Repair Work in Human-Computer Dialogue 
Alison Cawsey* 
Department of Artificial Intelligence, University of Edinburgh, Scotland 
ajc@uk.ae.ed.aipna 
Pirkko Raudaskoski 
English Department, University of Oulu, Finland 
ekl-pr@finfou.bitnet 
Abstracl;: If human-computer interaction is to be 
effective, it is vi~al that there are opportunities to check on 
understanding, and repair that understanding when it fails. 
This paper discusses this idea of repair in human-computer 
interaction, and provides a number of examples of different 
types of repair work in an interactive explanation system. 
1. Introduction 
The importance of repair in human interaction is 
increasingly recognised. If a dialogue is to pro- 
ceed smoothly it is vital that there are opportanities 
for checking understanding and providing clarification 
when misunderstanding does occur. Everyday interac- 
tion is full of such checks and repairs, though these may 
be so au~;omatic as to be almost transparent, rarely dis- 
turbing ~he flow of the interaction. 
In human-computer interaction, providing opportu- 
nities for clarification may be even more important. If 
the communication is to be robust and effective, then 
there must be opportunities for both parties to 'repair' 
the interaction when it fails. If the user is communicat° 
ing in natural language (or even in a complex command 
language) then there are many cases where the system 
may not 'understand'. If the system is giving complex 
instructions or explanations, then there are many cases 
where the user may not understand. 
These checks and repairs have been studied by peo- 
ple working in the field of conversation analysis (CA) for 
many years. For example, people have analysed pref- 
erences for different types of repair \[7\], and typical se- 
quences of repair moves. 
Recently, there has been some interest in checks and 
repafi" within Cognitive Science (though the approach 
*Supported by a post-doctoral fellowship from the Science and 
Engineering Research Council 
to the subject is often very different from that of CA). 
This includes work by Ringle and Bruce \[5\], who anal- 
yse checking moves and conversation failure, and Clark 
and Schaefer \[2\], who have recently proposed a model of 
dialogae based on contributions rather than single com- 
municative acts. These are the sections of discourse 
through which the participants arrive at the mutual 
knowledge that the conveyed message is understood, 
and may involve checking and repair work. 
Despite the prevalence of checks and repairs in hu- 
man interaction, there has been very little work within 
computational linguistics on these essential components 
of conversation. The rest of this paper will discuss the 
problem in more detail, and present some examples of 
different types of repair work in an implemented inter- 
active explanation system. 
2. Repair in Human Interaction 
In human conversation there are continual implicit 
acknowledgements that communication is proceeding 
smoothly. The speaker is monitoring the hearer in dif- 
ferent ways to see if they understand (for example, using 
checking moves such as 'Do you know what I mean?'), 
and the 'hearer' is often giving verbal acknowledgement 
to the speaker (e.g., 'yes', 'uhuh'). If the hearer takes 
over the conversation, she may acknowledge the last ut- 
terance implicitly by, for example, continuing the topic 
\[2\]. However, if the utterance is not understood, ~ re- 
pair may be initiated. We can examine this repair from 
several perspectives: 
Sequencing: A typicalrepair sequence may consist of 
a repair initiator by the hearer, a repair by the original 
speaker, and an acknowledgement by the hearer. How- 
ever, repair sequences in general may be much more 
complex. For example, the speaker may do a third turn 
327 
repair on realising, from the hearer's response that her 
original utterance was not understood. These different 
types of repair have been discussed in \[7\]. 
ttepalr Initiators: Repair initiators may take many 
forms in human interaction, including: facial expres- 
sion; verbal signals ('huh?') and clarification questions. 
In human dialogue, the speaker may frequently self- 
correct without hearer intervention. 
Source of Trouble: Communication may break 
down for many reasons, such as from lack of hearing, 
reference failure or from general misunderstanding of 
complex material. Although the form of the repair ini- 
tiator may indicate the source of the trouble, this is 
not always the case. It may therefore be necessary to 
guess at the likely source of trouble, possibly using dis- 
course context or assumptions about the hearer's knowl- 
edge to reason about likely problems \[3\]. Repair work 
both relies on, and shapes the context of the interac- 
tion. However, whatever the source of the problem, the 
basic interactional mechanism is the same \[6\]. 
3. Example Repair Work 
In order to illustrate some of these different aspects of 
repair, this section will give a number of examples of 
types of repair work in an interactive explanation sys- 
tem (the EDGE system, described further in \[1\]). These 
include repairs when the user fails to understand the 
explanation, as well as repairs when the system fails to 
understand the user. These latter are adapted from \[4\]. 
The EDGE system plans explanations of the be- 
haviour of simple circuits, depending on assumptions 
about the user's knowledge. These are interactive, with 
many opportunities for repair work when the user fails 
to understand the explanation. The user input to the 
system consists of one or two word commands or ques- 
tions, rather than arbitrary natural language utter- 
ances. However, even with this restricted input there 
in an obvious need for repair work which addresses the 
systems lack of 'understanding' as well as the users. 
3.1 User Misunderstandings 
First, we will illustrate how the system may repair user 
misunderstandings. We must consider both how to 
structure the dialogue, and how to plan the content of 
a specific repair sequence. 
Clarification Questions: Whenever the system 
pauses the user may ask a clarification question (us- 
ing a restricted command language). The system will 
normally reply to this question, then try and get back 
to what it was in the middle of explaining. This is 
achieved using discourse 'plans' to structure the clar- 
ification sub-dialogue, and a simple notion of focus to 
attempt to resume the previous discussion in such a way 
that it follows on from the topic introduced by the user. 
The following example illustrates this: 
S: The light detector circuit's components are: A 
light-dependent-resistor and a fixed-resistor. 
U: What-is-a light-dependent-resistor? 
S: A light-dependent-resistor is a kind of resis- 
tor. Its function is to provide an output resistance 
which depends on the input light intensity. 
S: Anyway, when its input light intensity is high, 
its output resistance is quite low .... 
In this example the system was planning to describe 
the detailed behaviour of the light detector circuit's 
components. Because of the interruption/clarification, 
the system chooses to first describe the behaviour of the 
light dependent resistor. 
Signalling Misunderstanding: The user may also 
signal that they are not following without mentioning 
the exact problem -- maybe they don't know why they 
don't understand. The system must then 'guess' at the 
likely source of trouble. This is done by maintaining 
a model of the discourse so far which includes any as- 
sumptions made about the user's knowledge. In at- 
tempting a repair the system identifies an assumption 
which may have been mistaken, then tries to fill in miss- 
ing knowledge or explain something another way. For 
example: 
U: Huh? 
S: Don't you know what sort of circuit a light- 
detector-circuit is? 
U: No. 
S: OK, A light-detector-circuit is a kind of 
potential-divider circuit. Anyway.. 
System initiated remediation: Sometimes the sys- 
tem can deduce that the user has a misunderstanding 
from the user's utterances (e.g., replies to questions). 
Then a remediation sequence is initiated by the sys- 
tem using strategies based on work on tutorial dialogues 
(e.g., \[8\]). For example: 
S: What's the output voltage of the light-detector- 
circuit? 
U: High. 
328 
S: No, a light-detector-circuit is like a heat- 
detector.circuit except its output voltage de- 
pends on the input \]ight-intentity instead of hear- 
intensity. So, what's the output voltage of the 
light-detector-circuit? 
'Fhese examples illustrate different ways repairs may 
be iniated, how repair sequences may be structured 
w:ithin an ongoing dialogue, and how the system may 
tress at problems or use standm'd remediation strate- 
gies. 
3.2 System ~Misunderstandlng ~ 
Within the dialogue there are also places where the user 
may ask a question, but the system may not be able to 
in~;erpret it. (\['he system must then choose an app,'opri.- 
at~ repair initiator. In these examples the user's input 
is in the form of a simple command hmguage --- for nat- 
urnl language input it is even more important to give 
helpful repair initiators. 
Object mimmderstood: If the object of the ques- 
tion is misunderstood, the repair initiator stmuld direct 
attention to that ~missing' object: 
U: What-is-a light-circuit ? 
S: What.d,~-~ what? 
U: light-detector-circuit. 
S: OK, A light-detector-circuit is a kind of .. 
Ql:mstion type misunderstood: If the question 
type is misunderstood, attention should be directed to 
that: 
U: Whatisa light-detector-circuit ? 
S: What about the light-detector-circuit? 
U: What-is-a light-detector-circuit. 
Both misunderstood: If the question is of the right 
form but both parts are not understood, the system 
simply says 'what?'. 
Wrong form: If the utter~nce is not of a recognisable 
form (e.g., it cannot be. 'parsed')j the system informs the 
user of acceptable forms (e.g., question-type question- 
obj). 
Repeated errors: Repair initiators for repeated er- 
rors give: further information, such ~ lists of relevant 
object and question types. 
These simple examples illustrate tile importance of 
nshtg an appropriate repair initiator when the system 
fails to understand. This is important for both com- 
mand and natural language based input. 
4. Conclusion 
This paper has illustrated ~tle importance, and some of 
the problems of repair work in human-computer dia- 
logues, hnportant issues include repair sequencing, se~ 
lecting and responding to different repair iniators, and 
reasoning about the possible source of the problem and 
helpful 'remediation' strategies. The example system is 
fairly simple, though in an evaluation of an early ver- 
sion with menu-based user input, the interactive/repair 
based approach to explanation generation proved use- 
ful. Future work on any practical natural language di- 
Mogue system should consider these issues. 

