(Semi-)Automatic Analysis of Dialogues
Mare Koit
Institute of Computer Science, University of Tartu, J. Liivi 2, Tartu, Estonia
Keywords: Dialogue, Dialogue Act, Dialogue Structure, Communicative Strategy, Analysis, Software.
Abstract: We study human-human and human-computer dialogues with the aim to determine which dialogue acts and
communicative strategies do the participants of interaction use, and which structural parts does a dialogue
include. We develop software that makes it possible to recognise and annotate the dialogue acts, the
dialogue structure and the communicative strategies. In order to recognise dialogue acts, a data-driven
method is implemented when determination of the dialogue structure and the strategies is based on rules.
The software tool is used by linguists in dialogue studies which further aim is to develop a dialogue system
that interacts with a user in natural language following norms and rules of human-human communication.
The contribution of the paper consists of integration of the existing approaches within a common platform
and adaptation to the Estonian language.
1 INTRODUCTION
The pragmatic analysis of a coherent text usually
follows to the morphological, syntactic and semantic
analysis of the sentences which form the text. The
output of preceding stages of analysis is used as the
input of the following stage.
We are studying a special kind of texts –
dialogues (transcripts of human-human spoken
dialogues and human-computer written dialogues).
We try to carry out the pragmatic analysis of
dialogue texts without the traditional preceding
stages of analysis (morphological, etc.). Therefore,
the input of the pragmatic analysis is a plain text.
First, we determine the dialogue acts (DA) in a
dialogue using a statistical method. After that,
recognition of the dialogue structure and dialogue
strategies can be carried out using the rules which
are based on the DAs.
Our aim is to build a software tool that can be
used by linguists for annotating the dialogues in
order to study and compare their structure.
Different typologies of DAs have been worked
out (e.g. Sinclair and Coulthard, 1975, Stenström,
1994, Bunt et al., 2012). The most well-known
typology, DAMSL (Allen and Core, 1997), is
proposed as the standard annotation scheme for
dialogue tagging by the Discourse Resource
Initiative. The main aim of DAMSL is to capture the
multiple function utterances can have, as well as the
interrelation of different speech acts.
We have worked out our own typology of DAs
which is based on the principles of organization of
conversation borrowed from the Conversation
Analysis, CA (Hutchby and Wooffitt, 1998) which
has been our main research method since 1990s. We
are using the act typology for annotating our
dialogue corpus. However, the main part of our
typology coincides with DAMSL.
Several data-driven methods have been used for
recognition of DAs: n-grams, Hidden Markov
Models, Bayes classifiers, neural networks, decision
trees,
transformation-based learning, memory-
based learning, etc
(Reithinger and Maier, 1995,
Wright, 1998, Keizer et al., 2002, Grau et al., 2004,
Levin et al., 2003, Samuel et al., 1998, Fernandez et
al., 2005). We are using Naïve Bayes classifier for
recognition of DAs.
The structural parts of a dialogue can be
determined using the rules formulated on the basis
of DAs. The dialogue manager of a dialogue system
uses information about the structure of
communication in order to understand the user’s
utterances and to generate its own responses (Field
et al., 2008).
Communicative strategies, or dialogue policies,
have been annotated and studied in information
seeking dialogues (Jokinen, 1996) and in negotiation
dialogues (Georgila et al., 2011). Reinforcement
445
Koit M..
(Semi-)Automatic Analysis of Dialogues.
DOI: 10.5220/0004818104450452
In Proceedings of the 6th International Conference on Agents and Artificial Intelligence (ICAART-2014), pages 445-452
ISBN: 978-989-758-015-4
Copyright
c
2014 SCITEPRESS (Science and Technology Publications, Lda.)
learning has been used to recognise argumentation
policies in negotiations. We are departing from the
notion of communicative strategy as introduced by
Jokinen in her Constructive Dialogue Model, CDM
(Jokinen, 1996, 2009). We use rules to assign the
communicative strategies to utterances.
