Knowledge Acquisition in the construction of ontologies: a case
study in the domain of hematology
Fabrício M. Mendonça1,* Kátia C. Coelho1 André Q. Andrade1,2 and Mauricio B. Almeida3
1
Graduate Program in Information Science, Federal University of Minas Gerais, Brazil
2 Institute for Medical Informatics, Medical University of Graz, Austria
3 Department of Information Theory and Management, Federal University of Minas Gerais, Brazil
ABSTRACT
The activities of organizing knowledge recorded in texts and obtaining 2 BACKGROUND
knowledge from human experts – the knowledge acquisition process –
are essential for scientific development. In this article, we propose 2.1 An overview of Knowledge Acquisition
methodological steps for knowledge acquisition, which have been
applied to the construction of biomedical ontologies. The methodologi- The KA activity generally includes the collection, analysis,
cal steps are tested in a real case of knowledge acquisition in the do- structuring and validation of knowledge for representation
main of the human blood. We hope to contribute to the improvement
of knowledge acquisition for the representation of scientific knowledge
purposes (Hua, 2008). It is an activity composed of a set of
in ontologies. tasks that employ computer-based and manual techniques
(Gaines, 2003; Boose & Gaines, 1989; Shadbolt, 2005). A
multitude of definitions for KA can be found (Shaw &
1 INTRODUCTION Gaines, 1996; Scott & Clayton, 1991; Payne et al, 2007) and
Ontologies have been proposed as an alternative for creating the theories and methods that support KA activities rely on
representations of reality suitable for computers. At least diverse academic research fields. Ways of acquiring and
four activities are essential in the development of ontolo- representing knowledge come from Computer Science
gies: specification, knowledge acquisition, conceptualizatio (Compton & Jansen, 1989), Cognitive Science (Hawkins,
and formalization. In knowledge acquisition (KA), the expe- 1983), Linguistics (Campbell et al, 1998) and Psychology
rience available in the literature of diverse fields mentions (Harris, 1976).
difficulties in communication between experts and profes-
sionals who deal with information (Boose, 1990). 2.2 Classification of KA techniques
This article investigates the activity of KA within the KA techniques can be classified into manual techniques and
scope of biomedicine. In order to explore the activity, we computer-based techniques (Boose, 1990). In general, the
propose procedures for KA employing the best practices manual techniques are rooted in Psychology (Kelly, 1955)
referenced in the literature. We systematize these proce- and computer-based techniques are classified as automatic
dures in a list of methodological steps with the aim of test- or semi-automatic. KA can be classified according to the
ing their feasibility in a real case. knowledge obtained in the process. The assumption that
The empirical research is conducted within the scope of a different methods of elicitation result in different types of
biomedical project, focused on human blood. The know- knowledge is known as the differential access hypothesis
ledge acquisition results have been used in the development (Hoffman et al, 1995). In addition, KA can be classified
of a knowledge base for scientific and educational applica- according to application methods such as protocol-
tions related to the human blood. Descriptions of different generation techniques, protocol-analysis techniques, matrix-
stages of research are provided as examples throughout the
based techniques and sorting techniques (Shadbolt & Swal-
article. The main contributions are the aforementioned list
low, 1993).
of steps and observations made in real situations with the
Protocol-generation techniques include interviews. The
aim of improving the KA performance.
The remainder of this paper is organized as follows: sec- most well-known technique for interviews is the teachback
tion 2 reviews the literature on KA. Section 3 explains the technique (Hua, 2008; Shadbolt, 2005). Protocol-analysis
theoretical rationale, the systematization and tools that com- techniques are used in the transcription of interviews in or-
pose the KA methodology. Section 4 presents comments of der to identify different knowledge types. Matrix-based
interest during the next phases of the research. Finally, sec- techniques involve the diagrammatic organization of prob-
tion 5 puts forward our final remarks. lems. The most well-known technique is the repertory grid
(Hua, 2008; Shadbolt, 2005). Sorting techniques are tech-
niques in which the domain entities are classified in order to
check how an expert classifies the knowledge. The most
* To whom correspondence should be addressed: fabriciommendon-
well-known technique is card sorting (Hua, 2007; Hoffman
ca@gmail.com
1
Fabrício M. Mendonça et al.
