Dialogue tagsets in oncology
Mary McGee Wood
Department of Computer Science
University of Manchester
Manchester M13 9PL U.K.
mary@cs.man.ac.uk
Abstract
Dialogue analysis is widely used in oncol-
ogy for training health professionals in
communication skills. Parameters and
tagsets have been developed indepen-
dently of work in natural language pro-
cessing. In relation to emergent stan-
dards in NLP, syntactic tagging is mini-
mal, semantics is domain-speci c, prag-
matics is comparable, and the analysis of
cognitive a ect is richly developed. We
suggest productive directions for conver-
gence.
1 Motivation
Dialogue analysis systems have been developed
in oncology as a tool for assessing and improv-
ing the communication skills of health profession-
als. Rates of psychiatric morbidity (clinical anxi-
ety and depression) in cancer patients are lowered
when health professionals have adequate commu-
nication skills to discover and address the patients’
concerns and worries. Health professionals inter-
viewing patients sometimes exhibit negative be-
haviours, such as \blocking" a certain line of in-
vestigation rather than encouraging the patient
to describe his or her problem. On the other
hand, a skilled interviewer uses active interven-
tions to direct the progress of the interview, as
well as more passive responses. Several oncology
research groups have demonstrated that these pat-
terns can be detected and quanti ed through anal-
ysis of conversations between health professionals
and patients. This in turn can form a basis for
more e ective training in communication skills.
The Psychological Medicine Group at Manch-
ester (PMG), funded by the Cancer Research
Campaign (CRC), is a leading group in dialogue
analysis in oncology. This paper describes the
parameters and tagsets (analogous to \Dialogue
Act" tagging (Stolcke et al 2000)), which they and
three other groups have developed for this highly
specialised domain.
This domain o ers an interesting contrast to the
\instructional" or \service" dialogues commonly
studied. The health professional is the \expert" in
the conventional sense, and at times conveys med-
ical information to the less knowledgeable patient
in a conventional way. At other times, the patient
should be seen as the \expert" with regard to his
or her own perceived physical and mental condi-
tion, and the task of the health professional is ef-
fectively that of \knowledge elicitation" as under-
stood in expert systems development. This  exi-
ble and dynamic shifting of participants’ roles in a
dialogue poses an interesting challenge, compared
to the clearly de ned and static roles assumed in
much work in dialogue analysis.
2 Parameters for dialogue tagging
Complete and accurate tagging of dialogue must
encode a number of independent aspects of each
utterance. These are represented as \layers"
in the DAMSL system (Core & Allen 1997).
Form-based tags (question, statement) are sup-
plemented with diacritics indicating other types
of information, such as task-management or
communication-management.
The four oncology dialogue tagging systems
considered here all share this basic principle, al-
though they di er in the speci cs. Butow et
al (1995:1115) cite the recognition as early as
1983 of \layers of meaning ... such as the con-
tent, the process, the emotion and the purpose".
Their own CN-LOGIT system encodes three \di-
mensions": \source" (who is speaking), \process"
(questions, responses, initiated statements), and
\content". A complete dialogue can be mapped
into a three-dimensional information space, and
measures can be applied such as how much time
was spent in each cell of the cube. Ong et al
(1998) use the Roter Interaction Analysis Sys-
tem (RIAS). Each utterance in a dialogue is cat-
egorised, and also rated on  ve distinct \global
a ect" scales. The Medical Interaction Process
System (MIPS) of Ford et al (2000) also stresses
the multi-dimensional nature of dialogue annota-
tion, using  fteen \content codes" and eight \af-
fective modes". PMG (Maguire & Faulkner 1988;
Maguire p.c.) have separate tagsets for Form,
Function, Content, Level, Cue, Cue Management,
Blocking, and Focus.
One can see an implicit consensus here that (to
use NLP terms) syntactic form, overt semantic
content, pragmatic force, and cognitive a ect are
distinct and are all signi cant. The di ering de-
grees of detail and prominence they receive in the
di erent systems are discussed under those head-
ings in the next section.
