4. Dialogue Acts

Introduction

Dialogue act annotation schemes are used to mark important characteristics of utterances. These annotations indicate the role of an utterance in a specific dialogue and make the relationship between utterances more obvious.

Most of the dialogue schemes nowadays are task-oriented as we will see later on in this report. This is due to reduce the amount of annotation tags to a capable size for annotation and to increase the analyzing rate of the NLP system in which the scheme is used. The information content (or the semantics) of task-oriented dialogues can be basically splitted into task / domain related information and information that addresses the communication process. To guarantee generality and therefore more flexibility both information levels should be kept separately in the notion-choice of tags. Schemes which cover only those two fields are said to be schemes for rather shallow analysis.

As an example of a scheme that allows deep analysis DAMSL can be mentioned. With its forward and backward looking functions it keeps track of how an utterance constrains the future beliefs and actions of the participants, and affects the discourse and how an utterance relates to the previous discourse, respectively.
 

Scheme Comparison

• Alparon

Developer: Delft University of Technology, Netherlands
Domain: information retrieving about the services offered by the Dutch public transport

• Chat

Developer: Carnegie Mellon University, Department of Psychology, USA
Domain: analysis of child language.

• Coding scheme of the University of Chiba

Developer: University of Chiba, Japan
Domain: route direction, scheduling, telephone shopping and tourist information.

• COCONUT

Developer: The University of Pittsburgh, USA
Domain: buying furniture for the living and dining rooms of a house
 
• Condon & Cech's coding scheme

Developer: University of Southwestern Lousiana, USA
Domain: decision-making
 
• C-STAR

Developer: C-STAR Consortium
Domain: travel-planning
 
• DAMSL

Developer: Discourse Representation Initiative
Domain: general

• Janus

Developer: Carnegie-Mellon University, USA
Domain: appointment-scheduling
 
• Flammia's coding scheme

Developer: Massachusetts Institute of Technology, USA
Domain: information-seeking
 
• LinLin

Developer: Linköping University, Sweden
Domain: information-retrieving

• Maptask

Developer: The University of Edinburgh, UK
Domain: route direction

• Nakatani et al.'s coding scheme

Developer: Harvard University and AT&T Bell Laboratories, USA
Domain: direction-giving

• SLSA

Developer: Göteborg University, Sweden
Domain: courtroom interaction

• SWBD-DAMSL

Developer: University of Colorado, USA
Domain: general/telephone conversation between strangers

• Traum's coding scheme

Developer: Université de Genève, Switzerland
Domain: general/task-oriented

• Verbmobil

Developer: DFKI, Germany
Domain: appointment scheduling

Another reason might be that they are not used very much which could, for example, lead to the assumption that they are too complicated.

However, in the following some criteria for schemes are detailed which can be used to scale the observed schemes.

• Task-Orientation: almost all schemes refer to dialogues that are task-driven, i.e. performed to accomplish a specific task. LE research is obviously oriented towards task-driven dialogues, since it is not particularly useful for this area to analyse ordinary conversations. However, some schemes are nevertheless devised for non-task-driven dialogues. Values for this category are TD (task-driven) and NTD (non-task-driven).

• Applications Orientation: again, this is a relevant parameter for LE research. The majority of task-driven dialogues are also application-oriented, i.e. geared towards industrial or commercial applications. On the other hand other dialogues, and thus their schemes, are not applications-oriented. Values for this category are positive (applications-oriented) and negative (non-application-oriented).

• Domain Restriction: most dialogues in LE research are restricted to a relatively highly specific domain of subject matter. This represents a relevant parameter since schemes are often influenced by the type of domain of the dialogues annotated. Values for this category are positive (restricted domain) and negative (unrestricted domain). The positive value can be further categorized according to the following typology of domains:

 

 

TR: travel
TS: transport
COS: computer operating systems
COU: courtroom interaction
BA: business appointments
DES: directory enquiry services
FUR: furnishing rooms interactively
DIR: giving directions
INST: giving instructions (e.g. about cooking)

• Activity Type: this category refers to the type of activity to which dialogues may belong, and is another dimension along which schemes can be classified. It seems that in current dialogue research, there is a major division between two leading paradigms: cooperative tasks between human participants (such as negotiating appointments) and information extraction tasks, in which a human agent interrogates a computer system (or a human surrogate for a computer system). A typology of activity types is the following (together with abbreviation used):

 

 

CN: cooperative negotiation
IE: information extraction
PS: problem solving
TI: teaching/instruction
CO: counselling
CH: chatting
 

• Human / Machine Participation:

HH: human-human dialogues. This category is further divided into:
MM: machine-mediated(computer, telephone)
NMM: non-machine-mediated
HM: human-machine dialogues. This category is further divided into:
S: simulated
NS: non-simulated

In order to develop a standard it is necessary to compare schemes with regard to their underlying task and dialogue acts. The following tables present domain-grouped schemes and show the equivalence between their dialogue acts. Entries with italic font represent higher order expressions which can't be annotated.
 

• There is no dialogue act information available on the Janus scheme.
• No specific dialogue acts are mentioned for SLSA.
• Nakatani et al.'s coding scheme is developed for discourse annotating only, therefore no dialogue acts are specified, too.
• All the other schemes, like COCONUT, Condon and Cech's and C-STAR are not stated in one of the tables above as they are designed for different tasks (e.g. problem solving, decision making or travel-planning, respectively). However, they are taken under consideration in the conclusion below.

Conclusion

The huge amount of coding schemes detailed in the Annexes shows the current research interests in dialogue act annotation schemes. There also seems to be a trend to shallow, task-oriented annotation as these schemes predominate those which focus on a general approach. The comparison of dialogue acts of schemes with equivalent domains reflect the similarities expected. But more surprisingly even a dialogue act comparison among all schemes regardless to their orientation shows quite a lot of parallelism although the general schemes are, of course, more comprehensive.

In order to decide, which schemes should be taken under consideration in MATE, schemes should have a coding book, they should be heavily used and should have good evaluation results. Also a scheme that is not related to a special task seems to be more appropriate as a task related and therefor possibly restricted one. If we look at the general comparison of schemes above, one can observe that all listed schemes provide a coding book. Amongst the schemes which are mostly used we can see Alparon, Chat, SWBD-DAMSL and Verbmobil. Unfortunately Chat hasn't been evaluated, but Alparon, SWBD-DAMSL and Verbmobil are judged to be good. As SWBD-DAMSL is the only one of these schemes which is not task related, it should be definitely supported in MATE. With regard to the MATE standard, of course, the dialogue acts of all schemes should be taken into account and analysed.

Last Modification: 26.8.1998 by Marion Klein