Other-Initiated Self-Repairs in Estonian Information Dialogues: 
Solving Communication Problems in Cooperation 
 
Olga Gerassimenko 
Department of  
Estonian  
University of Tartu 
50409 Tartu, Estonia 
gerro@ut.ee 
Tiit Hennoste 
Department of  
Estonian  
University of Tartu 
50409 Tartu, Estonia 
hennoste@ut.ee 
Mare Koit 
Institute of Computer 
Science 
University of Tartu 
50409 Tartu, Estonia 
koit@ut.ee 
Andriela Rääbis 
Department of  
Estonian  
University of Tartu 
50409 Tartu, Estonia 
andriela@ut.ee 
 
 
Abstract 
The paper gives an overview of repair se-
quences used in Estonian spoken information 
dialogues. 62 calls for information, travel bu-
reaus, shops or outpatients’ departments are 
analysed. Several repair types are considered. 
Our further aim is to develop a dialogue sys-
tem which can interact with the user in Esto-
nian following the norms and rules of human-
human communication 
1 Introduction 
Our goal is to build a dialogue system that would be 
able to interact with humans in Estonian using norms 
and rules of human-human communication. Human-
human communication is not fluent, problems that 
must be solved arise continuously. Similar problems 
occur in human-computer interaction. 
Several branches of linguistics study how to solve 
communication problems: discourse analysis, psy-
cholinguistics, foreign language learning, conversa-
tion analysis, dialogue modelling on the computer 
and determining dialogue acts (Allen, Core, 1997; 
Linell, 1998; Bunt, 1999; Traum, 1999; Allwood et 
al. 2001; O'Brien, 2002). Similar phenomena are 
called differently: structure shifts, feedback, ground-
ing, repair, communication strategies. There are at 
least the following difficulties with these treatments: 
1) repair is considered together with other dia-
logue management acts and not as a separate phe-
nomenon as it is in natural conversation 
2) different means of repair are considered instead 
of repair as a process and sequence 
3) non-natural utterances are used in dialogue sys-
tems for repairs (To Paris, is that correct? / I don't 
quite understand ) 
To overcome these problems, we analyse repair 
using the methods of the conversation analysis (Sche-
gloff et al 1977). There are 4 kinds of repairs differ-
enced in CA: self-initiated self-repair, other-initiated 
self-repair, self-initiated other-repair, other-initiated 
other-repair. All the repairs analyzed in this paper are 
other-initiated self-repairs. We differentiate three 
subtypes here: clarification (45 cases), non-
understanding (7), and reformulation (27). 
We have analyzed calls for information (asking 
phone numbers, addresses etc.), and calls to travel 
bureaus, shops and outpatients' departments. The 
analysed sub-corpus consists of 62 dialogues; 2472 
dialogue act tags were used in our sub-corpus, among 
them 79 repair initiations (about dialogue acts see 
Hennoste et al 2003). 
2 Other-Initiated Self-Repair 
In the case of clarification and non-understanding the 
hearer initiates the repair and the partner solves the 
problem. Non-understanding means that hearer didn’t 
hear or didn’t understand something in speaker’s 
previous utterance. There are two main possibilities 
to initiate repair: general question which only indi-
cates that there was a problem (e.g. mida/what, kui-
das/pardon), or wh-questions to localise the problem 
exactly. The problem is solved by repeating, elaborat-
ing or explicating the problematic information. 
In our dialogues, the non-understanding is initi-
ated mostly to indicate non-hearing (6 examples from 
7) by using general questions kuidas/ pardon or ut-
terance ei kuule/I don't hear. The problem is solved 
by repeating the problematic information, either ex-
actly or with modifications.1 
                                                           
1 In examples, the participants are C (client) and O (infor-
mation officer). Simplified transcription is used (see Ap-
C: aga kallis see tööluba on. (0.5) 
how much does this work permit cost 
>O: kuidas |PPE: NON-UNDERSTANDING| 
pardon 
>C: kallis tööluba on. |PPJ: RE-
PAIR| 
how much does the work permit cost 
O: ei, tö- tööluba ei ole vaja. 
no work permit is needed 
 
Clarification means that hearer is not sure whether 
she has understood some partner’s utterance cor-
rectly. Repeating of the problematic part of utterance 
+ tag-question particle (e.g. jah/yes) are used to initi-
ate the repair. The problem is solved by agreeing 
with it or rejecting it. 
Clarification is initiated in two ways in our data: 
 - a phrase or declarative sentence that repeats the 
problematic part (70%) 
- the same + a question-particle jah/yes, vä/or, 
ühesõnaga / in short (23%). 
The problems are solved mostly by using 
- jah, jaa /yes, mhmh (60%) 
- the same + repeating or adjusting the previous turn 
(12%). 
 