et al, 1995). The Diagram-based technique consists of the In the activity of codification we employed Sketch En-
creation and use of network representations, such as concep- gine2, an online tool for the creation and analysis of linguis-
tual maps (Corbridge et al, 1994). A methodology for KA tic corpora. The fragmentation of the text into morphemes
that combines card sorting and laddering can be employed and the identification of the grammatical classes are auto-
in the construction of ontologies (Wang et al, 2006). matically performed.
After the codification activity, we proceeded with the in-
2.3 KA in Biomedicine
formation retrieval from the corpus with the aim of identify-
Natural Language Processing (NLP) techniques are com- ing terms used to describe blood transfusion procedures. In
mon in the biomedical domain (Hersh, 2009; Verspoor et al, order to do so, we used word suffixes common of medical
2006). These techniques can be divided into two main terms (Lovis, Baud & Rassinoux, 1998) as such -apheresis,
streams: the rule-based approach (Friedman et al, 2004; -centesis, -desis, -ectomy, -opsy, to mention but a few. Then,
Hahn, Romacker & Schulz, 2002) and the statistical ap- we built regular expressions using the Sketch Engine corpus
proach (Taira & Soderland, 1999; Sebastiani, 2002). query language, in order to retrieve terms related to proce-
A comparison between the two methods involved the test- dures, as well as the absolute frequencies that occur in the
ing of systems using both approaches to the automatic cate- corpus.
gorization of MEDLINE abstracts (Humphrey et al, 2009) As a final task of the extraction phase, we analyzed the
and found comparable results for most evaluated items. The morphological productivity of the terms obtained using the
results favored the statistical approach, though the authors British National Corpus (BNC)3 as a reference. The analysis
suggested the combination of both approaches. consisted of comparing the frequency of each term in the
corpus with its frequency in the reference corpus. In order to
3 METHODS proceed with the morphological productivity analysis we
3.1 Case study: knowledge context and domain used the AntConC4 tool.
In the elicitation phase, we made use of the terms obtained
This work explores the best practices in an ongoing KA
in the extraction phase, which were employed as guidelines
scenario applied within the scope of the Blood Project (Al-
to start the contact with experts. This phase consisted of
meida, Proietti & Smith, 2011), an information organization
holding interviews and the application of KA techniques
initiative in hematology. The project is taking place in a
with experts, doctors, biologists and researchers. During the
medical institution responsible for hematology and blood
course of the interviews, sorting and matrix techniques were
transfusion research and that offers healthcare services for a
applied. The cycle that characterizes the clinical process,
population of around 20 million people.
ranging from the development of an infectious disease
3.2 Methodological steps through its treatment, was adopted to guide the approach
In this section, we describe the list of steps for KA. Then, taken with the experts. For modeling the domain, we
we present a synoptic table summarizing the tasks involved adopted the disease as disposition approach, as proposed by
and systematizing the steps in the list, which was divided (Scheuermann, Ceusters & Smith, 2009). The three major
into four main phases: extraction, elicitation, validation and stages that comprise that cycle are: etiological process,
refinement. course of disease and therapeutic response. In order to apply
In the extraction phase we applied NLP techniques and the described reasoning so far, a template was created in
tools in order to obtain candidate terms for the ontology. Protégé-Frames.
