor in T5 states the obligation explicitly, not allowing thestudent to overlook it.

Notes In DAMSL, an info request is any utterance that creates an obliga-tion for the hearer to provide information. It includes all questions: yes/no-questions, WHO-questions, assertive questions, indirect questions (“Tell me thetime”), requests for other actions that provide information (“Show me wherethe city is on the map”).4 However, in tutorial dialogues we want to distinguishbetween questions whose answer the tutor knows, and such whose answer hedoes not know. The former are called diagnostic queries (Core et al., 2002) (See3.1.4). We consider only the latter to be info request.

An info request performed by the tutor initiates a clarification subdialogue.

Examples

(3.11) T: ..., aber wie bringt das den Beweis weiter? (soc1p)

This utterance is an info request and not a diagnostic query, because thetutor really does not know the answer.

Similar example:

4We would rather consider those as action directives.

24 CHAPTER 3. DIALOG ANNOTATION SCHEME

(3.12) T2: ..., aber was hat die Potenzmenge mit diesem Beweis zu tun? (soc1k)5

Relation to other moves in the same dimension Different from diagnos-

tic query, where the tutor knows the answer. above.Different from understanding query, which just asks for a signalling of un-

derstanding.

Relation to moves in other dimensions Different from request clarification,which requests specific information about a previous utterance’s vagueness.

3.1.4 Diagnostic query

Definition An utterance by which ”the speaker is testing whether a listenerknows a piece of information by asking them to supply the information”. Thespeaker, normally the tutor, already knows the answer. They are often ques-tions, but can be also requests: ”tell me how electricity flows through the circuit”(Core et al., 2002).

Obligations Poses the obligation on the student to answer it (addressing itis not enough here).

Examples

(3.13) T: (Sie haben die Regel aber nicht richtig angewendet.) Wie haetten Siediese Regel anwenden muessen? (soc1p)

Counter-example

(3.14) T: Koennen Sie das noch genauer erklaeren? (soc20k) (It is not a Di-

agnostic query, because the tutor does not know what to expect as an an-swer. It is a check origin of problem realised as a request clarification andinto request or open option.

Relation to other moves in the same dimension Diagnostic query differsfrom Info request and Understanding query in that the speaker already knowsthe answer.

Relation to moves in other dimensions Can be hint, align and check

origin of problem in the task dimension.

Notes We distinguish between diagnostic query and info request for annota-tion. We’ll see if this distinction is necessary for the implementation.

5This is indeed a diagnostic query. The tutor knows that Potenzmenge has nothing to dowith this proof. It is a rhetorical question, part of explaining a misconception.

3.1. FORWARD-LOOKING FUNCTION DIMENSION 25

3.1.5 Understanding query

Definition An utterance that asks “the listener whether they understood,rather than making them prove it (Core et al., 2002). It takes a “yes” or “no”as an answer.

Obligations Poses the obligation to address it. It does not pose the obligationto explain more or to prove it.

Examples

(3.15) T5: Verstehen Sie jetzt, wie das mit der Implikation zusammenhaengt?(soc12k)

Relation to other moves in same dimension Difference to diagnostic query:the tutor does not know the answer. The realisation of the two can be the same(”Do you understand?”), but the obligations that they pose are different: diag-

nostic queries require more than ”yes” or ”no” as an answer.Difference to info request: It asks for signalling understanding.

Relation to moves in other dimensions It serves a pedagogical purposewhen it is done by the tutor, i.e., a prompt in the task dimension.

Notes The student can also perform it.

The annotators should look at the student’s next utterance for differentiatingbetween diagnostic query and understanding query.

3.1.6 Committing-speaker-future-action

Utterances which potentially commit the speaker to some future course of action.The commitment can be conditional on the listener’s agreement (offer) or not(commit) (Allen and Core, 1997).

3.1.6.1 Commit

Definition An utterance with which the speaker commits himself to a futurecourse of action. An utterance that accepts an action directive or open option

will typically be a commit (Allen and Core, 1997). “Okay” can be also a com-

mit. The speaker’s commitment does not depend on the acceptance of thecommitment by the hearer, e.g.,as in the case of a promise.

Obligations Commit does not pose an obligation on the hearer to respond,but poses a future obligation on the speaker to perform the action he has com-mitted himself to.

26 CHAPTER 3. DIALOG ANNOTATION SCHEME

Examples

(3.16) T4: Dies mache ich nun. (did16k) It marks the structure of discourse,discourse marking at communication management level.

3.1.6.2 Offer

Definition An utterance by which the speaker indicates willingness to committo an action, if the hearer accepts it (Allen and Core, 1997).

Obligations In the tutorial genre, it does not pose an obligation on the stu-dent to address it explicitly, because he is obliged to accept it. In the same way,the tutor is not obliged to wait for the student to accept it in order to performthe offered action.

If the student makes an offer, the tutor is obliged to address it.There is the possibility that the student rejects an offer of the tutor, e.g.,T:

Do you want me to help you? S: No, I’m ok. - This is also not likely in tutorialdialogues. However, the obligation to address the offer is fulfilled via rejectingit.

Examples

(3.17) T: Let’s look at an example (constructed example)

Notes In (Allen and Core, 1997), Offer poses an obligation on the hearer toaddress it (accept or reject also count as addressing), and on the speaker tobring about the committed action.

3.1.7 Conventional

This category captures the function of utterances as conventional communicativeactions such as greetings and saying goodbye.

Notes In (Allen and Core, 1997) also explicit performatives (“You are fired”),exclamations (“Ouch”), forward-looking functions not captured by the schemesuch as holding/grabbing the turn (“Right”, “Okay”).

Although they are not in corpus, we may need them for the implementation.

3.1.7.1 Conventional opening

Definition An utterance which is a phrase conventionally used to summonthe addressee and/or start the interaction (“Can I help you”, “Hi”) (Allen andCore, 1997).

Obligations None.

3.1. FORWARD-LOOKING FUNCTION DIMENSION 27

Examples Not in corpus.

3.1.7.2 Conventional closing

Definition An utterance which is a phrase conventionally used in a dialogueclosing or used to dismiss the addressee (“Good-bye”) (Allen and Core, 1997).

Obligations None.

Examples Not in corpus.

3.1.8 Apologise

Definition An utterance by which the speaker expresses regret. According to(Searle, 1975), an expressive speech act: the speaker expresses a psychologicalstate or reaction.

Obligations The tutor should address it by saying something like “There isno need to apologise”, for motivation reasons.

Examples

(3.18) S4: (Hab keine ahnung mehr.) Tut mir leid! (min14k).

(3.19) T: Das ist nicht richtig. S: Das tut mir leid. (constructed) (S performs anAddress statement (backward-looking function) and an apologise (forward-looking function))

(3.20) S: Entschuldigung, es gilt natuerlich ... (soc20p) (not a Self correction,because T points S at material and if we take it to be a self correction, thenevery S move after a hint would have to be a self correction. “Entschuldigung”is apologise, the rest is a Statement assert.)

Relation to moves in other dimensions The backward-looking functionis address statement.

Notes It is forward-looking, because it does not arise from a previous obliga-tion/intention.

There is a move Apology in the DARPA 9 Communicator system (Walkerand Passonneau, ) which captures utterances that the system makes to apologisefor misunderstandings, i.e., “I am sorry. I am having trouble understandingyou.”

3.1.9 Gratitude

Definition An utterance that expresses gratitude.

28 CHAPTER 3. DIALOG ANNOTATION SCHEME

Obligations No obligation to address it, but the listener may say “Bitte”.

Examples

(3.21) S: Danke. (As a reaction to encourage) (soc20k S9)

(3.22) Das ist nett. (constructed) (implicit gratitude)

“Danke” cannot be an Assert. “Das ist nett” can.

Notes It is forward-looking, because it does not arise from a previous obliga-tion/intention.

Relation to moves in other dimensions The backward-looking functionis address statement.

3.1.10 Signaling emotion

Notes We use this, instead of Exclamation in (Allen and Core, 1997) such as“Ouch”.

3.1.10.1 Frustration

Definition An utterance which expresses a negative emotion.

Obligations The tutor is obliged to do an encourage (Motivation Theory).

Examples

(3.23) S2: Wenn ich das wuesste! (soc20k) =Ich weiss es nicht.

Relation to moves in other dimensions It can realise a resign move attask level.

3.1.10.2 Satisfaction

Definition An utterance which expresses a positive emotion.

Obligations None.

Examples Possibly not in corpus, because the subjects would express thesethings verbally.