The paper is organised as follows. In section 2
we introduce our data – the Estonian dialogue
corpus and the dialogue act typology. Sections 3 to 5
describe the functionality of the software tool: semi-
automatic recognition of dialogue acts, automatic
determination of structural parts of dialogue and
communicative strategies. In section 6 we draw
conclusions.
2 DIALOGUE CORPUS AND THE
TYPOLOGY OF DIALOGUE
ACTS
2.1 Estonian Dialogue Corpus
The Estonian Dialogue Corpus (Hennoste et al.,
2008) currently includes three parts. The first part of
the corpus is formed by human-human spoken
dialogues recorded in authentic situations and
transliterated using the transcription of CA (Hutchby
and Wooffitt, 1998). In our corpus, there are
telephone calls as well as face-to-face conversations,
among them institutional as well as everyday
conversations. Most of them are institutional
information-seeking dialogues. The number of the
dialogues is over 1000. The main aim of recording
the dialogues has been the study of human-human
conversation. For that reason, the corpus includes
various types of dialogues: directory inquiries, calls
to travel agencies, bus stations, outpatients’ offices,
shops, etc. as well as face-to-face dialogues in shops,
services, travel agencies, guiding on the street, etc.
However, such diversity makes harder the automatic
analysis of dialogues. The corpus is open and
increasing, new recordings and transliterations will
be made and added into the corpus.
The second part of the corpus is collected in
Wizard-of-Oz (WOZ) experiments where a human
plays the role of the computer (Dahlbäck et al.,
1993, Bellucci et al., 2009). Custom software is used
for experiments. A user puts in his/her text (request
for information) from the keyboard and receives the
Wizard’s answers on the screen. The number of
WOZ dialogues is about 100.
The third part of the corpus is formed by actual
interactions with two web-based dialogue systems
(DS). One of them gives information about cinema
programmes and the other – dental information
(www.dialoogid.ee). The user puts in his/her texts in
Estonian from the keyboard and receives the
computer’s answers on the screen, similarly with the
WOZ experiments. The number of dialogues is
about 100.
Different kinds of dialogues have been collected
and used in dialogue studies for comparison. A part
of the corpus has been used for development of the
software. Still, the software is aimed for the
automatic analysis of the whole corpus which is
increasing in time.
2.2 The Typology of Dialogue Acts
Our main aim is to support the study of human-
human communication. For that reason, we have
worked out our own typology of DAs (s. an
overview in Appendix). The typology is based on
CA. In the typology, the DAs are divided into two
groups – adjacency pair (AP) acts where the first
pair part expects a certain second pair part like
question – answer, and non-AP acts like
acknowledgement.
On the other hand, the DAs are divided into
communication managing acts (e.g. greeting and
thanking), repair acts (e.g. other-initiated repair),
and information acts (e.g. different types of
questions). The name of a DA consists of two parts
separated by a colon (e.g. QU: Wh-question, VR:
Acknowledgement): the first part indicates the act
class (e.g. QU – questions, VR – voluntary
responses) and the second part is the proper name of
the act (e.g. Wh-question, Acknowledgement). The
total number of the acts is 126. The full list can be
found e.g. in (Hennoste and Rääbis, 2004). Fig. 1
demonstrates the transcript of a spoken dialogue
where DAs are annotated (DA tags are placed
between vertical strokes, some of the utterances
have double DA tags, i.e. they hold more than one
function). The transcription of CA is used in the
example.
In order to study communication, we annotate
the DAs in our corpus. So far, two persons annotated
the DAs manually, by using custom software that
simplifies to choose dialogues from the corpus and
DAs from a list and then a third person (an expert)
disambiguated the annotations. Automatic
annotation will make the job much easier. Further,
we are looking for the structural parts of dialogue
which can be simply determined on the basis of
adjacency pairs of DAs.
The communication participants use
ICAART2014-InternationalConferenceonAgentsandArtificialIntelligence
446
communicative strategies in order to achieve their
communicative goals. Our software is planned, first,
to recognise DAs in a dialogue transcript and after
that, to determine the dialogue structure and
communicative strategies.