3 Dialogue tagsets
Not surprisingly, the actual tagsets developed in
oncology re ect their domain more closely than
the parameter sets do. In comparison with NLP
work, syntactic classi cation is minimal and func-
tionally oriented, while communication manage-
ment and psychological / emotional loading re-
ceive prominent,  ne-grained analysis.
3.1 Form
Although all four oncology systems encode the
form of an utterance in some way, the classi ca-
tions have a strong pragmatic bias. Questions are
distinguished, not in traditional syntactic terms
as yes-no or wh-, but according to their e ect on
the  ow of the dialogue. The simplest set is that
of Butow et al: Open Question, Closed Question,
Response to Question, Statement, Other. PMG
add Directive Question (open), Directive Question
(closed), Screening Question, Leading Question,
Multiple Question.
Ford et al distinguish \modes" from \con-
tent codes", but even the modes encode coarse-
grained content information as well as a ective
classi cation. The form categories of Ong et al
are \instrumental" (Directions, Question-asking,
Information-giving, &c), and they specify that \if
a decision must be made between categorizing an
utterance in an instrumental or a ect category,
the a ect category should be used" - quite rea-
sonably, given the purpose of their analysis.
Even with a prior commitment to maintaining
separate and independent levels of analysis, some
leakage between levels can occur. (The set of
forty-two Dialogue Act labels used by Stolcke et al
(2000) shows some similar mixing of levels, includ-
ing both purely syntactic tags (such as Declarative
Yes-No Question) and a ective tags (such as Ap-
preciation).)
3.2 Content
The content of an utterance is also encoded in all
four systems, and the tagsets on this level are the
most domain-speci c. Butow et al cite seven con-
tent categories: Treatment, Diagnosis, Prognosis,
History, Other medical matters, Social matters.
Ford et al, with 15 content codes, and PMG,
with 38, are the most fully developed. Both
include Medical (further distinguished by PMG,
with four categories for diagnosis and two for prog-
nosis), Treatment, Psychological, Social, Lifestyle,
&c. PMG are particularly detailed in their cat-
egories for psychological and emotional issues,
shading into the a ect level: Concerns, Feelings,
Emotions, Religion, &c. Again, this is what one
would expect, given that their reason for carry-
ing out the analysis is to assess the health pro-
fessional’s success in getting the patient to talk
about exactly these issues.
Both Ford and PMG also include the opening
and closing of the interview under this heading,
where it sits oddly. A separate level of commu-
nication management, as in DAMSL, would ac-
commodate these and the open/ closed/ directive
question distinction currently made in the Form
tagsets, clarifying all three.
3.3 Pragmatics
As noted above, the Form classes used in the four
coding schemes express more pragmatic than syn-
tactic information. Ong et al’s \instrumental clus-
ters and categories" (Directions, Question- ask-
ing, Information-giving, Counselling) can be con-
sidered pragmatic. So can PMG’s \Function"
codes: eliciting, checking, acknowledgement (psy-
chological, general, cognitions); reassurance, ne-
gotiation, information giving. These are similar
to some of the Dialogue Act labels used in NLP
work: Stolcke et al’s (2000) agreement, response
acknowledgement, summarize, or VERBMOBIL’s
suggest, con rm, clarify (Jekat et al 1995).
3.4 A ect
Cognitive a ect - the psychological force, for a
patient, of an utterance or a complete dialogue -
is the focus of interest in oncology and thus the
most highly developed area. Ford et al pick out
eight of their \modes" as a ective, including the
expression of irritation, gratitude, apology, and
concern.
Ong et al rate both doctor and patient, by cod-
ing their utterances, on  ve distinct \global a ect"
scales: Anger/ irritation, Anxiety/ nervousness,
Dominance/ assertiveness, Interest/ engagement,
Friendliness/ warmth. Their \a ective clusters
and categories" comprise (with subheadings) so-
cial behaviour, verbal attentiveness, showing con-
cern, and negative talk.