3.1.11 Other forward function

When the speaker is performing an action not captured by any other forward-looking function (Allen and Core, 1997).

3.2. BACKWARD-LOOKING FUNCTION DIMENSION 29

3.1.12 Decision tree

See end of chapter6

3.2 Backward-looking Function Dimension

This dimension captures how the current utterance relates to the previous dis-course. In annotating this aspect, we identify the type of backward-lookingfunction and indicate the previous stretch of discourse that is being affected7

(the antecedent) (Allen and Core, 1997). An antecedent is a dialogue move inthe preceding discourse. Note that the antecedent need not be necessarily inthe immediately preceding discourse, but can be more distant. It can also bethe case that a move has multiple antecedents. In this case, we annotate for allantecedents. Only direct (explicit) speech acts are annotated.

1. Agreement

(a) Accept

(b) Accept part

(c) Maybe

(d) Reject

(e) Reject part

2. Understanding

(a) Signal non-understanding (SNU)

(b) Request clarification

(c) Signal understanding

i. Repeat rephrase

ii. Acknowledge

iii. Completion

(d) Correct misspeaking

3. Answer

4. Information-relation

(a) Address action dir

(b) Address info request

(c) Address statement

(d) Address SNU

(e) Address other

6The decision tree is partly based on the ones provided in (Allen and Core, 1997) and(Core et al., 2002).

7This is necessary for cases like in did15p, T5, where the tutor ignored previous studentutterance and addressed the one before. He indicated this.

30 CHAPTER 3. DIALOG ANNOTATION SCHEME

3.2.1 Agreement

This aspect codes how the utterance unit affects what the participants believethey have agreed to, typically, at the task level, or whatever the topic of discus-sion is. These relations occur in contexts where the one agent has made somekind of proposal, such as a request that the hearer does something, an offer thatthe speaker does something, or a claim about the world. The current utterancethen indicates the other participant’s view of the proposal. An agent may ex-plicitly accept or reject all or part of the proposal, or be non-committal, or leaveit open by requesting additional information, or exploring the consequences.

For instance, 3.24 is not explicitly accepting the student utterance (but isdoing that in the task dimension, see signal accept).

(3.24) T9: Sehr gut! (soc20k)T: Da fehlt etwas. (constructed) (This is accept at task level, because itaccepts what was said and indicates additionally that something is missing.)

An utterance may explicitly accept part, but implicitly reject part of the pre-vious utterance, however this aspect only codes what is explicitly accepted/rejectedby the response (Allen and Core, 1997).

The Annotators should tag accept, reject, answer and address-moves with aflag for the utterance that is their antecedent (see also (Allen and Core, 1997)and (Core et al., 2002)). With MMAX, this will be done by means of pointersfrom the move to its antecedent.

3.2.1.1 Accept

Definition An utterance that accepts a proposal, request, statement or infor-mation request.

Example from (Allen and Core, 1997):A: Can you tell me the time?B: Yes (accept)

Obligations Whenever a forward-looking move is accepted, there is the obli-gation to carry the accepted move through (Matheson et al., 2000). For instance,an accepted offer incurs an obligation to follow it. An acceptance of a suggestionor request to perform an action poses the obligation to attempt to achieve theaction (Traum and Allen, 1994).

Accept, however, does not apply to all forward-looking moves. Moreover, theacceptance of an Assert does not pose any obligations to perform any actions.

Examples

(3.25) T4: Der ansatz ist richtig. (soc2k)

3.2. BACKWARD-LOOKING FUNCTION DIMENSION 31

3.2.1.2 Accept part

Definition An utterance that accepts part of a proposal, request, statement orinformation request. Implicitly, it rejects another part of the utterance, but weonly code what is explicitly accepted. When an utterance does both explicitly,it should be segmented it into two units and labelled accordingly.

Obligations Same as for accept for the part that is accepted.

Examples

(3.26) T: P(C)∪P(A∩B)⊆P(C∪(A∩B)): das ist richtig! A∩B⊆P(A∩B): das istnicht richtig! (min11p T7)

3.2.1.3 Maybe

Definition An utterance with which the speaker ”explicitly states that hecannot give a definite answer at the moment” (Allen and Core, 1997).

Obligations None.

Examples

(3.27) I’ll have to think about it. (constructed example)

Notes Relates to offer. In our domain there is not likely to be a maybe,because it is clear that one has to think.

3.2.1.4 Reject

Definition An utterance that rejects a proposal, request, statement or infor-mation request.

It says nothing positive about the antecedent (possibly indicating error). Itcan be implicit or contain ”no” or “not” (Core et al., 2002).

Obligations The student is obliged to justify the rejection or elaborate on it.The tutor might signal which part is problematic (e.g., via domain contri-

bution evaluation). This depends on the teaching model and on the possibilityof realising the domain contribution evaluation.

Examples

(3.28) T3: Das ist keine aussage (did18d) (signal ill formed)

(3.29) T: Das kann man nicht so folgern (soc13k)

(3.30) T: Das ist nicht der richtige Weg (soc1k)

32 CHAPTER 3. DIALOG ANNOTATION SCHEME

(3.31) T: Was Sie geantwortet haben, ist nicht eindeutig. (soc12p)

(3.32) T3: Das ist keine vollstaendige aussage (did15p)

(3.33) T6: Das ist unvollstaendig oder nicht ganz korrekt. (min11d, also min11cT8): (2 moves, Reject and Reject part, shouldn’t be annotated, becauseonly in minimal feedback condition)

(3.34) T: Das ist kein vollstaendiger ausdruck. (did19p)

(3.35) T: Sie haben die Regel aber nicht richtig angewendet.(soc1p)

Relation to moves in other dimensions Reject can have the flag of thetype of domain contribution that is being rejected. This information will comefrom the task dimension move domain contribution evaluation.

Note: MPT say that when an offer is rejected, the person who rejects it isobliged to propose a counter-offer (Matheson et al., 2000). We want to considerif this is true in the tutorial dialogues genre.

”No” is very seldom used, and only when the student is totally wrong.

3.2.1.5 Reject part

Definition An utterance that rejects partly a proposal, request, statement orinformation request. It implicitly accepts another part of the utterance, but weonly code what is explicitly rejected.

Obligations Same as for reject for the rejected part.

Examples

(3.36) T3: Das ist nicht ganz richtig (did4p)

(3.37) T7: P(C)∪P(A∩B)⊆P(C∪(A∩B)): das ist richtig! A∩B⊆P(A∩B): dasist nicht richtig! (min11p)

3.2.2 Understanding

3.2.2.1 Signal non-understanding (SNU)

Definition An utterance that signals that the speaker has not understood theprevious utterance, i.e., did not hear it or could not make sense of it.

Instances for that move are “I don’t understand” and variants like “Whatdid you say?”.

Obligations SNU poses an obligation to address it (E.g., via a clarification).

3.2. BACKWARD-LOOKING FUNCTION DIMENSION 33

Examples

(3.38) Ich verstehe Ihre Frage nicht (did15c)

(3.39) Was wollen Sie damit sagen? (soc17k)

(3.40) Was meinen Sie? (soc17d)

Relation to other moves in the same dimension The difference to re-

quest clarification is following: An utterance is an SNU, when interlocutor didnot understand anything of the preceding contribution, and a request clarification

when, they understood something of the preceding contribution, but not all ofit (Core et al., 2002).

Relation to moves in other dimensions An SNU can have the forward-looking functions statement or info Request.

Notes They are “utterances that explicitly indicate a problem in understand-ing the antecedent”. An applicability test is the rough paraphrase “What didyou say/mean?” (Allen and Core, 1997). We have narrowed down this definitionto cases like “I don’t understand” and variants. (Allen and Core, 1997) pointsout that not all clarification questions should be labeled as SNU, they could bee.g.,hold.8

An SNU can be addressed (address SNU) or, when SNU is a question, andcan be answered (the response should be labeled as answer, and if not resolved,as address SNU).

SNUs introduce clarification subdialogues. An answer to it addresses attask level the previous utterance and at dialogue level the move with which thesubdialogue was initiated.

For realisation purposes we distinguish between SNU and request clarification.The formulation of one or the other can prove useful for letting the collocutorknow how much they need to clarify.

3.2.2.2 Request clarification

Definition A move which applies when some of the input has been understood(Core et al., 2002). It has basically the form of questions, but can be also animperative i.e., it can be an action directive in its forward-looking function..

Obligations It poses an obligation to address it (E.g., via a clarification).