References 

\[1\] A. Cawsey. Generating explanatory discom'se. I11 
R. Dale, C. Mellish, and M. Zock, editors, Cur- 
rent Research in Natural Language Generation, Aca- 
demic Press, 1990. 

\[2\] tI. Clark and E. Schaefer. Contributing to discourse. 
Cognitive Science, 13:259-294, 1989. 

\[3\] J. D. Moore. A Reactive Approach to Ezplanation 
in Expert and Advice-Giving Systems. PhD thesis, 
Information Sciences Institute, University of South- 
ern California, 1989. (published as ISI-SR-90-251). 

\[4\] P. Raudaskoski. Repair work in human-computer 
interaction. In P. Luff, D. Frohlich, and N. Gilbert, 
editors, Computers and Conversation, Academic 
Press, 1990. 

\[51 M. Ringle and B. Bruce. Conversation failure. In 
W. Lehnert and M. Ringle, editors, Strategies for 
Natural Language Processing, Lawrence Earlbaum, 
Hillsdale, New Jersey, 1981. 

\[6\] E. Schegloff. Some som'ces of misunderstanding in 
talk-in-interaction. Lecture to the Cognitive Sci- 
ences program, University of California, Berkeley. 

(71 E. Schegloff, G. Jefferson, and H. Sacks. The prefer- 
ence for self-correction in the organisation of repair 
in conversation. Language, 53:361-382, 1977. 

\[8\] B. Woolf and T. Murray. A framework for repre- 
senting tutorial discourse. In Proceedings of the lOth 
International Conference on Artificial fntelligence~ 
pages 189-192, 1987.