KA from texts consists of three main activities: construction In the stage called etiological process, there is a healthy
of a corpus related to blood transfusion, codification of this human body with characteristics that are normal according
corpus and information retrieval from the corpus. to medical parameters. In the pre-clinical manifestation of
The subset of the corpus related to blood transfusion uses the disease, the body develops disorders, which are bearers
the manual of the American Association of Blood Banking of dispositions. Such dispositions are naturally associated
(AABB) as a source. From the AABB website1 we down- with the entities’ existence, for example, the disposition of
loaded thirty-two chapters that comprise the seventeenth the human body to get sick (Smith, 2008). There are
edition of the manual. From this material, twenty-seven changes in the patient already, but not noticed. The etiologi-
chapters were processed by the tool used for codification. cal process stage can be represented as follows:
This material was select as a sample according to the stage ETIOLOGICAL PROCESS => produces => DISORDER
of the research underway when writing this paper. Certainly, => bears => DISPOSITION.
in future works, diseases processes and clinical finding will
be considered. 2
Available at: . Access: Dec. 15, 2010
3
Available at: . Access: Nov. 30, 2011
4
Available at: . Access: July 23, 2011
1
Available at: . Accessed: July 23, 2010
2
Knowledge Acquisition in the construction of ontologies: a case study in the domain of hematology
The course of disease stage starts with the clinical manife- Update data
Wiki Page
3.2 updating after each vali-
station of the disease (disposition). At this moment, the dis- dation
K. engineer
order manifests itself through symptoms, which the patient 4.1 integra- Characterize
-Template Protégé
is able to identify. Then, a doctor identifies the disease signs (4)
tion between related genes,
-K. engineer
through a physical exam or through a report of the patient. granularities proteins, etc
Refine-
4.2 connec- Connect data
In this stage, it is possible to determine the clinical pheno- ment -Template Protégé-
tion with top- with other on-
- K. engineer
type, that is, the principal observable characteristic of that level tologies
disease. The course of disease stage can be represented as Table 1: KA list of steps proposed
follows: DISPOSITION => realized in =>
PATHOLOGICAL PROCESS => produces => 4 RESULTS
ABNORMAL BODY FEATURES. One evident result is the methodological list of steps de-
In the therapeutic response phase, a sample is taken from scribed in the previous section, which has been tested and
the infected part of the body in order to perform laboratory improved over the course of the research (Table 1).
tests. At this point, it is possible to establish a treatment plan In the codification activity (extraction phase), from the texts
so that the body may return to normality. The plan is the selected 369,741 tokens were automatically identified and
result of a diagnosis founded in the interpretative process of related to parts-of-speech. Subsequently, in the information
a clinical framework. The clinical framework is composed retrieval phase, 57 terms related to blood transfusion proce-
of symptom representation records as well as physical and dures were identified. Table 2 depicts the top-five terms
laboratory exam results. The therapeutic response stage can from the set of 57 terms retrieved, which were used as a
be represented as follows: ABNORMAL BODY basis for starting interviews with experts:
CONDITION => recognized as => SIGN AND SYMPTOM Term Frequency
=> used in => INTERPRETATIVE PROCESS.
apheresis 124
The third phase of the proposed list of steps for KA, called
phlebotomy 32
the validation phase, uses wiki science tools for collabora-
cytometry 20
tive validation of candidate terms for an ontology. After the
elicitation phase, according to the knowledge obtained, can- cordocentesis 16
didate terms are transferred to a wiki to then be validated by plasmapheresis 15
experts online. Table 1: top-five terms retrieved and absolute frequency
The fourth stage of the proposed list of topics, called the
refinement phase, uses a second template, also created using The rationale applied in the elicitation phase made it poss-
Protégé-Frames. The goal was to record information about ible to understand the major stages of the disease manifesta-
how to integrate the different levels of granularity required tion. Table 2 presents an example of blood disease analysis
to understand a disease and its manifestations. This integra- following this rationale for Bernard-Soulier Syndrome:
tion involves obtaining the relations between parts of the
body that a certain disease affects, the related genes and the Etiological inheritance of a defect in the platelet membrane recep-
process tor that affects the hemostasis
related proteins.