8Hold in DAMSL is meant as a reaction to proposals. It is not clear what it is, but seemsto be an attempt to have an all encompassing move, instead of accounting for subdialogues.This is not sufficient for us.

34 CHAPTER 3. DIALOG ANNOTATION SCHEME

Examples

(3.41) T2: Was meinen Sie mit F (F a formula) (did15p)

(3.42) Meinten Sie wirklich F? (did16p)

(3.43) Was soll das x darstellen? (soc17p)

(3.44) T2: Was soll das heissen (did19c)

(3.45) T5: Ist das die antwort auf meine Frage oder ein neuer Loesungsversuch?(soc21p)

(3.46) Bitte erklaeren Sie Ihren Schritt genauer! (soc12p)

Relation to moves in the same dimension SNU applies when none of theinput has been understood.

Notes There is no move request clarification in DAMSL.Request clarification are only questions in (Core et al., 2002).

3.2.2.3 Signal understanding

An utterance which signals understanding. Any utterance that does not explic-itly signal non-understanding implicitly indicates understanding, so there is noneed to tag such cases (Allen and Core, 1997).

• Repeat rephrase

Definition This move is “...used for utterances that repeat or para-phrase what was just said in order to signal that the speaker has beenunderstood...[they] do not necessarily make any further commitment as towhether the responder agrees with or believes the antecedent.” (Allen andCore, 1997)

Obligations None.

Notes It is probably not useful for us. It is difficult to implement andin nature really an acknowledge.

These are Attitude Acceptance IRUs.

• Acknowledge

3.2. BACKWARD-LOOKING FUNCTION DIMENSION 35

Definition ‘Acknowledgements are utterances consisting of short phrasessuch as “ok”, “yes”, “uh-huh”, that signal that the previous utterancewas understood without necessarily signaling acceptance’ (Allen and Core,1997). They do not resolve the content of the utterance they address.9

Obligations When the student addresses the tutor, the tutor is obligedto acknowledge, even if it is only to reject the offer to talk about the issuethat the student has raised. The preferred way of doing the latter is in anexplicit manner that would be to condescending in other genres: ”Beforewe get to that,”(p. 8), this implicitly acknowledges it and is a reject.

An explanation for that is that, because the student does not give directanswers, he must indicate somehow that he is taking the students answerinto account, rahter than just ignoring it (Tsovaltzi and Matheson, 2002).

However the student is not obliged to do any acknowledge. Sometimesthey don’t do any implicit acknowledging either. They are just silent andassume that the tutor knows that they are listening and thinking aboutthe problem. Otherwise, the student would ask for clarification.

An explanation for that is that both student and tutor are aware of theobligations that ther respective social roles bring. Since it is the student’sobligation to take what the tutor says into account (expertise plays a role)in order to proceed with the task, the student does not feel that they needto indicate that that is what they are doing (Tsovaltzi and Matheson,2002).

Examples

(3.47) S5: schon klar. (soc20k)

Relation to other moves in same dimension Observation from theBE&E corpus: It might well be the case that ”yes”, ”OK” and ”right” areaccept. In the case they are accept, the tutor can move on with the taskand does not insist on eliciting the correct domain contribution.

”Good”, “very good”, ”that’s right” serve always as accept. Whereas”right”, ”alright” and ”ok” are ambiguous. They are commonly only an

9We distinguish the aspects resolvedness vs aboutness of answers to questions, follow-ing (Ginzburg, 1996). Resolvedness and aboutness are relations between questions and an-swers. An answer resolves a question when the answer provides information that positivelyor negatively resolves the question (Larsson, 2002). ’About’ is a relation that “captures therange of information associated with a question independently of factuality or level of detail.”(Ginzburg, 1994). An answer move always resolves the question. If the question is followedby an utterance which does not resolve it, but is about it, we tag it as address info request.For instance, “T: Das kann ich nicht beantworten” (min11d T2, min11k T6) is not an answer

according to the distinction resolve vs about: it does not answer, but rather refuse to answera question, it does nevertheless address the question (address info request).

36 CHAPTER 3. DIALOG ANNOTATION SCHEME

acknowledge when they are followed by a hint. ”OK” is commonly usedafter a not totally wrong answer.

It is up to the annotators to decide when an expression serves as an accept

or as an acknowledge.

Notes There is only a realisation difference from Repeat rephrase. Namely,acknowledge is only short responses like “ok”, “yes”, “aha”.

When the tutor does a check or problem, there is normally no acknowledge

preceding it.

• Completion

Definition An utterance that shows ”understanding by finishing or addingto the clause that a speaker is in the middle of constructing” (Allen andCore, 1997).

Notes Not possible to implement at the moment, as we do not allowbarge-in.

3.2.2.4 Correct misspeaking

Definition They are “... utterances that by offering a correction indicatethat the hearer believes that the speaker has not said what he or she actuallyintended” (Allen and Core, 1997).

It applies only to cases where the current speaker makes a correction to whatwas previously uttered, that is, to the utterance that is addressed by the currentcorrect misspeaking.

There is no dimension for annotating self corrections currently in DAMSL.

Obligations None.

Examples S: A is a sub-category of B. (where ’sub-category’ is wrong, non-existent terminology)T: That’s right, A is a subset of B. (correct misspeaking) But why? (con-structed)

3.2.3 Answer

Definition An utterances that complies with an info-request action in theantecedent and resolves it.10

Obligations None.

10See footnote on aboutness vs. resolvedness.

3.2. BACKWARD-LOOKING FUNCTION DIMENSION 37

Examples

(3.48) T3: Ist das noch derselbe Losungsweg wie in der vorigen antwort ?Info request

S3: (correct) Nein Answer(T3) ich habe mich umentschieden : Ich zerlegejetzt die Potenzmenge : P ( C ∪ ( a ∩ B ) ) ⊇ P ( C ) ∪ P ( a ∩ B ) (did10p)

Relation to moves in other dimensions Answers will always be Asserts

Forward-lookingly (even if they are an imperative, since the forward lookingfunction is to provide information, not to influence future action) (Allen andCore, 1997).

Notes S can answer T question by a question: soc1k T6-S6.A speaker can answer his own question, e.g.,did22k T7-T8, min9c S2.

3.2.4 Information-relation

This category “captures how the content of the current utterance relates to thecontent of its antecedent” (this category is not further elaborated in DAMSL,but is left for future study) (Allen and Core, 1997).

We see the utility of this category in its connection to the theorem proverand/or in disambiguating between what a current utterance and its relation tothe overall context is, eg., in case of mediation of a subdialogue.

The annotators should add flags for the antecedent, that is, create pointersin MMAX.

As pointed out before, address sth applies when neither of accept, reject andanswer apply. The latter entail an implicit address anyhow.

The following categories are all new, defined due to the DIALOG corpus.

3.2.4.1 Address action directive

Definition Any utterance that addresses a previous (not necessarily in imme-diately preceding turn) action directive move, but is not a reject, reject part,accept or accept part.

Obligations None.

Examples

(3.49) T1: Bitte zeigen Sie : K ( ( a ∪ B ) ∩ ( C ∪ D ) ) = ( K ( a ) ∩ K ( B )) ∪ ( K ( C ) ∩ K ( D ) ) ! Action directive

S1: (correct) K ( ( a ∪ B ) ∩ ( C ∪ D ) ) = K ( a ∪ B ) ∪ K ( C ∪ D )Address action directive(T1) (did10d)

3.2.4.2 Address info request

Definition Any utterance that addresses a preceding info request move with-out resolving it.

38 CHAPTER 3. DIALOG ANNOTATION SCHEME

Obligations None.

Examples

(3.50) S1: was ist K ( a ) Info request

T2: Das kann ich nicht beantworten . Address info request(S1)

3.2.4.3 Address statement

Definition Any utterance that addresses a preceding statement, assert orreassert, without being an explicit reject, reject part, accept or accept part.

Obligations None.

Examples

(3.51) S6: das stimmt schon. verstehe die definition nicht, einfaches Beispielwuerde mir weiter helfen Statement

T7: Sei die Menge X={1, 2}. Dann ist die Potenzmenge von X die MengeP(X)= {0, {1}, {2}, {1, 2}}. (did15p) Address statement(S6)

Relation to moves in other dimensions In the task level, such questionscan be (pragmatic) hints, e.g.,elicit discrepancy, or a check origin of problem

Notes We use this tag instead of followup which is a backward-looking movein BE&E, e.g.,“Why do you think that?”, “How are you going to do that?”.followup is defined as a reaction of the tutor to a student answer in form of afollowup question asking for more details: “Asking a question about an answer”(Core et al., 2002). Examples from WoZ corpus: “S2: F. T3: Warum?” (soc13k).In WoZ corpus the followup is asking for the next step in proof.

