Despite increasing relevance in research- and industrial contexts, the implementation of knowledge-based systems (KBS) still is a challenging task. We motivate, that patterns|basically a renowned means for providing reusable solutions for similar problems|can drastically leverage development e orts and time. In this paper, we propose a framework for pattern-driven, encompassing KBS development, consisting of: Tailored usability criteria for a clear delimitation of KBS solutions, a basic pattern speci cation template, and a collection of foundational KBS UI patterns. We further describe practical experiences with the approach, entailing: The reference implementation of several patterns with the tailored development tool ProKEt, their design- and usability-related evaluation, and empirical evidence of applying pattern-driven KBS development in actual projects.
Despite increasing relevance in research- and industrial contexts,
knowledgebased systems (KBS) engineering still denotes a challenging task. In general
software engineering, patterns are renowned for describing proven solutions and
preventing common pitfalls, thus fostering reuse and strongly leveraging overall
development e orts. To date, various pattern collections for general UI- and
interaction design are proposed, including [
In contrast to standard web pages or query forms, KBS do not solely build on strictly prede ned question sequences; rather, they characteristically live from follow-up items| exibly included interview items that become relevant only during the questioning session and based on the concrete user input. Also, KBS often require the prominent integration of additional information for elaborating the knowledge base / interview items more clearly and of (in-place) explanations of the results. This calls for tailored UI and interaction solutions that best support those requirements.
We motivate, that also KBS engineering can strongly pro t from tted
patterns that subsume such proven (and optimally evaluated) KBS solutions; this
can strongly support and re ne requirements engineering, and leverage
encompassing KBS development. First steps into that direction have been already taken
with regards to the knowledge base, e.g., [
The rest of the paper is organized as follows: In Section 2, we introduce a basic KBS pattern speci cation framework: A short de nition of relevant terms, tailored (usability-related) classi cation criteria, and a KBS UI pattern speci cation template. A collection of four framing patterns, that can be ne-tuned into a total of ten pattern variants, is proposed in Section 3. Practical experiences related to pattern-driven KBS development are reported in Section 4, and a summarizing conclusion and promising future work are presented in Section 5. 2
KBS UI Pattern Speci cation Framework Before proposing a set of usability-related KBS delimitation criteria in Section 2.1 and sketching a basic pattern speci cation template in Section 2.2, we clarify some basic terms used in the remaining paper.
empty solution set; from one or more init questions, such a KBS then questions in all directions, depending on the particularly implemented indication mechanisms. In contrast, a backward progression KBS is initialized with a target solution and poses only those questions that contribute to the nal state of that chosen solution.
KBS use forward progression, whereas clari cation KBS base on backward progression. Clari cation KBS can further be used with two application focuses: Consultation focussed|i.e., the clari cation KBS is started empty, and all contributing questions are posed. Or justi cation focussed|then, such a system is called for justifying a solution that already has been derived in the preceding session, thus corresponding to an elaborate, interactive solution explanation. 2.1
Today, diverse UI design- and usability guidelines and standards are available,
such as Nielsen's heuristics [
5. Explorability (Participation): Are users enabled to deviate from the suggested questioning sequence, are potential expert shortcuts provided? 6. Intuition (usage): Are the applied presentation/interaction forms familiar or otherwise self-descriptive? If not, are particularly novice users supported (instructions, tutorials, examples)?
7. Transparency: Is the current state (i.e., state of questions, results, overall progress) clearly and comprehensibly mediated at any time?
8. Clear Arrangement/Aesthetics: Does the overall design exhibit certain aesthetics, e.g., by using a small number of virtual lines & basic symmetry? 2.2
We propose four basic KBS UI patterns: Questionnaire, Interview, Clari er, and
Clari er Hybrid, along with variants. In earlier research, we already introduced
three basic interaction categories for KBS, see [
Problem Statement: The KBS should compactly display a greater part of the KB, o er a highly transparent UI, intuitive usage, and a certain extent of explorability ; comprehensibility, is no key requirement for the core UI itself. Solution: Questionnaire KBS resemble paper- or web-based questionnaire forms. Depending on the particular UI style, many to all indicated interview objects are displayed simultaneously and typically ordered in some form of grid-based layout. Questionnaire may suggest (visually), but does not necessarily prescribe, an optimal interrogation sequence and thus fosters explorative usage. A certain comprehensibility can be achieved by adding auxiliaries|such as informative popups with additional explanations for interview items. Per default, Questionnaire uses forward progression, c.f., Section 2 Variations: Box-, Daily-, and CheckList Questionnaire.
a. Box Questionnaire closely adheres to the design of standard questionnaires by using established, familiar question presentation forms|e.g., checkboxes and radio buttons for choice questions, see Figure 1, I; thereby, each question is rendered within a distinct box, resulting in a very regular layout, but some waste of UI space.