platelets with a glycoprotein Ib complex (GP Ib) ab-
Finally, the steps put forward so far are gathered together, Disorder normality, either quantitative (absence of GP Ib) or
thus creating the list of steps for KA. qualitative (mutation of GP1BA, GP1BB, GP9)
Resources and Disposition Bernard-Soulier Syndrome (A, B or C)
Phase Task Description
people involved
1.1 build a Create a corpus -Medical texts Pathological abnormal platelet adhesion to the extracellular matrix
corpus from texts -K. engineer process during the initial phase of plug formation
(1) Symptoms bleeding, hematomas
1.2 codifica- Automatically -Sketch Engine tool
Extrac-
tion fragment texts -K. engineer excessive bleeding, gingival bleeding, menorrhagia,
tion Signs
1.3 informa- Obtain terms -Sketch Engine tool purpura, epistaxis, gastrointestinal bleeding
tion retrieval through suffixes - K. engineer Table 2: KA reasoning applied to a blood disease
-Template Protégé
2.1 obtain Hold interviews
and teachback;
knowledge with experts An example of a Protégé-Frames template related to Ber-
-K. engineer, experts
-Matrix Techniques nard-Soulier syndrome is depicted in Fig. 4.
(2) 2.2 know the Identify ex-
-K. engineer and
Contact terminology perts’ rationale
expert
2.3 see ad- Understand how
-Sorting techniques -
hoc organiza- experts sort
-Experts
tion concepts
(3) Obtain approval
3.1 validate -Wiki Page
Valida- of terms ac-
knowledge -Expert
tion quired
3
Fabrício M. Mendonça et al.
worth noticing that the difficulties in the validation stage did
not occur among experts validating their own prior know-
ledge. Rather, the majority of cases of non-validity occurred
when an expert evaluated the knowledge provided by anoth-
er expert. However, the differences did not seem irreconcil-
able. In many cases, experts suggested referring to their own
scientific publications to resolve outstanding issues.
iv) The refinement stage was conducted in the same way
as the contact stage. Indeed, it was conducted as an inter-
view merged with work to understand the rationale behind
and organization of the experts’concepts. When analyzing
the results, one can conclude that this stage provides useful
insights into the building of ontologies in terms of interope-
rability. This is because the refinement stage is based on the
premise of connection to top-level ontologies.
Observations made over the course of all these stages al-
lowed us to identify problems that occur in the KA process
Fig. 4. Protégé-Frames template with example about blood for which solutions have been sought as the research has
disease continued. These problems are the result of the influence of
the following factors:
Finally, it is worth mentioning that at the time this article i) factors related to the expert profile, such as: training,
was being written, the ontology developed in OWL had experience and previous participation in similar projects,
more than 300 classes and 50 properties, and practically all limitations in expertise;
the methodological steps were up and running, providing ii) contextual factors, such as: cultural, geographical, po-
data for different ontology parts. litical and financial issues, lack of access to information
sources and deficiency in organizational structure;
5 DISCUSSION iii) factors related to the interaction between expert and
In each stage of the KA process, as depict in Table 1, it is knowledge engineer, such as: short-term outlook (KA is
possible to identify issues to be discussed: seen as “additional work”) and domain complexity;
i) The extraction stage was undertaken mainly by a know- iv) factors that make recording results difficult, such as:
ledge engineer using NLP tools applied to sources suggested non-approval by the expert of the results of the activity and
by experts. As a means of producing a list of relevant terms constant advancement in the scientific field.