A : ((summons)) | RIF: Summons |
B: `Estmar=`info, | RIS: Answer
|
| RS: Introduction |
Estmar info
`Leenu=kuuleb | RS: Introduction |
Leenu is hearing
tere | RIF: Greeting |
good morning
A: tere `päevast. | RIS: Greeting |
good morning
(.) ee kas te `ütleksite mulle takso num-
`telefoninumbri e `tellimiseks.
| QUF: Open yes/no |
could you tell me a phone number for
ordering a taxi
(0.5)
B: neli kaks `null, neli kaks `null on
`Eepee auto. | QUS: Giving information |
four two zero four two zero is Eepee car
(0.5)
A: jah. | VR: Neutral acknowledgement |
yes
(.) neli kaks null neli kaks null jah?
| QUF: Offering answer | | RPF: Checking |
four two zero four two zero yes
B: jah? | QUS: Yes | | RPS: Repair |
yes
A: no suur `tänu teile. | RIS: Thanking |
thank you very much
B: palun? | RIS: Please |
you are welcome
Figure 1: A directory inquiry from the Estonian Dialogue
Corpus (A – client, B – official). Dialogue acts are
annotated (s. Appendix).
The next sections 3 to 5 are dedicated to the
description of the software tool.
3 RECOGNITION OF DIALOGUE
ACTS
3.1 Method
As a result of previous observations, our first aim
was to choose a suitable method for automatic
recognition of DAs. After the DAs are annotated in
dialogues, the rules for recognition of the dialogue
structure can be formulated on the basis of DA tags.
Further, there exists a close relation between DAs
and communicative strategies (in the sense of CDM)
therefore rules can be formulated for recognition of
communicative strategies on the basis of DAs.
In this way, DAs prove to be good indicators for
determining the dialogue structure as well as the
dialogue strategies.
We have tested several methods for recognition
of DAs: multi-layered perceptrons, decision trees,
suffix trees, Bayes classifier. An overview of the
results can be found in (Koit, 2011, Aller, 2012). No
method was considered sufficient for fully automatic
recognition of DAs. There are at least two reasons of
that – the complexity of the typology of DAs and the
diversity of our (relatively small) corpus which does
not offer necessary training material. That is why we
decided to implement semi-automatic annotation of
DAs in our software: the programme finds DAs for
every utterance in a dialogue and then a human
annotator corrects the annotation errors if needed.
We have chosen the most robust and simplest
method from the set of the tested methods – Naïve
Bayes (Manning and Schütze, 1999).
3.2 Implementation
The semi-automatic annotator splits a dialogue text
into utterances and assigns up to five most probable
DAs to every utterance. After that, a human can
correct mistakes and then to repeat the automatic
annotation if needed. The input is a .txt file – a
dialogue where turns (but not the utterances) are
located in separate rows. The output is a .txt file
where turns are splitted into utterances placed into
different rows and DA tags are assigned to every
utterance. The annotator implements Naïve Bayes
classifier. In the experiments with the classifier, the
following features were chosen in order to achieve
the best results: the probability of trigrams of words,
the length of utterances (number of words) and the
geometric mean of the probabilities of the DA tags
(Fishel, 2007).
The annotator itself includes two parts: training
and annotation. Also cross-validation can precede to
the training (Fig. 2). When training, a new session is
initiated, a model is created and used for annotation
of new data. The classifier is implemented as Perl
scripts. The annotator was trained on 800 dialogues,
ten-fold cross-validation was used. The average
recall of 64.7% and precision of 33.0% were
received. The calculations were made on the basis of
the most probable tag for every utterance but the
interface offers up to five tags in decreasing order of
the probability. Actually, a human annotator does
not need to search a suitable tag from the list of all
DA tags but the right tag is mostly located among
the five.
(Semi-)AutomaticAnalysisofDialogues
447
Figure 2: Dialogue act annotator. The numbers indicate
different processing states: (1) start of new session, (2 to
4) cross-validation, (5 to 7) training, (8) preprocessing
new data, (9) applying model to new data.
A detailed description of the dialogue act
annotator can be found in (Aller, 2012).