O: nii on pakkuda (0.5) kümme kaks-
kümend.(.) 
I can propose (0.5) ten twenty 
>C: kümme kakskümend. |PPE: CLARI-
FICATION| 
ten twenty 
>O: jah. |PPJ: REPAIR| 
yes 
 
In reformulation the hearer initiates the repair and 
suggests her own interpretation of the problematic 
issue. The partner agrees with it or rejects this inter-
pretation. The most frequent ways for repair initiation 
are:  
- word, phrase, sentence (22% of cases) 
- the same + a questioning particle siis/then, jah/ yes, 
eks/eh, vä/or (30%) 
- et/that, tähendab/it means that + word, phrase, sen-
tence (33%). 
The problem is solved by 
- jah, jaa / yes, mhmh / yes (41%) 
- jah, jaa / yes + repeat or explanation (33%) 
- explanation (22%). 
 
O: ja m präegu no ütleme keskmine 
hind on oleneb nüüd hotellist ekso-
                                                                                       
pendix). Only repair acts are marked. Repair initiations are 
marked by PPE: NON-UNDERSTANDING  / CLARIFI-
CATION / REFORMULATION, problem solving acts are 
marked by  PPJ: REPAIR. 
le see on kuskil öheksa (.) tuhat 
(.) keskmine hind.  
and now let’s say the average price 
is it depends of hotel it is about 
nine (.) thousand (.) average price 
>C: nädal. |PPE: REFORMULATION| 
[for] a week 
>O: jaa. |PPJ: REPAIR| 
yes 
3 The Structure of Repair Sequence 
3.1 Prototypical Repair Sequences 
In the prototype case, the client asks and the officer 
answers in information dialogues (the roles can be 
reversed in some cases, e.g. when the officer asks 
client's personal data). There are two prototypical 
locations of repair sequence: after client’s ques-
tion/request and officer’s answer. They give us two 
prototypical repair sequences. 
 
   A. Problem in question/request 
MAIN LINE: question/request of the client 
REPAIR SEQUENCE: 
> Repair initiation by officer 
> Problem solving by client  
BACK TO MAIN LINE: officer's answer  
 
   B. Problem in answer 
MAIN LINE: officer's answer  
REPAIR SEQUENCE: 
> Repair initiation by client 
> Problem solving by officer 
BACK TO MAIN LINE: client’s response (mhmh/ 
ahah / jah) + (new) question / request / finishing con-
versation 
 
69% of repairs that concern question/request belong 
to prototype A. This forms 14% from all repairs. 44% 
of repairs regarding answer belong to prototype B. 
This makes up 27% of all the repairs (75% of them 
are clarifications). In sum, 71% of non-under-
standings, 40% clarifications and 33% reformulations 
are prototypical.  
3.2 Untypical Repair Sequences 
First, there exist peripheral variants of prototypical 
repair structures. The remaining examples in group A 
do not constitute clear sub-groups. In group B there 
are some more clearly deviating repair sub-groups. 
 
1. A participant initiates repair regarding information 
in some earlier turn. 
Repairs are initiated regarding the immediately 
previous turn in 90% of cases, but there are some 
examples, where client clarifies information which 
(s)he got earlier in the conversation and breaks the 
process of giving information by consultant, as in the 
following example. After repair C gives response by 
continuer mhmh. Only after this O continues to give 
information. 
O: kolmkümmend viis ruutmeetrit 
dushsh on seal olemas, (.) mõõturid 
on seal olemas (1.0) [korras] 
it has got thirty five square me-
ters, a shower, it has got water 
meters [in order] 
>C: [ja tuhat kolmsada.] |PPE: CLA-
RIFICATION| 
[and one thousand three hundred] 
>O: jah. (.) |PPJ: REPAIR| 
yes 
C: mhmh  
mhmh 
O: elektri boiler on seal olemas, 
(0.5) ja ta on tühi. 
it has got a boiler (0.5) and it's 
empty 
 
The structure is as follows: 
O giving information 
>C repair initiation regarding earlier information 
>O problem solving 
C continuer mhmh 
O continues giving information 
 
2. The answerer/consultant returns to the main line, 
not the client, as it is in prototypical B type repair. 
For example, the client initiates a repair regarding 
the beginning of an e-mail address and breaks the 
officer's dictation of the address. In this case, the of-
ficer (O) must return to the main line, not the client 
(C). There are two possibilities to do this: 
- O carries out a repair and continues giving informa-
tion in the same turn 
- O carries out a repair, C gives back the turn using 
particle mhmh and O continues to give information. 
Therefore, there are two structures: 
 
O giving information 
>C initiation of repair 
>O problem solving + continuation of giving infor-
mation 
 
O giving information 
>C initiation of repair 
>O problem solving 
C mhmh 
O continuation of giving information 
 
3. Some turns are added after the repair before return-
ing to the main line (8 cases). 
The problem is that officer believes that his/her 
answer given before repair initiation was finished, 
but client does not. So the client does not return to 
the main line after problem solving but sends the turn 
back by using jah/yes or mhmh/yes. Then the officer 
sends turn back again and only after that the client 
returns to the main line. The structure is as follows: 
O: answer (giving information) 
>C: initiation of repair 
>O: problem solving 
C: jah / mhmh 
O: jah / mhmh / adjusting or repeat 
C: (mhmh +) new question / finishing conversation / 
adjusting or justifying the request 
 
Second, there are some repair sequences with re-
versed roles - the officer asks questions and the client 
answers. 
 