in a domain, the extraction was useful in preventing the Concerning the proposed elicitation technique (section
knowledge engineers from having to start interviews from 3.2), which is based on Scheuermann, Ceusters & Smith
scratch. In general, the terms selected were useful for de- (2009), one can argue that there is a methodological pitfall
scribing the domain according to the opinions of experts. when using a formal disease model to acquire knowledge. It
ii) The contact stage is the heart of KA processes, since it could be argued that relevant domain knowledge could be
is within this stage that experts share their knowledge. This missed by doing so, because what would be acquired is
stage was conducted as a cycle that involved interviews something of a pre-conceived frame of meaning. However,
interspersed with attempts to understand the rationale used we observed that some sort of structure was required to
by experts to understand the phenomena in the domain. As conduct the activity and save time, mainly considering the
part of this attempt, the knowledge engineer employed sort- limited availability of the experts. According to our expe-
ing and matrix techniques. Regarding the interview based on rience in this study of case, knowledge missed for this rea-
an ontological disease model, it is worth reporting that the son may be dealt with using complementary techniques. The
results were very reasonable, insofar as the experts approved interviewees were not constrained when talking and teach-
of the framework organized in the etiological process, back techniques were employed to give them the chance to
course of disease and therapeutic response proposed by clear up misunderstandings and flaws. In addition, the onto-
Scheuermann, Ceusters, & Smith (2009). logical disease model was used only to organize the inter-
iii) The validation stage was conducted, in many cases, view and to make notes, not in an attempt to formalize
during the interviews, mainly in the beginning of the knowledge directly.
process when experts didn´t have experience with Wiki The NLP techniques applied aimed at collecting candidate
pages. In general, the validation confirmed the interviews terms for the ontology, instead of trying to populate it di-
and the teachback technique performed previously. It´s rectly. In this sense, the use of those techniques was impor-
tant to obtain a first list of candidate terms. Even though
4
Knowledge Acquisition in the construction of ontologies: a case study in the domain of hematology
NLP is not considered a good source for ontological know- Gaines, B.R. Organizational Knowledge Acquisition. In: Handbook on
ledge, it may be useful when dealing with a large volume of knowledge management. Birkhäuser: Springer. 2003, 700 p.
material. Another issue when using NLP was the size of our Hahn, U., Romacker M., Schulz, S. Medsyndikate - a natural language
sample: in order to build a significant corpus, one should system for the extraction of medical information from findings reports.
International Journal of Medical Informatics. 2002;67:63-74.
have at least 10 million words, which are not available to us.
Harris Z. On a theory of Language. The Journal of Philosophy, v. 73, n. 10,
p. 253-276 1976.
6 CONCLUSION Hawkins, D. An analysis of expert thinking. International Journal of Man-
This article has proposed a list of steps for KA, which are Machine Studies. v. 18, p. 1-47, Jan. 1983
based on techniques found in the literature. The steps in the Hersh, W. Information Retrieval: A Health and Biomedical Perspective
list has been tested, proving their viability. The work de- 3ed: Springer 2009.
scribed includes a project in which research was conducted Hoffman, R.R., Shadbolt, N.R., Burton, A.M., Klein, G. Eliciting know-
ledge from experts. Organizational Behavior and Decision Processes.
to identify the best practices for and difficulties in perform-
v. 62, n.2, 1995. pg 129-158.
ing the KA activities with hematology experts within the
Hua, J. Study on Knowledge Acquisition Techniques. 2nd Inter. Symp. on
scope of creating an ontology. The list of steps is a partial Intelligent Information Technology App. 2008.
result that has been improved based on direct observation. Humphrey, S.M., Neveol, A., Browne, A., Gobeil, J., Ruch, P., Darmo-
One conclusion we could draw from the overall expe- ni,SJ. Comparing a Rule-Based Versus Statistical System for Automat-
rience is that KA is a very time-consuming and expensive ic Categorization of MEDLINE Documents According to Biomedical
process. This may explain why it is neglected in many cas- Specialty. J Amer Soc for Inf Sci and Tech. 2009 Dec; 60:2530-9.
es. In future work, we intend to further clarify in which con- Kelly, G.A. The psychology of personal constructs. New York: Norton,
text each technique is most suitable. This could be done 1955.
with assistance from experts, taking in account their time Lovis, C., Baud, R., Rassinoux, A.M., Michel, P. A., Scherrer, J.R..(1998).
limitations. Regardless, in this case study, some techniques Medical dictionaries for patient encoding systems: a methodology. Art.
Int.in Medicine. Vol. 14, Issue 1, pp. 201-214.
were chosen, as was mentioned in last column of Table 1.