PMG do not represent a ect as a separate pa-
rameter, as such. Their function codes include
a ective functions such as Empathy and Reassur-
ance. Many of their content codes can also repre-
sent a ect, as noted above. Topics such as Con-
cerns, Feelings about health care, Religion / spiri-
tual issues can be addressed at any level from sim-
ply factual to deeply emotional, blurring the pic-
ture: this would be clari ed if the a ect level were
explicitly factored out. The most direct represen-
tation of a ective level comes in the two codes
Psychological explicit and Psychological implicit.
Each utterance in a dialogue can be given several
content codes, commonly including one of these
two, as seen in the sample dialogue below.
Cognitive a ect has barely been touched on by
NLP research in dialogue tagging. It is clearly
more subtle and di cult than syntactic, seman-
tic, or pragmatic analysis, and also less signi cant
in instructional or service dialogues than in the
highly charged, life-critical domain of cancer care.
It is, however, an important aspect of dialogue and
speaker modelling, and of the design of appropri-
ate responses. In this area, NLP could learn some
valuable lessons from oncology.
4 An example
Here is a brief typical example from a PMG anno-
tated dialogue. Notice the multiple and somewhat
diverse content codes, and the classi cation of
cue management (somewhat counter-intuitively
attached to the cue utterance itself, not the
response).
P26: I said there’s only another thing that I hope I
never have to have and that’s selectron treatment.
Content: 23 (Psychological implicit)
24 (Treatment)
Level: 1 (Hint)
Cue: 1 (Patient cue)
Cue management: 4 (Cue explored)
N27: Mmmm.
Form: 02 (Response)
Function: 06 (Acknowledgement - general)
Content: 24 (Psychological implicit)
P27: But I says ... ... if I have to I will do, I said
whatever you say, I said.
Content: 24 (Treatment)
14 (Information)
N28: Now why did you say that about selectron?
Form: 04 (Open directive question)
Function: 04 (Clari cation - general)
Content: 24 (Treatment)
P28: No it were me that said that.
Content: 24 (Treatment)
N29: Right,
Form: 02 (Response)
Function: 06 (Acknowledgement - general)
Content: 24 (Treatment)
have you had it before?
Form: 05 (Directive question (closed)
Function: 04 (Clari cation - general)
Content: 24 (Treatment)
13 (History)
P29: I said, I, I don’t mind what you do but I hope
I’ve never to have selectron treatment again, but I
said if I have to, if it’s a necessity then I will.
Content: 23 (Psychological implicit)
24 (Treatment)
33 (The future)
Level: 1 (Hint)
Cue: 1 (Patient cue)
Cue management: 2 (Minimal acknowledgement)
N30: Right.
Form: 02 (Response)
Function: 06 (Acknowledgement - general)
Content: 24 (Treatment)
14 (Information)
P30: But I hope I never have.
Content: 23 (Psychological implicit)
24 (Treatment)
33 (The future)
Level: 1 (Hint)
Cue: 1 (Patient cue)
Cue management: 4 (Cue explored)
N31: And why was that, because you were isolated
or what was it....?
Form: 08 (Multiple question)
Function: 03 (Clari cation - psychological)
Content: 23 (Psychological implicit)
24 (Treatment)
5 Implementations
All the dialogue tagging work in oncology has been
done manually. A few primitive software tools
have been developed to support the tagging pro-
cess and to analyse the results.
Ong et al (1998) have developed a Turbo Pas-
cal \computerized version" of the RIAS coding
system. The advantages they claim for it give one
some idea of average state of its  eld:
\With this program, classi cation of utterances
can be done directly on computer. As a result,
the extra step of entering paper and pencil data
into the computer is omitted. Also, sequential in-
formation is kept. Moreover, because the ten last
classi cations are constantly visible on the screen
there is direct feedback about the ongoing conver-
sation. This provides an important memory aid
with respect to which utterance has to be coded
next. As a consequence, the task becomes less at-
tention demanding and therefore less error- prone.