3.2.4.4 Address SNU

Definition Any utterance that addresses a previous SNU.

Obligations None.

Examples

(3.52) T5: Ich verstehe Ihre antwort nicht , denn das ist kein vollstandigerdeutscher Satz : Den Durchschnitt der Menge K ( a ) und der Menge K (B ) , also K ( ( a ∪ B ) ∩ ( C ∪ D ) ) SNU

S4: (wrong) Ist x der Durchschnitt der Menge K ( a ∪ B ) und der MengeK ( C ∪ D ) , dann ist K ( ( a ∪ B ) ∩ ( C ∪ D ) ) Address SNU(T5)

3.3. TASK DIMENSION 39

3.2.4.5 Address other

Definition Utterances addressing other moves apart from the ones definedabove and apart from reject, reject part, accept or accept part, which also ad-dress utterances. E.g., address request clarification, address diagnostic query,address understanding query, address open option, address commit, address

offer, address emotion, address apologise, address gratitude, address agreement,address signal understanding, address correct misspeaking, address answer.

Notes Only after exhaustive search for other backward-looking moves shouldthe annotators use these tags.

3.2.5 Decision tree

See end of chapter11

3.3 Task Dimension

This dimension captures functions that are particular to the task at hand and itsmanipulation. It comprises moves that require special treatment in the tutorialgenre or moves that only acquire the specific function in the genre. These movesare tagged for also when they are not explicitly signalled, i.e., they can be alsoimplicit.

Note: The moves in this dimension can only be planned by use of theProof Manager (PM) or after consulting the current student progress (hintingsession status: HSS). To capture this difference, we enumerate the moves in thisdimension as members of two super-categories: proof task vs. tutoring task. Ifit turns out through the annotation that there is enough cross super-categoriestagging, we will consider making two dimension out of the two super-categoriesin the task dimension.

Proof Task

1. Domain Contribution

Tutoring Task The list of the dialogue moves in this dimension alsoincludes preliminary indications (in brackets) about communication betweenmodules, necessary for the deliberation of each move.

1. Hint (Hinting Algorithm: HA, HSS and PM)

2. Domain contribution evaluation (HA, HSS and PM)

(a) Signal accept (correct answer)

11The decision tree is partly based on the ones provided in (Allen and Core, 1997) and(Core et al., 2002).

40 CHAPTER 3. DIALOG ANNOTATION SCHEME

(b) Signal reject (incorrect answer)

i. Signal unknown

ii. Signal missing basic knowledge

iii. Signal wrong linguistiv term

iv. Signal time out

v. Signal wrong

vi. Signal irrelevant

vii. Signal stepsize

viii. Signal misconception

ix. Signal ill formed

x. Signal complete partially accurate

xi. Signal complete inaccurate

xii. Signal incomplete accurate

xiii. Signal incomplete partially accurate

xiv. Signal near miss

xv. Signal other

3. Check origin of problem (both HSS and PM; e.g., disallow after so manynon-correct answers)

4. Request evaluation (HA, HSS and PM)

5. Prompt (HA, HSS)

6. Encourage (HA, HSS)

7. Request assistance (HA, HSS and PM)

8. Resign (HA, HSS and PM)

9. Align (HA, HSS)

Proof Task

3.3.1 Domain contribution

Definition An utterance that attempts to directly resolve the proving task ora proving task related issue to the best of the speaker’s capacity. It counts assuch even if wrong or irrelevant.

Obligations The tutor is obliged to acknowledge it and also perform one ofthe domain contribution evaluation dialogue moves.

3.3. TASK DIMENSION 41

Examples

(3.53) S2: (correct) Wegen A ⊆ K ( B ) = A ∩ B = ∅ ( obiger Schritt ) undanalog B ⊆ K ( A ) = B ∩ A = ∅ und A ∩ B = ∅ = B ∩ A ( offensichtlich) folgt die Behauptung A ⊆ K ( B ) = B ⊆ K ( A )

(3.54) T: Is A a subset of B?S: yes (= A is a subset of B) (constructed example)

Counter-examples

(3.55) S5: Habe Probleme mit Potenzmenge... (did15p): not a domain con-

tribution although it mentions powerset; it can be handled by DM withoutconsulting the PM.

(3.56) S: What is a K(A)? (constructed example) Not a domain contribution,but a request assistance, but has to be sent to the PM anyway.

(3.57) Hint: give-away-relevant-concept) Das ist nicht richtig! Sie mussen alserstes die wenn-dann-Beziehung betrachten. (soc23k) This is a hint, as thetutor is not advancing the proof task to the best of their ability. Thatwould mean to give the whole proof. They are, however, contributing tothe tutoring task by giving this piece of information and hoping to effectlearning.

Notes The tutor does not contribute to the task, but to the tutoring process!It probably contains domain knowledge, but this is not a decisive factor

either way.The fact that the initiative might lie with the other dialogue participant is

irrelevant here.Any other utterance from the student that does not contain domain knowl-

edge, but needs to be evaluated (eg. T: Is A a subset of B S:yes) would take adifferent flag for what counts as an answer to it, (in the example yes, no).

Originally, we expected that only student utterances will be annotated asdomain contributions, because we see a qualitative difference between domain

contributions and tutor utterances that advance the task. However, becauseof the design of the experiment through which we collected the data, there aredomain contributions by the tutor in the corpus. Namely, when the tutor at theend of a tutoring session gives the whole proof away. That is, he is contributingto the task to the best of his ability.

Tutoring Task

3.3.2 Hint

This dialogue move was already annotated for during experiment. For every-thing that is not annotated yet, it can be annotated as hint, if it is consideredas such.

42 CHAPTER 3. DIALOG ANNOTATION SCHEME

Definition A hint is a tutorial strategy that aims to encourage active learn-ing. It can take the form of eliciting information that the other interlocutor isunable to access without the aid of prompts, or information which he can ac-cess but whose relevance he is unaware of with respect to the problem at hand.Alternatively, a hint can point to an inference that the other interlocutor isexpected to make based on knowledge available to him, which helps the generalreasoning needed to deal with a problem (Hume et al., 1996).

The initiation of hints can be due to various reasons:

(i) the speaker (tutor) observes that the other interlocutor (student is notmaking any progress in the task, or that he is taking steps in the wrongdirection

(ii) the other interlocutor (student) asks a question and the speaker (tutor)does not want to answer it directly

(iii) the other interlocutor (student) gives the wrong answer or asks the wrongquestion in response to a speaker’s (tutor) question

Obligations Partial answers, in the form of hints, from the tutor discharge theobligation to address the student’s questions or utterances. There are examplesin the BE&E corpus where the tutor explicitly states her method or the studentshows he is aware of it, such as ”Very good. You answered your own question”or ”I’ll give you another hint.”, from the tutor and ” I need another hint”,from the student. Statements like these constitute additional support for theobligations involved. They show the dialogue participants’ awareness of theirrespective roles in this special social context (Tsovaltzi and Matheson, 2002;Hulstijn, 2003).

Examples (soc23k)

Tutor (1): Bitte zeigen Sie : Wenn A ⊆ K(B), dann B ⊆ K(a)!

Student (1): (wrong) A ⊆ B

Tutor (2): Das ist nicht richtig! (Hint: give-away-relevant-concept) Sie mussen als erstesdie wenn-dann-Beziehung betrachten.

Student (2): (wrong) A ⊆ K(K(a))

Tutor (3): Das ist zwar richtig, aber im augenblick uninteressant. (Hint: elaborate-

domain-object) Wissen Sie, wie sie die wenn-dann-Beziehung behandeln mussen?

Student (3): (wrong) Nein

3.3. TASK DIMENSION 43

Tutor (4): (Hint: explain-meta-reasoning) Sie mussen die Beziehung auflosen, indem Siedie Gultigkeit der Voraussetzung annehmen und daraus die Gultigkeit der Behaup-tung folgern.

Student (4): (unknown) A ⊆ K(B) ≡ B ⊆ K(a)

Tutor (5): (request clarification) Was wollen Sie mit dem Zeichen ≡ ausdrucken?