Milton, N., Clarke, D., Shadbolt, N. Knowledge engineering and psycolo-
The list of topics has been successfuly applied in other re-
gy. Int. J. of Human-Computer St., v. 64, n. 12, p. 1214-1229. 2006.
lated domains. It appears to be a systematized alternative for Payne P.R, Mendonça E.A, Johnson S.B, Starren J.B. Conceptual know-
creating ontologies using a rational means of approaching ledge acquisition in biomedicine: a methodological review. J Biomed
experts. Inform. 2007. v 40, n. 5, p. 82–602.
Sebastiani, F. Machine learning in automated text categorization. ACM
ACKNOWLEDGEMENTS Comput Surv. 2002 Mar; 34:1-47.
This work is partially supported by Fundação de Amparo à Scott, A.C., Clayton, J.E., Gibson, E.L. A practical guide to knowledge
acquisition. Addison–Wesle, 1991, 509 p.
Pesquisa do Estado de Minas Gerais (FAPEMIG), Governo
Shadbolt, N. Eliciting Expertise. In: Evaluation of Human Work. Ed. Tay-
do Estado de Minas Gerais, Brazil, Rua Raul Pompéia,
lor & Francis. 2005.
nº101 - São Pedro, Belo Horizonte, MG, 30.330-080, Brazil. Shadbolt, N., Swallow, S. Epistemics: Knowledge Acquisition. Shaw,
This work partially supported by CNPQ (Conselho Nacional M.LG. and Gaines, B.R. Soft.Engineering Journal, 1996. p. 149-165.
de Pesquisa e Desenvolvimento) Smith, B. New Desiderata for Biomedical Terminologies. In Munn, K.;
Smith, B. (Ed.). Applied Ontology. Frankfurt: Verlag. 2008. pp. 21-39
REFERENCES Scheuermann, R.H., Ceusters, W., Smith, B. Toward an ontological treat-
Almeida, M.B., Proietti, A.B., AI J., Smith, B. The Blood Ontology: an ment of diasease and diagnosis. Proceedings 2009 Summit on Transl.
ontology in the domain of hematology. (2011) Proccedings of ICBO. Bioinf., San Francisco, CA, pp. 116-120.
Boose, J.H. Knowledge acquisition tools, methods, and mediating represen- Taira, R.K and Soderland, S.G. Statistical natural language processor for
tations. In: Japanese Knowledge Acq. for KBS, JKAW, 1. 1990. medical reports. J. Am. Med. Infor. Ass. 1999:970-4.
Boose, J.H and Gaines, B.R. Knowledge Acquisition for Knowledge-Based Verspoor, K., Bretonnel, C. K., Goertzel, K., Mani, I. Linking natural lan-
Systems. 1989. Machine Learning, v. 4, p. 377-394. guage processing and biology: towards deeper biological literature
Campbell, K.E. et al. Representing thoughts, words, and things in the analysis. BioNLP ’06: 2006; Morristown USA: ACM; 2006. p. iii-iv.
UMLS. J Am Med Inform Assoc., v. 5, n. 5, p. 421–31, 1998. Wang Y, Sure Y, Stevens R, Rector, A. Knowledge elicitation plug-in for
Compton, P. and Jansen, R. A philosophical basis for knowledge acquisi- Protégé: Card sorting and laddering. In: The Semantic Web, ASWC, v.
tion. Knowledge Acquisition. European KA for KBS. 1989. 4185, p. 552-565, 2006. Available from:
Corbridge C, Rugg G, Major N, Shadbolt N.R, Burton A. Laddering: tech- http://dx.doi.org/10.1007/11836025_53
nique and tool use in knowledge acquisition. J. of Knowledge Acqui-
sition, 1994, p. 315-341. Available from: https://blog.itu.dk/SLR-
F2010/files/2010/07/paper-1-pages-1-12-15.pdf
Friedman, C, Shagina, L., Lussier, Y., Hripcsak, G. Automated encoding of
clinical documents based on NLP. Journal of the American Medical In-
formatics Association. 2004 Sep-Oct;11:392-402.
5