4 ANNOTATION OF THE
DIALOGUE STRUCTURE
A typical dialogue consists of three parts: (1) a
conventional beginning, (2) the main information
part, and (3) a conventional ending. The kernel of
the information part is an adjacency pair directive –
grant or question – answer.
Sub-dialogues can occur in the main part after a
request (or question) and/or answer, respectively: an
adjusting/specifying question is asked and answered,
or a repair for solving a communication problem is
initiated and performed.
The corpus analysis suggests to use adjacency
pairs of DAs as the main cues for recognition of
different parts and sub-dialogues of a dialogue.
The conventional opening and closing parts can
be recognised looking for APs of rituals and the
single conventional act RS: Introduce in the
beginning or at the end of a dialogue, respectively.
The main part begins with a request or question
immediately after the opening part and continues
until the closing part begins. Sub-dialogues in the
main part can be recognised by double-tags:
information-sharing initiated by the responder before
giving information begins with the act tag ACF:
Adjusting the conditions of answer and ends with
the act tag ACS: Adjusting the conditions of answer.
Other-initiated repair begins with the act tag RPF:
Reformulation, RPF: Checking or RPF: Non-
understanding and ends with RPS: Repair (Fig. 3, cf.
Koit, 2012).
[Opening]
A: ((summons)) RIF: Summons
B: RIS: Answer RS: Introduce [B
introduces the service company]
( RS: Introduce [B introduces him/herself] )
( RIF: Greeting )
A: RIS: Greeting
[Main part]
A: DIF: Request / QUF: Wh-question/
Open yes-no
( [information sharing initiated by B]
--> B: ACF: Adjusting the conditions of
answer
<-- A: ACS: Adjusting the conditions of
answer
)
( [other-initiated repair]
--> B/A: RPF: Reformulation/ Checking/
Non-understanding
A/B: RPS: Repair
<-- ( B/A: VR: Repair evaluation )
)
B: ( VR: Neutral acknowledgment
QUS/DIS: Deferral ) DIS/QUS: Giving
information
( [other-initiated repair]
--> A/B: RPF: Checking/ Non-
understanding/ Reformulation
B/A: RPS: Repair
<-- ( A/B: VR: Repair evaluation )
)
( A: VR: Neutral acknowledgment /
Neutral bounder / Neutral change of state )
[Closing]
A: RIF: Thanking ( RIF: Greeting )
( B: RIS: Please RIS: Greeting )
Figure 3: The structural parts of information dialogue:
opening, main part, closing. A – client, B – official. Sub-
dialogues are marked by ‘-->’ (begin) and ‘<--’ (end).
The dialogue acts between ‘(’ and ‘)’ can be missed. An
overview of the DA typology is given in Appendix.
The automatic annotator of the dialogue
structure (implemented by S. Aller) takes as input
the dialogue where DAs are annotated (.txt file) and
uses rules for recognition of different parts of
dialogue. The parts are distinguished by different
colors. The output is given in two formats: .txt and
.xml. The programming language is PHP.
An example output is presented in Fig. 4. The
main part of the dialogue includes a sub-dialogue –
repair initiated by the client (participant A) and
performed by the official (participant B).
ICAART2014-InternationalConferenceonAgentsandArtificialIntelligence
448
Opening
A : ((summons)) | RIF: Summons |
B: Estmar info
| RIS: Answer | | RS: Introduction |
Leenu is hearing | RS: Introduction |
good morning | RIF: Greeting |
A: good morning | RIS: Greeting |
Main part
could you tell me a phone number for
ordering a taxi | QUF: Open yes/no |
(0.5)
B: four two zero four two zero is Eepee
car | QUS: Giving information |
(0.5)
A: yes | VR: Neutral acknowledgement |
Sub-dialogue: other-initiated repair
--> four two zero four two zero yes
| QUF: Offering answer |
| RPF: Checking |
<-- B: yes | QUS: Yes | | RPS:
Repair |
Closing
A: thank you very much | RIS: Thanking |
B: you are welcome | RIS: Please |
Figure 4: The structural parts of information dialogue:
opening, main part, closing (cf. Fig.1).