Student (5): (wrong) A ⊆ K(B) daraus folgt B ⊆ K(a)

Tutor (6): Das ist genau das, was Sie beweisen sollen. (Hint: give-away-performable-step)Wir setzen nun voraus, daß A ⊆ K(B) gilt. Daraus wollen wir dann schließen, daßdie Behauptung B ⊆ K(a) gilt. Verstehen Sie, warum wir annehmen mussen, daßA ⊆ K(B) gilt?

Student (6): (wrong) Nein

Tutor (7): (Hint: explain-meta-reasoning) Wir mussen dies annehmen, weil wir die wenn-dann-Beziehung auflosen wollen. Bei einer solchen Beziehung muß man stets zeigen,daß aus einem bestimmten ausdruck ein anderer ausdruck folgt. Deshalb nehmenwir zuerst die Gultigkeit des ersten ausdrucks an, weil wir sonst keine Basis hatten,von der wir ausgehen konnen. Verstehen Sie, warum wir zuerst die wenn-dann-Beziehung auflosen wollen?

Student (7): (wrong) Nein

Relation to other moves in the same dimension A hint is attemptingto resolve the proving task or proving task related issues indirectly and it char-acteristically not to the best of the speakers ability. This is the difference to adomain contribution.

3.3.3 Domain contribution evaluation

This category comprises dialogue moves which are the task dimension counter-part of the categories of accept, accept part, reject and reject part (backward-

looking).This move belongs to the tutoring task rather than the proof task, since this

DM is actually not necessary for the Proof Task at all, as the tutor might chooseto do another move instead of it, based on the tutoring strategy followed.

This aspect is reserved for moves that are depending on the answer categori-sation scheme. We annotate the categories of accept, accept part, reject andreject part (backward-looking) with the special kind of move they are in thetask dimension.

44 CHAPTER 3. DIALOG ANNOTATION SCHEME

The annotators will be asked to annotate for these, but if they find it hardto do it, we can deal with them at a later stage. We need them in order to filterout the relevant ones for the generation in the end. The Categorisation modulecan provide the information, and the generator can have heuristics for makinguse of them when/if it signals the relevant information.12

Relation to moves in other dimensions These moves may be accompaniedby a reject, reject part, accept or accept part at the level of backward-lookingfunctions.

The obligations are taken care of at the level of the backward-looking func-tions.

• Signal accept

Definition Signal accept is the task-level counterpart of accept whenwe are accepting a domain contribution, vs accepting an offer, which doesnot have to do with the domain/task.

Examples

(3.58) Das ist richtig!

• Signal reject The task-level counterpart of reject when the speaker (thetutor) is rejecting a domain contribution. On this level, the tutor alsosignals the reason for rejecting the domain contribution:

– Signal irrelevant

Definition An utterance that signals that the domain contributionof the student is not relevant for the proof/task (at the particularstage or at all). The domain contribution can be correct or incorrect.

– Signal complete partially accurate

Definition An utterance that signals that the domain contributionis complete with some accurate parts, but not all of the parts areaccurate.

– Signal complete inaccurate

Definition An utterance, which signals that the domain contribu-tion is complete, but all parts in it are inaccurate.

– Signal incomplete accurate

12One such case might be as a response to request evaluation

3.3. TASK DIMENSION 45

Definition An utterance that signals that the domain contributionis incomplete, and all present parts are accurate.

– Signal incomplete partially accurate

Definition An utterance that signals that the domain contributionis incomplete, but some of the present parts are accurate, howevernot all of the present parts are accurate.

– Signal near miss

Definition An utterance that signals that the domain contributionis very close to the expected correct answer, albeit, not completelycorrect.

– Grain of truth

Definition An utterance that signals that the domain contributionis far from correct, but there is some kind of correct information init, if very minimal.

Relation to moves in other dimensions It can be a reject orreject part move in the backward-looking dimension.

Examples

(3.59) Das bringt uns nicht weiter im Beweis (soc17p)

(3.60) Das bringt auch nichts (soc17p)

(3.61) T8: Das ist nicht die Begruendung fuer Ihre Behauptung!(did22k) (example is ambiguous when out of context)

(3.62) T2: Die anwendung der Potenzmenge bringt fuer diesen Be-weis nichts (soc21k T2).

(3.63) T3: Das ist zwar richtig, aber im Moment uninteressant. (soc23k)

– Signal stepsize

Definition An utterance that signals that the domain contributiondoes not have the expected granularity, i.e., it makes too big or toosmall a step in the proof.

Examples

(3.64) T5: Das ist zu zeigen (did19c)

– Signal misconception

46 CHAPTER 3. DIALOG ANNOTATION SCHEME

Definition An utterance that indicates that the domain contribu-tion was based on a misconception and which points at what themisconception is.

Examples

(3.65) T2: Sie koennen nicht einfach voraussetzen, dass C die leereMenge ist. (soc5p)

– Signal ill formed

Definition An utterance that signals that the (domain) contribu-tion of the student was ill-formed.

Examples

(3.66) T3: Das ist keine aussage (did18d)

• Signal other

There may be utterances that signal other kinds of rejecting or acceptinga domain contribution. If the annotators feel that more such moves areneeded, they should contact the creators of the annotation scheme.

The above categories include more than what the tutor in the WOz ex-periment annotated. Since there is not necessarily a one to one corre-spondence between the formal definition of a category and its possibleverbalisation, or what the tutor decides to signal, we include here all theformal categories, hence allowing the annotators’ interpretation of whatis being signalled. In analogy, we even include non-formalised categories,which might be more intuitive for the annotators to understand and cap-ture. This might give us a correspondence between intuitive definitionsand formal definitions.

We might need to define more of these moves, depending on the finalcategorisation scheme and the categories that will appear in it.

3.3.4 Check origin of problem

Definition A question that the speaker (tutor) asks in order to find out if andwhat problem the hearer (student) has. It does not advance the proving task.

Obligations The student has to address it (taken care of at the forward-looking

level).This move discharges the obligation of the tutor to address a preceding

question of the student.

3.3. TASK DIMENSION 47

Examples

(3.67) T: Koennen Sie das noch genauer erklaeren? (soc20k) (check origin of

problem realised as a request clarification).

(3.68) T: What is it that you do not understand? (constructed example)

Relation to moves in other dimensions The forward-looking function canbe info request, understanding query or offer.

In the backward-looking dimension, it could be realised by nothing, in case itis produced after time-out, and by address sth, where sth stands for whateverDM a previous resign was realised by, when the check or problem follows a re-

sign, commonly by address SNU . It can even be realised by request clarification.

3.3.5 Request evaluation

Definition An utterance with which the speaker (the student) asks explicitlyabout his progress, e.g.,“How am I doing?”, “Was that right?” (constructedexamples). “Why was that not right?” is not a request evaluation, but re-

quest assistance

Obligations The tutor has to address it definitely and probably answer italso by, e.g., either an encourage (”You are doing fine”), a domain contribution

evaluation (”Your answer is mainly correct, but ...”) or a combination of thetwo (”You are doing fine, but you have to concentrate on the ...”).

3.3.6 Prompt

Definition An utterance that requests the hearer explicitly to proceed withthe task and provide further information.

Obligation The student is obliged to address it, by trying to proceed withthe task.

Examples

(3.69) T: Wie geht es weiter? (did15d)

(3.70) T: Wie koennte das gehen? (did16k)

(3.71) T: Wie koennte es weitergehen? (did15k).

Relation to moves in other dimensions It can be realised as diagnos-

tic query or an info request in the forward-looking dimension (e.g., “Wie gehtes weiter?”, depending on whether the tutor knows the answer to the question)and encourage in the task dimension.

48 CHAPTER 3. DIALOG ANNOTATION SCHEME

Notes It is usually the tutor who would utter prompts. For the student suchan utterance will be a resign at the task level and info request forward-looking.

3.3.7 Encourage

Definition The speaker (tutor) gives some positive feedback to the hearer(student) in order to encourage them to continue. It does not presuppose thatthe speaker accepts the domain contribution as in (3.72), but it might, as in(3.73).

Obligations None.

Examples

(3.72) T[1]: ...What are the instructions asking you to do?S[2]: Make a circuit between the source (battery) and the resistance

(rheostat) and then attach the miliampmeter to measure the re-sistance in ohms.

T[3]: OK, you’re close (encourage). But keep in mind that the miliamp-meter is a special case of an ammeter. Do you remember what anammeter measures?

(3.73) Das ist vollkommen richtig! (soc17d) Where the adverb “vollkommen”indicates that it is an encourage.