b. Daily Questionnaire, originally inspired by daily newspapers, realizes a more at, juxtaposed presentation style for questions and answers, c.f., Figure 1, II; therefore, each question along with its answer options is placed in a line, implying a less regular yet more compact layout than the Box variant.
c. CheckList Questionnaire mirrors paper-based check lists by representing answer options by columns that simply are ticked o , see Figure 1, III. Therefore, all choice questions need to be mappable to a xed answer set; including further types, e.g., numerical questions, is possible, yet best results regarding e cient interaction and compactness are achieved with choice questions only. 3.2
Problem Statement: The KBS UI should be highly intuitive and easily comprehensible, thus speci cally supporting novice users/domain laymen; in turn, compactness, descriptiveness, e ciency, explorability, as well as UI transparency can be neglected. Solution: Interview imitates human conversations by presenting always a single question|or a group of related questions|at a time; additional information are available anytime quickly and easily, e.g., by directly integrating it near to the corresponding interview object. Interview typically prescribes the interrogation sequence in a rather xed manner. The basic lack of transparency can be alleviated by integrating auxiliaries such as an interview object history|listing the already processed interview items|and a progress information display. Variations: Strict-, Grouped-, and Hierarchical Interview.
d. Strict Interview displays only a single question at a time together with its additional information, see Figure 2, I. Thus, optimally, the KB should provide suitable auxiliary information for each interview item. Further, a sophisticated indication mechanism is advisable for keeping the possible interrogation paths at solid lengths, especially regarding large KBs.
e. Grouped Interview sequentially displays groups of (optimally topically related) questions or single questions; thus, it is a bit more e cient than Strict Interview, and o ers more explorability, as the particular sequence within question groups is not prescribed. The UI uses a similar basic frame as Strict Interview, where Questionnaire variants are used for rendering the groups, c.f., Figure 2, II.
f. Hierarchical Interview o ers an interactively navigable tree UI speci
cally for decision tree knowledge, c.f., [
Problem Statement: Selected single solutions in highly expertise domains should be investigated exclusively and in-depth. The KBS UI should be compact, transparent, descriptive, and o er skill-building ability induced by a high comprehensibility of the contents; users should be enabled to increase e ciency in contributing their personal knowledge, e.g., for using shortcuts regarding the interrogation sequence (explorability/participation ). Intuitive usage, in contrast, is no key requirement.
Solution: Clari er characteristically uses backward knowledge, see Section 2, for investigating only a single issue at a time. Therefore, Clari er renders the target solution and all contributing questions|i.e., that potentially alter the solution rating in any way|simultaneously and o ers means to adapt answers and thus investigate the consequences on the solution.
Variations: Hierarchical Clari er and Form Add-on Clari er.
) I I ( : y r o t s i w e i v r e t
o In C 0 7 9 e 1 l
a ta = m a = e D h f R o I ( ) I ( a n r e e o i n g e e
G R h t s e a n A r a e Y h t e n t a o D M n o i t a r e
e p
p O
O ,
, g
g r a r
g o r t n n
n o o e e u u u S S S , ) I I I ( e i v r e t n I l a c i h c r a r e i e i v r e t n I t t S : 2 H , ) .
I ( h
e x i | t s r n I e n , K t : E 2 g. Hierarchical Clari er exhibits its strengths optimally with backward knowledge that is re ned over several abstraction levels. It displays question and answers options within one node of its tree-style UI, see Figure 2, IV. The topmost node corresponds to the target solution, and is followed by nodes that correspond to the directly solution-relevant questions. Each question node again can be followed recursively by further question levels where the children denote a more ne-granular partition|one or several questions|of its parent. Thus, users decide whether to answer the more abstract top-level questions; or whether to implicitly answer them by expanding them and processing the children|child answers then are propagated recursively back to the parents.
h. Form Add-on Clari er adds minimalistic consultation widgets to static base justi cation presentation forms, such as nding lists or rule graphs, c.f., Figure 1, IV a&b. Clicking the interview objects in the justi cation automatically triggers compact (popup) consultation widgets; those contain all answer options for the respective question, potentially also indicating the value that is added to the solution. This allows for interactively adapting answers and thus exploring and clarifying a selected solution based on its general justi cation view. 3.4
Problem Description: A more intuitively usable and easily comprehensible UI representation for using clari cation knowledge is desired.