5 RECOGNITION OF
COMMUNICATIVE
STRATEGIES
We are using the notion of the communicative
strategy, introduced in (Jokinen, 1996) as a part of
CDM.
A communicative strategy is used by a
participant to build up the next utterance as a
reaction to the partner’s previous utterance.
Four context factors are used in CDM to determine
communicative strategies:
1) expectations – is the partner’s turn expected or
not
2) the central conception – does the partner’s turn
keep the topic or not
3) initiatives – has the speaker initiative or not
4) goals – are the speaker’s goals fulfilled or not.
All the context factors have binary values in the
CDM which results in 16 communicative strategies
(e.g. finish/start, follow-up old, somethingelse, etc.,
Table 1).
Every strategy can be represented as a vector
with the values of the coordinates of 0 or 1, e.g.
0000 (strategy notrelated) means that the partner’s
turn is unexpected, does not keep the topic, the
speaker does not have the initiative and there are
unfulfilled goals (cf. Table 1).
Donotannotate
1. PS: Uninterpretable
2. Rituals, except of
RIF: Preclosing – specify-new
RIS: Accept – follow-up-new
RIS: Reject – somethingelse
Adjacencypairacts
Directives,questions,opinions
Firstpairpart
3. The first A: DIF/QUF – finish/start
4. Later A: DIF/QUF (with a single DA tag)
a. If B: DIS/QUS does not precede then –
backto
b. If B: missing information or topic change
precedes then – specify-new
c. If B: giving information precedes then –
new-dialogue
5. B: QUF: Alternative/Open yes/no or DIF: Offer –
new-dialogue
6. A: DIF+TCF – specify-new
7. A: DIF+RP – specify-new (self-repair changes topic)
8. OPF: Opinion – new-dialogue
Secondpairpart
9. DIS/QUS: Giving information/Accept – follow-up-old
10. DIS/QUS: Missing information/Reject – continue
11.
DIS: Agreeing no – continue (like Missing
information)
12. OPS: Other – continue (like Reject)
Withdoubletags(thesecondpairandthefirstpairparts):
13. QUS+ QUF – new-question
14. DIS+DIF – new-request
Contactcontrol
15. CCF – specify-new
16. CCS – follow-up-new
Sub‐dialogues
17. ACF and RPF – subquestion,X
18. ACS and RPS – follow-up-old
Non‐adjacencypairacts
Additionalinformation
19. A: AI: Specification – backto
20. B: AI: Specification/Explication/Emphasize after
giving information by the same participant – follow-
up-new
Responses
21. A: VR: neutral/evaluative continuer – continue
22.
All the remaining VR – somethingelse, except of
a. VR: neutral/evaluative bounder
i. If giving information follows
then – object,X
ii. If not then – specify
b. VR: neutral/evaluative change of state –
repeat-new
Primarysingleacts
23. SA: giving information – follow-up-old
24. SA: other – somethingelse
Figure 5: Relations between dialogue acts (s. Appendix)
and communicative strategies.
(Semi-)AutomaticAnalysisofDialogues
449
Table 1: Communicative strategies in CDM.
Communicative strategy Vector
Notrelated 0000
New-st-request 0001
Objekt,X 0010
Specify-new 0011
Continue 0100
Somethingelse 0101
Subquestion,X 0110
New-dialogue 0111
New question 1000
New-request 1001
Repeat-new 1010
Specify 1011
Follow-up-old 1100
Follow-up-new 1101
Backto 1110
Finish/start 1111
Communicative
strategy
Vector
A : ((summons))
| RIF: Summons |
- -
B: Estmar info
|RIS: Answer|
|RS: Introduction|
- -
Leenu is hearing
|RS: Introduction|
- -
good morning
|RIF: Greeting|
- -
A: good morning
|RIS: Greeting|
- -
could you tell me
a phone number for
ordering a taxi
|QUF: Open yes/no|
Finish/start 1111
(0.5)
B: four two zero
four two zero is
Eepee car
|QUS: Giving
information|
Follow-up-old 1100
(0.5)
- -
A: yes
|VR: Neutral
acknowledgement|
Somethingelse 0101
four two zero four
two zero yes |QUF:
Offering answer|
|RPF: Checking|
Subquestion,X 0110
B: yes |QUS: Yes|
|RPS: Repair|
Follow-up-old 1100
A: thank you very
much
|RIS: Thanking|
- -
B: you are welcome
|RIS: Please|
- -
Figure 6: Communicative strategies in information
dialogue (cf. Fig.1).
