PII: 0950-7051(93)90039-V Towards competent information acquisition interactions between an expert system and its user E T Keravnou, J Washbrook* and F Dams* Expert systems are consultative, highly interactive systems, and hence the quality of interaction between system and user is important for the acceptability of the system. Acquisition o f data about the problem in hand is a central feature of the interaction between system and user, and it makes a major contribution to the user's perception of the system. It is there- fore crucial that the acquisition of this data (both as individual data items and as a sequence of data inputs) is perceived by the user as competent. This paper identifies central, domain-independent, design goals for the information-acquisition interactions between an expert system and its user, including mixed-initiative interac- tion, flexibility in user input, and system competence in query- ing the user. The paper discusses the realisation of these goals in a diagnostic expert system, Skeletal Dysplasias Diagnos- tician, through an explicit data model which allows for the representation of data with temporal and spatial (locality) aspects and the decide-status function which operates on the data model. Keywords: information-acquisition interactions, intelligent data handling, medical diagnostic expert systems, competent expert systems, skeletal dysplnsins, background knowledge E x p e r t systems, especially those t h a t dispense advice in critical d o m a i n s such as medicine, engineering an d finance, m u s t be perceived by their users as c o m p e t e n t if they are to be accepted. R e c o m m e n d i n g solutions fo r real p r o b l e m s which are subsequently p r o v e n in real life to be c o r r e c t is o f course indisputable evidence o f a Department of Computer Science, University of Cyprus, Kallipoleos 75, PO Box 537, Nieosia, Cyprus *Department of Computer Science, University College London, Gower Street, London WC1E 6BT, UK Paper received 7 January 1992. Revised paper received 7 September 1992. Accepted 8 October 1992 system's competence. H o w e v e r , fo r a system to progress f r o m the research l a b o r a t o r y into a real-life environ- ment, it m u s t n o t o n l y be able to solve test p roblems co rrect l y b u t also to interact intelligently with the user. It is t h r o u g h the dialogue with the system t h a t the user forms i m p o r t a n t first impressions. T h u s the i n t eraction between the system an d its user is o f prime i m p o r t a n c e if the system is to inspire confidence in p o t en t i al users whilst still in the l a b o r a t o r y , a n d to retain their confi- dence once it has progressed to a real-life setting. T h e principal spheres o f interaction are i n f o r m a t i o n acqui- sition a n d explanation, b o t h o f which m u st be carried o u t in a w ay which is acceptable to, an d tailored to the needs of, the p art i cu l ar user ~. Acquisition o f i n f o r m a t i o n a b o u t the p r o b l e m in h a n d m a y be t h r o u g h the system asking questions, o r the user volunteering i n f o r m a t i o n . A t the beginning o f a consultation, initial i n f o r m a t i o n is acquired t h r o u g h a p r e d e t e r m i n e d sequence o f general questions asked b y the system a n d items o f d a t a volun- teered b y the user. T h e initial i n f o r m a t i o n acquisition process is thus mixed-initiative, in t h a t b o t h the system an d the user m a y initiate a p a r t i c u l a r interaction. It is possible f o r an expert system t o require t h a t all p r o b l e m d a t a be entered at the beginning o f the consul- tation, w i t h o u t the necessity (or possibility) o f subse- q u e n t i n f o r m a t i o n acquisition. This, however, is based o n a simplified an d restrictive m o d el o f the co n su ltation process. A m o r e sophisticated m o d el w o u l d allow for system-initiated i n f o r m a t i o n acquisition at subsequent stages in the consultation. H o w e v e r it is arguable (see below) t h a t i n f o r m a t i o n acquisition should c o n t i n u e t h r o u g h o u t the consultation, a n d should be potentially mixed-initiative (i.e. b o t h the user a n d the system should be able to initiate i n f o r m a t i o n input) at all stages. In addition, t o be perceived as c o m p e t e n t , the system should take into a c c o u n t all relevant i n f o r m a t i o n pre- viously acquired b efo re asking questions. This m a y require the m a k i n g o f substantial chains o f inferences. An expert system which accurately models the reason- 0950-7051/931030141-16 © 1993 Butterworth-Heinemann Ltd Knowledge-Based Systems Volume 6 Number 3 September 1993 141 Towards competent information acquisition interactions: E T Keravnou et aL ing and knowledge o f d o m a i n experts will also provi d e a model o f their information-acquisition strategies. This can be used as the basis for system information-acqui- sition interactions. Such interactions are dynamically constructed, being driven by the derivation o f the solu- tionL T h e y can be analysed at two levels. At a high level the sequence is structured in terms o f the hierarchy o f f o r e g r o u n d problem-solving strategies that have been instantiated. At a low level each s u b g r o u p o f interactions c o r r e s p o n d i n g to a terminal problem-solving strategy can be analysed