(3.74) T9: Sehr gut! (soc20k)

(3.75) T4: Der ansatz ist richtig (soc2k)

Relation to moves in other dimensions Most utterances that are an ac-

cept as a backward-looking function will be an encourage, e.g.,(3.74).It is probably always a statement in the Forward Looking dimension, but it

can also be a satisfaction.

Notes: This was moved from forward-looking dimension, as , for certain cases,it might only acquire this task function in the particular tutorial genre. TheDM might need to consult the session status for when to produce an explicitencourage.

The annotators must decide for themselves if the dialogue move performed isan encourage or a domain contribution evaluation. If they cannot, they shouldannotate for both.

3.3.8 Request assistance

Definition An utterance with which the speaker requests assistance in or-der to be able to proceed with the task. It concerns only specific informationrequests, eg. about concepts. It does not include request for the following step.

3.3. TASK DIMENSION 49

Obligations The tutor is obliged to address it , but will really only answer

request assistance, based on teaching model considerations.In any other case, the tutor’s obligation is to encourage.

Examples

(3.76) Das verstehe ich nicht. (soc13p, soc1p) (also an SNU backward-looking)

(3.77) S2: Ich habe die aufgabenstellung nicht verstanden. (soc5k) (also anSNU backward-looking)

(3.78) S1: Ich kann nicht anfangen. (soc5k)

(3.79) S: ”What is a P?” request assistance and info request

Relation to moves in other dimensions The forward-looking function isinfo request or statement or action directive.

Notes It is the student who requests assistance.“I don’t know” (= I cannot answer) is a reject (and statement) in DAMSL,

not an answer, whereas in (Core et al., 2002) it is an answer, because accordingto the speaker of the answer they are answering the question. But accordingto what we have in mind as a definition, an answer is resolving a question andis not just about a question, so it is address info request. So, ”I do not know”effect utterances are not answers, but rather resigns.

3.3.9 Resign

Definition An utterance with which the speaker (the student) indicates thathe gives up.

Obligations The tutor has to address it and try to make the student continueby an encourage every time. The rest is dependent on the teaching model.

Examples

(3.80)

(3.81) S3: Gebe auf. (soc5p)

(3.82) S: Habe keine ahnung mehr (min14k)

(3.83) Keine Loesung mehr vorhanden (min9k)

(3.84) Keine ahnung mehr, wie der Nachweis korrekt erbracht werden kann(min9p)

(3.85) S7: Komme nicht weiter (did15p)

50 CHAPTER 3. DIALOG ANNOTATION SCHEME

(3.86) S5: Ich moechte die antwort wissen. (soc5k) (resign realized as Open

option)

(3.87) S2: Wenn ich das wuesste! (soc20k) =Ich weiss es nicht.

Relation to other moves in the same dimension It is different fromrequest assistance, because the student states that they are giving up, one wayor the other. When the student asks for the following step it also means thatthey are also giving up.

3.3.10 Align

Definition An utterance with which the person who knows more tries to findout if/get evidence that the interlocutor shares indeed the same knowledge, asthe speaker believes they should.

Obligations Imposes the obligation on the student to answer it (taken careof at the forward-looking function), which is diagnostic query. Therefore, ”yes”does not discharge the obligation to answer it.

Examples

(3.88) T6: ... Warum haben wir zuerst die wenn-dann Beziehung betrachtet?(soc21k)

This is an align when the tutor is trying to re-elicit that information afterhaving given it once.

Relation to moves in other dimensions In the forward-looking dimensionalign is probably always a diagnostic query.

3.3.11 Decision tree

3.4 Conventional Task Management Dimension

This dimension captures utterances that explicitly address the management ofthe structure of the task. We tag also implicit dialogue moves in this dimension.

Note: In (Allen and Core, 1997), this dimension captures utterances that“explicitly address the problem solving process and experimental procedure”,including utterances that involve coordinating the activities of the interlocutors(“Let’s work on getting the train to Avon first”), asking for help on the pro-cedures (“Do I need to state the problem?”) or asking about the status of theprocess (“Are we done?”). An example from our domain would be “Let’s lookat an example”. DAMSL distinguishes between utterances that are part of thetask, and utterances involving the problem solving process (task management).Note that we have two tasks, proof task and tutoring task.

3.4. CONVENTIONAL TASK MANAGEMENT DIMENSION 51

1. Initiate task

2. Close task

3. Initiate subtask

4. Close subtask

There are no obligations in this dimension.

3.4.1 Initiate task

Definition An utterance with which the task is initiated.

Examples

(3.89) Bitte zeigen Sie: ...

Relation to moves in other dimensions In our dialogues, initiate task

can be at the same time also initiate dialogue, if there is no explicit initi-

ate dialogue.

Notes Normally, all domain contributions address this utterance in additionto whichever utterance their backward-looking function addresses.

3.4.2 Close task

Definition An utterance which indicates that the task is completed/closed.

Relation to moves in other dimensions In our dialogues, close task canbe at the same time also close dialogue, if there is no explicit close dialogue.

Notes The last move of the hinting session can also serve as close task, ifthere is no explicit move for those (e.g.,did15p T8).

Examples

(3.90) Damit ist die aussage Bewiesen (did15d)

(3.91) S: das wars (soc20k)

(3.92) Ich wiederhole noch einmal...

(3.93) Ich gebe Ihnen den vollstaendigen Beweis

52 CHAPTER 3. DIALOG ANNOTATION SCHEME

3.4.3 Initiate subtask

Definition An utterance with which a subtask13 is initiated.

Examples A potential example that resolves or illustrates a problem (mis-conception) that the student has.

3.4.4 Close subtask

Definition An utterance which indicates that a sub-task is completed/closed.

3.5 Conventional Communication Management

Dimension

This dimension concerns utterances that explicitly manage the structure of thedialogue.

The moves in this dimension might be realised by forward looking conven-tional moves (See Section 3.1.7).

1. Initiate dialogue

2. Close dialogue

3. Initiate subdialogue

4. Close subdialogue

5. Discourse marking

No obligations. This fact is special to human-computer interaction.

3.5.1 Initiate dialogue

Definition An utterance with which the dialogue starts.

Relation to moves in other dimensions In our dialogues, it can be thesame utterance with which the task is initiated, namely initiate task.

3.5.2 Close dialogue

Definition An utterance which indicates that the dialogue is finished/closed.

Relation to moves in other dimensions In our dialogues, close task canbe at the same time also close dialogue.

13A subtask is defined as a task that is secondary, complementary to and aiding the maintask (proof task). It can be used, for instance, in order to resolve a student misconceptionbefore moving on with the main task.

3.6. COMMUNICATIVE STATUS DIMENSION 53

3.5.3 Initiate subdialogue

Definition An utterance with which a sub-dialogue starts.It can be a request clarification, SNU etc. in the backward-looking dimen-

sion.

Notes Initiate subdialogue and initiate subtask may coincide in the same waythat initiate dialogue and initiate task may.

3.5.4 Close subdialogue

Definition An utterance with which a sub-dialogue is ended.

3.5.5 Discourse marking

Definition Utterances which mark the discourse structure, i.e., indicate howthe current utterance relates to the preceding or succeeding context.

Obligations None.

Examples

(3.94) T4: Dies mache ich nun. (did16k)

Notes Open and Close Segment IRUs will also have this function.

3.6 Communicative Status Dimension

This dimension captures features of an utterance unit such as whether it wasintepretable. These features mark exceptional cases. Therefore, most utteranceunits won’t have tags in this dimension (Allen and Core, 1997). The featuresare:

• Uninterpretable

• Abandoned

• Self talk

3.6.1 Uninterpretable

Definition Utterances which are not comprehensible are tagged with this la-bel. These utterances are usually word fragments, or utterances containingmisspelled/mispronounced words such that it is impossible to understand.

54 CHAPTER 3. DIALOG ANNOTATION SCHEME

3.6.2 Abandoned

Definition An utterance or utterance fragment which does not provide con-tent to the dialogue, i.e., “the import of the dialogue would not change, if theseutterance units were removed” (Allen and Core, 1997).

Notes In typed dialogue, such utterances are likely not to be submitted, i.e.,to be deleted before submission.

3.6.3 Self talk

Definition “The utterance unit consists of one speaker talking to himself.”(Allen and Core, 1997). It does not occur in typed dialogue.

3.6. COMMUNICATIVE STATUS DIMENSION 55

Figure 3.1: Forward-looking function: decision tree

Does the speaker make a claim about the world?