We have manually annotated the strategies in 60
information dialogues, occasionally taken from the
Estonian dialogue corpus. The study of the dialogues
has given as a result the following algorithm for
determination of communicative strategies on the
basis of DAs and the participants signs (Fig. 5, A –
client, B – official). The automatic annotator of
communicative strategies takes as input a dialogue
file (.txt) where DAs are annotated and gives as
output a .txt file where communicative strategies and
the corresponding vectors are assigned to the
utterances.
Some of the utterances remain without tags
because the strategies in CDM are mainly related to
requesting and giving information, i.e. to the main
part of dialogue (Fig.6).
The annotator is implemented by S. Aller in
PHP. A user can choose a dialogue from the corpus
and then annotate the DAs, or s/he can choose a
dialogue where the DAs are already annotated and
then optionally to annotate the dialogue structure
and/or communicative strategies.
6 CONCLUSIONS
We have introduced the Estonian dialogue corpus
and the dialogue act typology used for annotation of
the corpus. Our initial aim was to create software for
automatic annotation of DAs in the corpus. No
method was found which would give sufficient
practical results in the case our complex typology of
DAs and the diverse corpus. For that reason, we
implemented a semi-automatic annotator of DAs
which splits the dialogue text into utterances and
assigns up to five most probable DA tags to every
utterance using the Naïve Bayes classifier. Then a
linguist can confirm the right tags and/or correct
annotation errors.
The structural parts of dialogue are determined
using the DA tags. Different colours visualize the
different parts and make it possible to observe sub-
dialogues (information-sharing and other-initiated
repair) in the main information part.
Communicative strategies determined on the
basis of DAs add a new annotation layer to dialogue.
The values of context factors (coordinates of vectors
which correspond to different strategies) make it
possible to study how the initiative is moving from
one participant to another, where and under which
conditions the strategies are used which are not
topic-related, etc. Taking into account the relation
between DAs and communicative strategies, the
typical structure of the main information part of a
dialogue can be represented also by the strategies
(Fig.7, cf. Koit, 2003).
ICAART2014-InternationalConferenceonAgentsandArtificialIntelligence
450
∗
:/
∗
:,|
:|
:
|
∗
:;
:
|
:
Figure 7: The structure of the main part of information
dialogue: communicative strategies. Notations: [
dialogue or its part; { strategy or sequence which can be
missed; * strategy or sequence which can be repeated; |
variants of strategies; a comment.
Our further work includes the study of the
Estonian conversations by using the software tool.
Our further aim is to develop a DS which interacts
with a user in Estonian and follows norms of human-
human communication.
ACKNOWLEDGEMENTS
This work was supported by the European Regional
Development Fund through the Estonian Centre of
Excellence in Computer Science (EXCS), the
Estonian Research Council (grants SF0180078s08,
ETF9124 and ETF8558), and the Estonian Ministry
of Education and Research (grant EKT11005).
REFERENCES
Allen, J., Core, M. 1997. Draft of DAMSL: Dialog Act
Markup in Several Layers http://
www.cs.rochester.edu/research/cisd/resources/damsl/R
evisedManual/RevisedManual.html.
Aller, S. 2012. Dialoogiaktide märgendamine Eesti
dialoogikorpuses: ülevaade ressurssidest ja
tarkvaraarendus. [Recognition of Dialogue Acts in
the Estonian Dialogue Corpus: Overview of Resources
and Software Development.] Master’s thesis.
University of Tartu. http://comserv.cs.ut.ee/
forms/ati_report/
Bellucci, A., Bottoni, P., Levialdi, S. 2009. WOEB: Rapid
Setting of Wizard of Oz Experiments and Reuse for
Deployed Applications. Dipartimento di Informatica,
Università Sapienza di Roma, Italy.