in terms o f b a c k g r o u n d auxiliary strate- gies. T h e f o r e g r o u n d strategies represent d o m a i n - e x p e r t problem-solving processes, whilst the b a c k g r o u n d strate- gies provide indirect b u t indispensable s u p p o r t to these processes by organising the p r o b l e m d a t a and subse- quent system-initiated i n f o r m a t i o n acquisition 3. F o re- g r o u n d problem-solving strategies m a y be domain-speci- fic, while b a c k g r o u n d data-handling strategies, which often represent common-sense knowledge and reason- ing, are usually domain-independent. O t h e r aspects o f an information-acquisition interac- tion model, such as those related to the specifics o f con- ducting a particular task, or technical aspects related to the design o f user interfaces, are outside the scope o f this paper. Although the w o r k presented here is in the context o f a diagnostic expert system, c o m p e t e n c e in the informati o n - acquisition interactions is relevant to every consultative expert system, diagnostic or other. This p a p e r discusses information-acquisition interactions in the context o f the diagnostic system S D D (Skeletal Dysplasias Diagnos- tician) 4. T h e design goals identified constitute essential aspects o f the overall information-acquisition interaction model o f this system s . T h e first section identifies central, domain-indepen- dent, design goals for the information-acquisition inter- actions between an expert system and its user. These design goals have been realised in the medical diagnostic system, Skeletal Dysplasias Diagnostician 4, t h r o u g h b a c k g r o u n d strategies based on an explicit d a t a model and the decide-status function which operates on the d a t a model. The second, third and f o u r t h sections give a detailed description o f the decide-status operation; some o f this material has been a d a p t e d f r o m Reference 6. T h e fifth section discusses h o w the d a t a model an d decide-status function are used to achieve the required design goals for the information-acquisition interactions between an expert system and its user. I N F O R M A T I O N - A C Q U I S I T I O N I N T E R A C T I O N S : D E S I G N G O A L S I m p o r t a n t design goals for the information-acquisition interactions between an expert system and its user are listed below. Achieving these goals is necessary if the system is to be deemed competent. • Mixed-initiative interaction • Flexibility in user-volunteered information: o T h e user must be allowed to use the entire d o m a i n vocabulary. T h e user must be able to revoke information. T h e system must detect a conflict in the user- volunteered i n f o r m a t i o n as soon as it occurs a n d resolve it with the user. • System competence in querying the user." 0 It must be evident to the user t h at the system has t ak en notice o f the i n f o r m a t i o n volunteered by the user. 0 Th e system questions must be focused a n d meth- odical. 0 T h e r e should be n o n - r e d u n d a n c y in system questions. 0 Th e system must detect and eliminate a n y redun- d an cy in the user information. M i x e d - i n i f i a t v e interaction Th e information-acquisition interaction between the system an d the user must be allowed to be mixed-initia- tive where the user wishes it to be so. Th e user must be able to volunteer i n f o r m a t i o n initially a n d be given the o p p o r t u n i t y to volunteer fu rt h er i n f o r m a t i o n at subse- q u en t stages. Similarly, the system m u st be able to request fu rt h er i n f o r m a t i o n by querying the user if the i n f o r m a t i o n available at a n y stage is n o t sufficient for the system to p ro ceed to the next stage, o r to m a k e firm recommendations. W h ere necessary, a p p r o p r i a t e guidance should be given to the user b o t h when volun- teering i n f o r m a t i o n an d responding to a system question 7. T h e system should o f course be able to p e r f o r m a consultation even i f the user does n o t wish to volunteer information. This can be initiated t h r o u g h a predeter- mined sequence o f questions. Th e ex p l an at i o n model o f an expert system is separate f r o m the information-acquisition interaction model s . As m e n t i o n e d above, the sequence o f questions raised by the system is intrinsically related to the system's reasoning 2. Traditionally, the central role o f the ex p l an at i o n model is to reveal this reasoning 9. However, the explanation model has a subsidiary role in relation t o i n fo rm ation- acquisition interactions. This role concerns individual items o f i n f o r m a t i o n rat h er t h an the system reasoning processes. T h e user needs to be able to ask, n o t only why the system is asking a particular question (i.e. h o w does it relate to the reasoning process), b u t also w h at the parti- cular question means. Also, the system needs to be able to recognise when i n f o r m a t i o n input by the user is either unclear o r meaningless in the co n t ex t o f the consultation, and m u st have the ability to q u e r y the user, asking for clarification. F l e x i b i l i t y in user-volunteered information F o r the user to take full ad v an t ag e o f the facility to volunteer i n fo rm at i o n , the system m u st allow the user flexibility in the means by which i n f o r m a t i o n is entered. 