Yes

Is the speaker asking for information?Is the speaker making a new claim about the world?

Does the speaker indicate that the claim has already been made?

Tag as ReAssert Tag as other

Does the speaker alreadyknow this information?

Is the speaker requesting an action?

Tag as Diagnostic query Is the speaker askingwhether the hearer understood?

Tag as Understanding query

to the addresseeIs the speaker suggesting potential actions

Tag as Info request

Tag as Assert

Tag asTag asAction_directive Open_option

Yes No

No

Yes No

NoYes

Yes No

Yes No

Yes No

Is the speaker potentially committingto intend to perform a future action?beyond answering a request for information?

Yes No

Is the speaker asking the hearer to performan action, as opposed to simply suggesting it? Do not tag

NoYes

Is the commitment contingent on theaddressee’s agreement?

Tag as CommitTag as Offer

Yes No

conventional communicativeaction?

Is the utterance expressing a

NoYes

Is the speakerapologising?

Yes No

Tag as Apologise Is he expressinggratitude?

Tag as Gratitude Is he signalling emotions?

Is he signalling positive emotions?

Tag as Satisfaction Tag as Frustration

Do not tag

Is it summonning the addressee to start the interaction?

Tag as Conventional_opening

YesYes No

Yes No

Yes No

Yes No

Is it used as dialogue closing?

No

Tag as Conventional_closing Do not tag

Yes No

56 CHAPTER 3. DIALOG ANNOTATION SCHEME

Figure 3.2: Backward-looking function: decision tree

Is the utterance directly addressing previous utterance(s) by another speaker?

Do not tag

Tag as Signal_non_understanding Is the speaker signalling any lack of understanding?

Tag as Request_clarification Is the speaker explicitly signalling understanding?

Does the speaker repeat or paraphrase what was just said?

Does the speaker finish or add to the clause that a speaker is in the middle of constructing?

Does the speaker indicate that the previous speakerhas not said what he actually intended?

Tag as Acknowledge

Tag as Correct_misspeaking Do not tag

Tag as Repeat_rephrase

Tag as Completion

Yes No

Yes No

Yes No

Yes No

Yes No

Yes NoYes No

Yes No

Is the speaker disagreeing with part of the proposal, request or claim?

Tag as Reject_part

Is he addressing a statement?

Is he addressing a SNU?

Tag as Address_other

Is he addressing another move?

Do not tag

Tag as Reject

Tag as Address_SNU

Tag as Address_statement

Yes No

Yes No

Yes No

Yes No

Yes No

Is the speaker addressing a previous proposal, request, question or claim/statement?

No

Do not tag

Yes

Is the speaker addressing a question?

Does the utterance (positively or negatively) resolve it?

Yes

Tag as Answer Tag as Address_info_request,

Diagnostic _queryUnderstanding_query or

Is the speaker explicitly agreeing to part of the proposal, request or claim?

Tag as Accept Tag as Accept_part

Yes No

request or claim?Is the speaker agreeing to all the proposal,

Yes No

No

Yes No

Is the speaker disagreeing with all of theproposal, request or claim?

NoYes

Does the speaker explicitly statethat he cannot give a definite answer?

Is the speaker signalling that they understood none of the previous utterance(s)?

Tag as Address_action_directive

NoYes

Tag as Maybe

Is the speaker addressing an action_directive?

Is the speaker merely signalling either understanding or non−understanding?

3.6. COMMUNICATIVE STATUS DIMENSION 57

Figure 3.3: Tutoring task dimension: decision tree

Is the speaker asking the addressee to confirm orreconfirm that the addressee has the sameshared knowledge?

Tag as Align

Yes

Does the speaker attempt to deal with a problemto progress with the task?

No

Does the speaker try to find outwhether the addressee has a problem?

Tag as Check_origin_of_problem

Is the speaker willing to proceed with the task?

Does the speaker indicate he

Do not tag

Tag as Request_assistance Tag as Resign

Yes

Yes No

Yes No

Yes No

provide a domain contribution?

No

cannot proceed with the task?Does the speaker try to elicit it, either by providing it or asking for it?

Tag as Hint Does the speaker only invitethe addressee to proceed?

Tag as Prompt Do not tag

Yes

Yes No

Yes No

Is the speaker in the process of providing a summaryof the complete proof or of part of the proof?

No

Tag as Recapitulation Does the speaker give feedback on a prior domain contribution?

Does the speaker try to elicit feedback?Does he evaluate it?

Yes No

Yes No

No Yes No

Do not tagTag as Request_evaluationDoes he encourage theaddressee?

Do not tagTag as Encourage

NoYes

Is he signalling acceptance?

Tag as Signal_irrelevant

contribution was not accepted?Does he indicate why the Tag as Signal_accept

Tag as Signal_rejectDoes he signal that the

Does he signal that the contribution is jumping ahead,or something has not been done beforehand that it should?

Tag as Signal_stepsize Does he signal that a common mistake has been made?

Tag as Signal_misconception Does he signal a badly phrased input?

Tag as Signal_ill_formed Tag as Signal_other

Yes

Yes No

Yes No

Yes No

Yes No

Yes No

Yes No

contribution is not relevant?

Does the speaker indicate he expects the hearer to

58 CHAPTER 3. DIALOG ANNOTATION SCHEME

Chapter 4

Annotation tool

As an annotation tool we use MMAX (Muller and Strube, 2001; Muller andStrube, 2003).

Instructions on how to use MMAX can be found underhttp://www.coli.uni-sb.de/cl/courses/Anar-02/mmax-instr.phtml. A manual onusing MMAX can be viewed underhttp://www.coli.uni-sb.de/cl/courses/Anar-02/mmax-manual.html.

For questions on MMAX please contact Magda [email protected]).The data to be annotated is in the directory dialog/experiments/woz-jan-

02/Annotations/DMoves. Subdirectories Utterances and Data2Annotate con-tain the separate annotation projects for all dialogues for the 2 annotation tasks,utterance segmentation and dialogue move annotation respectively. For each di-alogue, there is the .anno file and the words and text files, and an initial (emptyor pre-filled) markable file) plus all the ”surrounding” files (anno scheme, xlsand dtd files etc). Each annotator will make a copy of this into a new directoryby their name. To hand in their annotation, the annotators will hand in thewhole directory.

The first annotation task will be segmenting coordinated clauses into utter-ances. That is, we separate main clauses that are connected by the coordinatingconjunctions aber, oder, denn and und. For instance, “Dann nehmen wir einbeliebiges Element x ∈ B und zeigen , daß dieses auch in K ( A ) sein muß”will be segmented as follows: [ Dann nehmen wir ein beliebiges Element x ∈ B][und zeigen , daß dieses auch in K ( A ) sein muß ]. Further examples:

“Das ist richtig, aber wie geht es weiter?”“Deswegen liegt kein Element von B in A , aber dafur in K ( A ) .”

This will be done with MMAX by creating separate markables for eachcoordinated clause and tagging them as utterances.

We split the annotation projects in two groups, one comprising all didactic(did) dialogues (Andrea), and one all min and soc dialogues (Ann).

The second annotation task is annotating for dialogue moves. Note that youshould look at the already annotated forward-looking dialogue acts when youannotate for the backward-looking acts in order to be able to identify and mark

59

60 CHAPTER 4. ANNOTATION TOOL

(via pointers) their antecedents.For both annotation tasks, the annotators should keep a log of the time they

spent per dialogue.

Appendix A

Font Coded Taxonomy

We use the following font coding for more information on the origin and occur-rence of the dialogue moves: Normal font: DAMSL moves in DIALOG corpus,italic: DAMSL moves not in corpus, sc: BE&E moves, bf: new moves.