By giving the opportunity to save the content of
the last coded utterance, an additional memory
aid is provided for shorter and longer breaks."
(Ong et al 1998:400)
Butow’s group have developed a \computer-
based interaction analysis system" with three
parts: \(i) micro level analysis coded in real time
and retaining the sequence of events, (ii) event
counts and (iii) macro level analysis of consulta-
tion style and a ect" (Butow et al 1995:1116). \At
the micro level the aim is to break the consul-
tation down into its components and to charac-
terise, count and/or time them... At the macro
level, the aim is to characterise the consulta-
tion in a more holistic way, such as patient-
centred vs doctor-centred, authoritarian vs a l-
iative or friendly vs hostile." (ibid:1115) All three
forms of analysis depend on counting and timing
utterance-events classi ed according to the three-
dimensional model described above, although Bu-
tow et al stress that they also \retain the sequence
of events". \In future analyses we will explore se-
quential information e ects" (ibid:1120). This is
evidently a signi cant innovation in its  eld. The
fundamental concept of a grammar of dialogue
is simply missing from the oncology work. On
the other hand, their techniques for \macro-level"
analysis of dialogues may well have something to
o er, especially in the subtle areas of modelling
and adapting to speakers’ attitudes and underly-
ing intentions.
6 Prospects
All this work has been developed with care, in the
light of experience, to serve a speci c and unusual
purpose. However, it shows no awareness of dia-
logue tagging work in NLP. Both  elds can bene t
from collaboration.
The author, together with Prof. Peter Maguire
of PMG, has recently been granted support by
CRC to develop practical software support for
the PMG oncology dialogue annotators. This
paper presents a preliminary analysis, part of a
feasibility study for that project. An associated
PhD studentship, awarded by the University of
Manchester Department of Computer Science,
ensures that the NLP perspective will be rep-
resented and the theoretical issues addressed.
We look forward to presenting more detailed
analyses, and original proposals, in the future.
Acknowledgments
Prof. Peter Maguire and Ian Fletcher of PMG
have been generous with their time and support
in the research leading to this paper. We also
gratefully acknowledge the support of the Can-
cer Research Campaign and of the Department of
Computer Science, University of Manchester.

References
Butow, PN, SM Dunn, MHN Tattersall, & Q. J.
Jones. 1995. Computer-based interaction analy-
sis of the cancer consultation. British Journal of
Cancer 71:1115-1121.
Core, M. & J Allen. 1997. Coding Dialogs with
the DAMSL annotation scheme. AAAI fall sym-
posium on Communicative Action in Humans and
Machines.
Ford, S, A Hall, D Ratcli e, & L Fallow eld.
2000. The Medical Interaction Process System
(MIPS): an instrument for analysing interviews of
oncologists and patients with cancer. Social Sci-
ence and Medicine 50:553-566.
Jekat, S, A Klein, E Maier, I Maleck, M Mast, &
J Quantz. 1995. Dialogue acts in VERBMOBIL.
Verbmobil-Report 65, Universitat Hamburg et al.
Maguire, P & A Faulkner. 1988. How to im-
prove the counselling skills of doctors and nurses
in cancer care. British Journal of Medicine 297,
847.
Ong, LML, MRM Visser, IPM Kruyver,
JM Bensing, A van den Brink-Muinen, JML
Stouthard, FB Lammes & JCJM de Haes. 1998.
The Roter Interaction Analysis System (RIAS) in
Oncological Consultations: Psychometric Proper-
ties. Psycho- Oncology 7:387-401.
Stolcke, A, K Ries, N Coccaro, E. Shriberg, R
Bates, D Jurafsky, P Taylor, R Martin, C van Ess-
Dykema, & M Meteer. 2000. Dialogue Act Mod-
elling for Automatic Tagging and Recognition of
Conversational Speech. Computational Linguis-
tics 26(3):339-373.