Bunt, H., Alexandersson, J., Carletta, J., Choe, J.-W.,
Chengyu Fang, A., Hasida, K., Lee, K., Petukhova, V.,
Popescu-Belis, A., Romary, L., Soria, C., Traum, D.R.
2012. ISO 24617-2: A semantically-based standard for
dialogue annotation. In Proc. of LREC-2012,
European Language Resources Association (ELRA),
Istanbul, Turkey, 430–437.
Daelemans, W., Zavrel, J., van der Sloot, K., van den
Bosch, A. 2004. TiMBL: Tilburg Memory-Based
Learner Reference Guide. Technical Report ILK 04-
02. Tilburg University and University of Antwerp.
Dahlbäck, N., Jönsson, A., Ahrenberg, L. 1993. Wizard of
Oz studies: why and how. In Knowledge-Based
Systems, 6, 4, 258–266. doi:10.1016/0950-
7051(93)90017-N.
Fernandez, R., Ginzburg, J., Lappin, S. 2005. Using
Machine Learning for Non-Sentential Utterance
Classification. In Proceedings of the 6th SIGdial
Workshop on Discourse and Dialogue. Lisbon,
Portugal, 77–86.
Field, D., Worgan, S., Webb, N., Wilks, Y. 2008.
Automatic Induction of Dialogue Structure from the
Companions Dialogue Corpus. In Proc. of the 4th
International Workshop on Human-Computer
Conversation, Bellagio, Italy.
Georgila, K, Artstein, R., Nazarian, A., Rushforth, M.,
Traum, D.R., Sycara, K. 2011. An annotation scheme
for cross-cultural argumentation and persuasion
dialogues. In 12th Annual SIGdial Meeting on
Discourse and Dialogue. Portland, Oregon, USA,
272– 278.
Fishel, M. 2007. Complex Taxonomy Dialogue Act
Recognition with a Bayesian Classifier. In
Proceedings: DECALOG'2007 Workshop on the
Semantics and Pragmatics of Dialogue. Rovereto,
Italy, 161–162.
Hennoste, T., Gerassimenko, O., Kasterpalu, R., Koit, M.,
Rääbis, A., Strandson, K. 2008. From Human
Communication to Intelligent User Interfaces: Corpora
of Spoken Estonian. In Proceedings of the LREC-2008
(CD): 6th International Conference on Language
Resources and Evaluation; Marrakech; 28-30 May
2008. (Ed.) Calzollari, N., Chouki, K., Mangaard, B.,
Mariani, J., Ojdik, J., Piperidis, S., Tapias, D.
Morocco: ELRA, 2008, 2025–2032. http://www.lrec-
conf.org/proceedings/lrec2008/pdf/518_paper.pdf.
Hennoste, T., Rääbis, A. 2004. Dialoogiaktid eesti
infodialoogides: tüpoloogia ja analüüs. [Dialogue acts
in Estonian information dialogues: a typology and
analysis.] Tartu: TÜ Kirjastus.
http://dspace.utlib.ee/dspace/handle/10062/18995.
Hutchby, I., Wooffitt, R. 1998. Conversation Analysis.
Principles, Practices and Applications. Cambridge,
UK: Polity Press.
Jokinen, K. 2009. Constructive Dialogue Modelling:
Speech Interaction and Rational Agents. John Wiley
& Sons Ltd.
Jokinen, K. 1996. Cooperative Response Planning in
CDM: Reasoning about Communicative Strategies. In
TWLT11. Dialogue Management in Natural Language
(Semi-)AutomaticAnalysisofDialogues
451
Systems, S. LuperFoy, A. Nijholt, G. Veldhuijzen van
Zanten, ed. Enschede: Universiteit Twente, 159–168.
Keizer, S., Op den Akker, R., Nijholt, A. 2002. Dialogue
Act Recognition with Bayesian Networks for Dutch
Dialogues. In Proceedings of the 3rd SIGdial
Workshop on Discourse and Dialogue. Philadelphia,
USA, 88–94.