4. Simple structures with reversed roles 
Such repairs occur first of all at the end of conversa-
tion where the client and the officer negotiate the 
next meeting and the officer asks client's address or 
phone number. Two types occurred.  
First, conversation continues with reversed roles 
after repair, the officer asks a new question: 
C giving information 
>O initiation of repair 
>C problem solving 
O repeat + new question 
 
O: ä (.) kes te olete.  
what's your name 
C: ja nimi on Saabas. (.) 
and the name is Saabas 
>O: Saabas jah? |PPE: CLARIFICA-
TION| 
Saabas yes 
>C: jah. (.) |PPJ: REPAIR| 
yes 
O: Saabas. a ku- kus te asute (0.8) 
see on Tartus jah?  
Saabas. where do you live (0.8) it 
is in Tartu yes? 
 
Second, officer wants to return to normal roles and 
gives the turn back to client using the particle mhmh 
or repeating earlier information: 
C giving information 
>O initiation of repair 
>C problem solving 
O mhmh /repeat 
C new question / finishing conversation 
 
O: ja nimi on  
and the name is 
C: Saabas, Arvo. Saabas Arvo 
(6.5) ((O writes down the name)) 
>O: Arvo jah? |PPE: CLARIFICATION| 
Arvo yes 
>C: jah. |PPJ: REPAIR| 
yes 
O: teeme siis nimodi et kümme kaks-
kümend. (0.5)|  
let's do so that (let’s meet) at 
ten twenty 
C: aitäh?  
thank you 
 
5. The second group are sequences where officer 
cannot answer the client’s question immediately but 
needs some additional information (6 examples). So 
he/she uses one or two adjusting questions at first and 
then initiates repair. Client solves the problem. And 
then the answer to the client’s question follows. The 
structure is as follows: 
C question / request 
(O adjusting conditions of answer) 
(C giving information) 
O adjusting conditions of answer 
C giving information 
>O initiation of repair 
>C problem solving 
(O adjusting conditions of answer) 
(C giving information) 
O answers to the C's question/request 
4 Conclusion 
We are not able to direct the behaviour of humans 
who communicate with the computer. Nevertheless, 
we can design our dialogue system in such a way as 
to minimize the reasons of repairs by the user. The 
primary strategy of the computer would be to provide 
clear and exhaustive information so that the need to 
clarify or re-formulate would be unnecessary. The 
computer itself must initiate repairs to clarify what 
information the user wants. The repair mechanism of 
the computer that simulates information officer can 
be prototypical (type A) because this occurs mostly 
in human-human communication. 
Our further work will concentrate on the analysis 
of a bigger corpus in order to find out and explain 
more repair structures which could be modelled in a 
dialogue system. 
Acknowledgement 
This work was supported in part by Estonian 
Science Foundation under Grant No. 5685. 

References 
James Allen and Mark Core 1997. Draft of DAMSL: 
Dialog Act Markup in Several Layers; 
http://www.cs.rochester.edu/research/cisd/resource
s/damsl/RevisedManual/RevisedManual.html 
Jens Allwood, Elisabeth Ahlsen, Maria Björnberg, 
and Joakim Nivre, 2001. Social activity and com-
munication act-related coding. Cothenburg Papers 
in Theoretical Linguistics 85. Dialog Coding – 
Function and Grammar. Göteborg Coding Sche-
mas. Ed by Jens Allwood. Goteburg. Pp 1- 28. 
Harry Bunt, 1999. Dynamic Interpretation and Dia-
logue Theory. The Structure of Multimodal Dia-
logue II. Ed. by M. M. Taylor, F. Neel, and D. G. 
Bouwhuis. John Benjamins, Philadelphia/ Amster-
dam, pp. 139-166. 
Tiit Hennoste, Mare Koit, Krista Strandson, Andriela 
Rääbis, Maret Valdisoo, and Evely Vutt, 2003. 
Developing a Typology of Dialogue Acts: Some 
Boundary Problems. Proc. of the 4th SIGdial 
Workshop on Discourse and Dialogue, Sapporo, 
5-6 July, 2003, pp 226-235. 
Cormac O’Brien, 2002. Grounding Strategies in Dia-
logue Systems. 
http://www.ida.liu.se/~nlplab/gslt/papers/cormac.ps 
Emanuel Schegloff, Gail Jefferson, and Harvey 
Sacks, 1977. The preference for self-correction in 
the organization of repair in conversation. Lan-
guage 52 (2): 361-382. 
David R. Traum, 1999. Computational Models of 
Grounding in Collaborative Systems. Working Pa-
pers of the {AAAI} Fall Symposium on Psycho-
logical Models of Communication in Collabora-
tive Systems. American Association for Artificial 
Intelligence. Menlo. Park, California. Ed. By 
Susan E. Brennan, Alain Giboin, and David 
Traum. Pp 124-131. 