Solution: Clari er Hybrids merge intuitive, comprehensible KBS UIs with backward knowledge for supporting especially also novice or laymen users in using clari cation KBS. Both Questionnaire and Interview patterns are suitable for using backward knowledge. The base implementation of Clari er Hybrid then corresponds to the variants described in Sections 3.1 and 3.2; in contrast to that, the targeted backward knowledge is processed, which might|depending on the actually selected UI variant|require some additions to the base UI; for example, widgets for explicitly navigating the hierarchical re nement levels. Variations: Clari cation Interview and Clari cation Questionnaire. 3.5
Table 2 summarizes the ne-grained delimitation of the proposed patterns according to the tailored classi cation criteria, introduced in Section 2.1; the extent of their realization is rated from low (L) to high (H). Further, corresponding target user characteristics for each pattern are speci ed by classifying the domain-expertise and the frequency of usage. Thus, the delimitation table serves as quick reference which pattern to apply in what context. If, e.g., a KBS solution is requested that is both highly intuitive usable, and|given the appropriate knowledge|highly comprehensible, also and especially for rst-time users, then the Strict Interview pattern suits best, c.f., Table 2.
iacompact
recomprehensible
irdescriptive
t
Ce cient
tyexplorable
i
i intuitive
l
b
atransparent
sUclear arranged
Our practical experiences related to KBS UI patterns encompass: Support of
pattern-driven KBS development by the tailored tool ProKEt [
Table 3 summarizes the questionnaire-based results of the comparative study (rating scale: 1/very good { 6/very bad). The rst major nding is a strong correlation between KB quality Q4 and each of KB utility Q1 (0.9968), KBS e ciency Q7 (0.9813), and perceived correctness of the result Q3 (0.9571), correlation coe cient given in parentheses. Further, KB quality Q4 correlates quite strong with the overall KBS success (0.8325); thus overall, the KB quality can be assumed one major in uencing factor regarding the overall KBS perception. This in turn explains the bad overall results for both Interview variants, despite their way more positive rating in the expert usability assessment: Both variants used a qualitatively rather unfavorable statistical KB|con rmed strongly also by subjective remarks. Yet, regarding Strict Interview, at least the basic tendency of the expert assessment|which con rmed a highly intuitive overall UI/interaction|was con rmed, see Q6 and Q9.
Box Questionnaire obviously received the best ratings, closely followed by the Daily variant; along with provided subjective remarks this indicates, that the more structured presentation of the Box variant was favored over the compact Daily layout, thereby consenting with the basic expert assessment ndings. Apart from underlining this tendency, however, subjective remarks speci cally criticized the more space consuming presentation of Box Questionnaire and the general lack of structure in Daily; those comments also revealed more approval for Daily regarding its ease of use, simplicity, and e ciency. Thus, we suspect an even further increase in the overall rating of Daily in case it is further enhanced regarding its presentation|including, e.g., a clearer distinction of question and answer items, a more prominent highlighting of answered items, and providing overall visual structure by indicating separators between question/answer lines.
Regarding Hierarchical Clari er, the ratings may seem improvable; yet, this KBS addressed a highly complex KB from the domain of protection against unlawful dismissal, with an equally comprehensive problem description of the dismissal conditions, the correctness of which was to be rated by the KBS. Thus, an utility value of 2.68 and even the more a success rate of 88.31 % are particularly good results in the given context of a highly expertise domain but domain laymen users. Especially the descriptive and transparent style of Hierarchical Clari er, mirroring the derived question/solution ratings directly in the UI may have supported that result; it most likely fostered the overall trustworthiness Q3 (compared to the Interview variants).
As a general important insight it excelled clearly, that the evaluation of a KBS UI always is inherently coupled with the applied KB and the respective problem to be solved. 4.3
Pattern-driven development along with the tool ProKEt already proved highly
bene cial in actual projects. First, we noticed a strong support of the
requirements engineering process. In the Mediastinitis project|where a documentation
KBS for the structured input of operation data is realized, c.f., [
Another advantage is the fostered reuse of KBS solutions. In the EuraHS
project, c.f. also [
Similarly, in the JuriSearch project, see [
In this paper, we motivated the bene ts of pattern-driven development in the context of KBS. For practical support, we proposed a pattern speci cation framework, based on tailored KBS (usability) criteria and a pattern template, and we introduced a collection of 10 fundamental KBS UI/interaction patterns. Further, we reported practical experiences with the approach: Using the tailored KBS engineering tool ProKEt for creating reference implementations of the patterns, evaluating ve of them regarding their design and usability, and empirical experiences with pattern-driven KBS development from current projects.
Despite denoting an exciting approach for leveraging encompassing KBS development, there are several aspects worth investigating in the future: First, the general extension of the tool ProKEt as to entirely support those patterns that are speci ed theoretically only so far. Second, a thorough usability- and design-related base evaluation of the not yet assessed patterns. Third, an indepth inquiry of one assumption that has emerged in the conducted study: That a structural enhancement of Daily Questionnaire may entail even better results compared to Box Questionnaire. Fourth, follow-up studies for investigating patterns in direct comparison for delimitating core characteristics even clearer|e.g., the pure Clari er variants vs. one or both Clari er Hybrids. Another goal is the actual application of further selected patterns in current (and potential future) projects; e.g., using Clari cation Interview as a rst-time user alternative for the Hierarchical Clari er in the JuriSearch project.