Koit, M. 2012. Towards automatic recognition of the
structure of Estonian directory inquiries. In Proc. of
5th Int. Conf. on Human Language Technologies: the
Baltic Perspective: Baltic HLT 2012, Tartu, Oct. 2012.
(Ed.) A. Tavast, K. Muischnek, M. Koit. IOS Press,
2012, 120– 128.
Koit, M. 2011. Automatic Recognition of Dialogue Acts
in Complex Typology. In Proc. of INISTA:
International Symposium on INnovations in Intelligent
SysTems and Applications, Istanbul. (Ed.) Akyokuş, S.
et al.. Istanbul: IEEE, 2011, 485–489.
Koit, M. 2003. The structure of information dialogues: a
case study. In 10th International Conference
Knowledge-Dialogue-Solution. Proceedings: 10th
International Conference Knowledge-Dialogue-
Solution, Varna, Bulgaria. Sofia: FOI-COMMERCE,
2003, 307–314.
Levin, L., Ries, K., Thyme-Gobbel, A., Levie, A. 1999.
Tagging of Speech Acts and Dialogue Games in
Spanish Call Home. In Proceedings of the ACL
Workshop “Towards Standards and Tools for
Discourse Tagging”. Somerset, NJ, USA, 42–47.
Manning, C.D., Schütze, H. 1999. Foundations of
Statistical Natural Language Processing. MIT Press.
Reithinger, N., Maier, E. 1995. Utilizing Statistical
Dialogue Act Processing in VERBMOBIL. In
Proceedings of the 33rd Annual Meeting of the
Association for Computational Linguistics.
Cambridge, Massachusetts, 116–121.
Sinclair, J., Coulthard, M. 1975. Towards an Analysis of
Discourse. Oxford:Oxford University Press.
Stenström, A.-B. 1994. An Introduction to Spoken
Interaction. London and New York: Longman.
Wright, H., Poesio, M., Isard, S. 1999. Using High Level
Dialogue Information for Dialogue Act Recognition
Using Prosodic Features. In Proceedings of an ESCA
Tutorial and Research Workshop on Dialogue and
Prosody. Eindhoven, The Netherlands, 139–143.
APPENDIX: OVERVIEW OF THE
DIALOGUE ACT TYPOLOGY
I. Adjacency Pair Acts
DIALOGUE MANAGING ACTS
1. Conventional acts (greeting, thanking, etc.), e.g.
RIF: Greeting, RIS: Greeting, RIF: Wish, RIS:
Thanking. .
2. Topic change acts (are used to start a new topic
or sub-topic), e.g. TCF: Initiation, TCS: Accept.
3. Contact control acts (typically occur in phone
conversations and are used as formulas which
can be presented as lists), e.g. CCF: Initiation,
CCS: Confirmation).
4. Adjusting the conditions of answer (ACF:
Adjusting the conditions of answer, ACS:
Adjusting the conditions of answer).
REPAIR ACTS
5. Repairs initiated and made by different
participants, e.g. RPF: Non-understanding, RPS:
Repair.
INFORMATION ACTS
6. Directives and grants (request, proposal, offer,
etc.), e.g. DIF: Request, DIS: Giving
information.
7. Questions and answers, e.g. QUF: Closed
yes/no, QUS: Yes, QUS: No.
8. Opinions and responses (assertion, etc.), e.g.
OPF: Assertion, OPS: Accept, OPS: Reject.
II. Non-Adjacency Pair Acts
DIALOGUE MANAGING ACTS
1. Conventional (contact, call, etc.), e.g. RS:
Introduce.
REPAIR ACTS
2. Repairs initiated and made by the same person,
e.g. RP: Self-repair.
INFORMATION ACTS
3. Primary single acts (narration, promise,
rhetorical question, etc.), e.g. PS: Promise.
4. Additional information (specification, softening,
etc.), e.g. AI: Specification.
5. Responses (continuer, acknowledgement, etc. –
acts that traditionally are considered as narrow
feedback), e.g. VR: Neutral continuer.
ICAART2014-InternationalConferenceonAgentsandArtificialIntelligence
452