A.1 Forward-looking

1. Statement

(a) Assert

(b) Reassert

(c) Other statement

2. Influencing-addressee-future-action

(a) Open option

(b) Action Directive

3. Info Request

4. Diagnostic query

5. Understanding query

6. Committing-speaker-future-action

(a) Commit

(b) Offer

7. Conventional

(a) Conventional opening

(b) Conventional closing

61

62 APPENDIX A. FONT CODED TAXONOMY

8. Apologise

9. Gratitude

10. Signaling emotion

(a) Frustration

(b) Satisfaction

11. Other forward function

A.2 Backward-looking

12. Agreement

(a) Accept

(b) Accept part

(c) Maybe

(d) Reject

(e) Reject part

13. Understanding

(a) Signal non-understanding (SNU)

(b) Request clarification

(c) Signal understanding

i. Repeat rephrase

ii. Acknowledge

iii. Completion

(d) Correct misspeaking

14. Answer

15. Information-relation

(a) Address action dir

(b) Address info request

(c) Address statement

(d) Address SNU

(e) Address other

A.3. TASK 63

A.3 Task

A.3.1 Proof task

(a) Domain Contribution

A.3.2 Tutoring task

(a) Hint

(b) Domain contribution evaluation

i. Signal accept

ii. Signal reject

A. Signal wrong

B. Signal irrelevant

C. Signal stepsize

D. Signal misconception

E. Signal ill formed

F. Signal complete partially accurate

G. Signal complete inaccurate

H. Signal incomplete accurate

I. Signal incomplete partially accurate

J. Signal near miss

K. Grain of truth

L. Signal other

(c) Check origin of problem

(d) Request evaluation

(e) Prompt

(f) Encourage

(g) Request assistance

(h) Resign

(i) Align

A.4 conventional task management

(a) Initiate task

(b) Close task

(c) Initiate subtask

(d) Close subtask

64 APPENDIX A. FONT CODED TAXONOMY

A.5 conventional communication management

(a) Initiate dialogue

(b) Close dialogue

(c) Initiate subdialogue

(d) Close subdialogue

(e) Discourse marking

A.6 communicative status

(a) Uninterpretable

(b) Abandoned

(c) Self talk

References

Jan Alexanderson et al. 1997. Dialogue acts in verbmobil-2. Technical reportverbmobil-report 204, DFKI GmbH Saarbrcken, Universitt Stuttgart, Tech-nische Universitt Berlin,Universitt des Saarlandes.

James Allen and Mark Core. 1997. Draft of DAMSL: Dialogue act markup inseveral layers. DRI: Discourse Research Initiative, University of Pennsylva-nia.

James Allen et al. 2001. Towards conversational human-computer interaction.AI magazine.

Chris Benzmuller, Armin Fiedler, Malte Gabsdil, Helmut Horacek, IvanaKruijff-Korbayova, Manfred Pinkal, Jorg Siekmann, Dimitra Tsovaltzi,Bao Quoc Vo, and Magdalena Wolska. 2003. Tutorial dialogs on mathe-matical proofs. In Proceedings of the IJCAI Workshop on Knowledge Rep-resentation and Automated Reasoning for E-Learning Systems, pages 12–22,Acapulco.

Mark G. Core, Johanna D. Moore, and Claus Zinn. 2002. Draft: Tutorialannotation scheme. September 11.

Armin Fiedler and Dimitra Tsovaltzi. 2003a. Automating hinting in an intelli-gent tutorial system. In Proceedings of the IJCAI Workshop on KnowledgeRepresentation and Automated Reasoning for E-Learning Systems, pages 23–35, Acapulco.

Armin Fiedler and Dimitra Tsovaltzi. 2003b. Automating hinting in mathemat-ical tutorial dialogue. In Proceedings of the EACL-03 Workshop on DialogueSystems: Interaction, Adaptation and Styles of Management, pages 45–52,Budapest.

Jonathan Ginzburg. 1994. An update semantics for dialogue. In Proceedings ofthe International Workshop on Computational Semantics, Tilburg.

A.6. COMMUNICATIVE STATUS 65

Jonathan Ginzburg. 1996. Interrogatives: Questions, facts and dialogue. InThe Handbook of Contemporary Semantic Theory. Blackwell, Oxford.

Joris Hulstijn. 2003. Roles in dialoguemarked. In Ivana Kruijff-Korbayova andClaudia Kosny, editors, Proceedings of DiaBruck’03, the 7th Workshop onthe Semantics and Prgmatics of Language, Saarbrucken, Germany.

Gregory D. Hume, Joel A. Michael, Rovick A. Allen, and Martha W. Evens.1996. Hinting as a tactic in one-on-one tutoring. Journal of the LearningSciences, 5(1):23–47.

Joern Kreutel and Colin Matheson. 1999. Modelling questions and assertionsin dialogue using obligations. In Proceedings 3rd International Workshopon the Semantics and Pragmatics of Dialogue, Amstelogue, University ofAmsterdam.

Joern Kreutel and Colin Matheson. 2001. Context-dependent interpretationsand implicit dialogue acts. In BI-DIALOG, Bielfeld.

Staffan Larssom. 2003. Interactive communication management in an issue-based dialogue system. In Ivana Kruijff-Korbayova and Claudia Kosny, ed-itors, Proceedings of DiaBruck’03, the 7th Workshop on the Semantics andPrgmatics of Language, pages 75 – 82, Saarbrucken, Germany.

Staffan Larsson. 2002. Issue-based Dialogue Management. Ph.D. thesis, De-partment of Linguistics, Goeteborg University, Sweden.

Eng Leong Lim and Dennis W. Moore. 2002. Problem solving in geome-try: Comparing the effects of non-goal specific instruction and conventionalworked examples. Journal of Educational Psychology, 22(5):591–612.

Colin Matheson, Massimo Poesio, and David Traum. 2000. Modelling ground-ing and discourse obligations using update rules. In Proceedings NAACL2000, Seattle.

Johanna Moore. 1993. What makes human explanations effective? In Pro-ceedings of the Fifteenth Annual Meeting of the Cognitive Science Society,Hillsdale, NJ.

Christoph Muller and Michael Strube. 2001. Mmax: A tool for theannotation of multi-modal corpora. In Proceedings of the 2nd IJ-CAI Workshop on Knowledge and Reasoning in Practical Dialogue Sys-tems, pages 45–50, Seattle, Wash., 5 August. http://www.eml.villa-bosch.de/english/Research/NLP/Publications.

Christoph Muller and Michael Strube. 2003. Multi-level annotationin mmax. In Proceedings of the 4th SIGdial Workshop on Dis-course and Dialogue, Sapporo, Japan, 4-5 July. http://www.eml.villa-bosch.de/english/Research/NLP/Publications.

Natalie K. Person and Arthur Graesser. 2003. Fourteen facts about humantutoring: Food for thought for ITS developers. In aied2003, SupplementaryProccedings, pages 355–344, Sydney.

J. R. Searle. 1975. A taxonomy of illocutionary acts. Language, Mind andKnowledge, Minnesota Studies in the Philosophy of Science, pages 344–369.

Nicole Shechtman and Leonard M. Horowitz. 2003. Media inequality in con-versation: how people behave differently when interacting with computers

66 APPENDIX A. FONT CODED TAXONOMY

and people. In Proceedings of the CHI 2003 conference on human factors incomputing systems, pages 5–10, Florida, USA.

D. R. Traum and J. F. Allen. 1994. Discourse obligations in dialogue processing.In Proceedings 32nd Annual meeting of the Association for ComputationalLinguistics (ICSLP92), pages 1–8.

Dimitra Tsovaltzi and Armin Fiedler. 2003. An approach to facilitating reflec-tion in a mathematics tutoring system. In Proceedings of AIED Workshopon Learner Modelling for Reflection, pages 278–287, Sydney, Australia.

Dimitra Tsovaltzi and Elena Karagjosova. 2004. A dialogue move taxonomyfor tutorial dialogues. In Proceedings of 5th SIGdial Workshop on Discourseand Dialogue, Boston, USA.

Dimitra Tsovaltzi and Colin Matheson. 2002. Formalising hinting in tutorialdialogues. In Proceedings of EDILOG: 6th Workshop on the Semantics andPragmatics of Dialogue, pages 185–192, Edinburgh, Scotland, UK.

Dimitra Tsovaltzi, Helmut Horacek, and Armin Fiedler. 2004. Building hintspecifications in a NL tutorial system for mathematics. In Proceedings ofthe 16th International Florida AI Research Society Conference (FLAIRS-04), Florida, USA.

Marilyn Walker and Rebecca Passonneau. Date: A dialogue act tagging schemefor evaluation of spoken dialogue systems. http://www.citeseer.nj.nec.com/walker01date.html.

Brent Wilson and Peggy Cole. 1996. Cognitive teaching models. In D.H.Jonassen, editor, Handbook of Research for educational communications andtechnology. MacMillan.

Magdalena Wolska, Bao Quoc Vo, Dimitra Tsovaltzi, Ivana Kruijff-Korbayova,Elena Karagjosova, Helmut Horacek, Malte Gabsdil, Armin Fiedler, andChristoph Benzmuller. 2004. An annotated corpus of tutorial dialogs onmathematical theorem proving. In Proceedings of LREC. In print.