142 Knowledge-Based Systems Volume 6 Number 3 September 1993 Towards competent information acquisition interactions: E T Keravnou e t al. Flexible vocabulary A menu-driven interface facilitates the entry of user information and ensures that every piece of information is meaningful to the system. However the user may also wish to volunteer information 'freely', without having to go through a menu, especially if the user has domain expertise and does not need guidance in deciding what information to enter and how to word it. This facility is also useful in cases where the domain vocabulary requires a complex network of menus for covering all the possible expressions. The flexibility to enter information directly does not necessarily imply a completely unconstrained natural language interface (which is anyway beyond the capabili- ties of the current technology), but rather the use of the entire domain vocabulary (which will undoubtedly include redundant terminology) albeit within specific syntactical constraints which should be simple and con- venient. The system should be capable of dealing with redun- dant terminology by correlating different expressions which have the same meaning. Internally the system will probably use a non-redundant set of terms, into which it will translate user terms. It is important, though, that when the system interacts with the user it either uses the user's terms, or makes it clear that it has translated the user's terms. Conflict detection and resolution and retraction Another aspect of flexibility in user interaction is for the system to be able to detect and resolve conflicting infor- mation input, and for the user to be able to retract previously volunteered information. Conflicting information input may occur if the user information summarises lower-level data (e.g. sensor readings, images, direct observations etc.). The inherent uncertainty of this lower-level data can lead to misinter- pretations. The user may wish to enter possibly conflicting data, provided that a facility to retract previously volunteered information is available. The system should be able to detect and reveal an inconsistency as soon as it occurs and should attempt to resolve it by consulting the user. This should give the user more confidence in the system; the user who is aware of the inconsistency would be surprised if the system did not detect it, whilst the una- ware user would be favourably impressed. In both cases it would be evident to the user that the system is taking notice and 'making sense' of the information entered. The user should therefore be able to volunteer infor- mation either directly, in the form which he or she is used to in reporting findings, or through menus, depending on inclination and convenience, and should also be given the option to retract information subsequently. System competence in querying user Being unable to use the entire domain vocabulary or revoke information could reduce the usability of the system without necessarily affecting the confidence of the user in the system's judgement. However, that confi- dence will be seriously undermined if the system's ques- tioning is perceived as incompetent or unintelligent. For a system to be deemed as competent in its questioning, the following requirements must be satisfied. Use of previously given information The system must be able to show the user that inform- ation entered has been accepted and processed by the system, and is being used both in the diagnostic process and as a basis for any subsequent information acqui- sition. For this to happen, the generation of (partial) solu- tions (in a diagnostic domain this would be the gene- ration of hypotheses) must be driven by input infor- mation; thus subsequent questioning based on the currently entertained (partial) solutions would relate back to what the system had already been told, making the questioning focused and methodical. (The details of hypothesis generation and exploration refer to the fore- ground problem-solving strategies rather than the back- ground data-handling strategies and hence are outside the scope of this paper.) One significant way of showing the user that the system is utilising input information is for the system to ask questions which naturally follow on from the data volunteered, or to reveal inconsistencies in the user- volunteered information. Furthermore, if some infor- mation volunteered by the user needs clarification or triggers other questions, these clarifying or prompting questions should usually be raised by the system imme- diately after receiving this information; this constitutes intelligent behaviour. Non-redundant questions It is important that questions raised by the system should be non-redundant, in that their answers should not be deducible from information already entered. In response to a redundant question the user is justified in saying 'but I told you so already', or 'I've told you I don't know this', or 'I couldn't possibly know this at the present point in time'. If many redundant questions are asked, the user will get justifiably irritated. To ensure non- redundancy, the system should be able to make intelli- gent inferences on the given information; these should include the correlation of statements with equivalent meanings. Another aspect of intelligent inferencing on input information is the ability to recognise whether a particu- lar item of information would be unknown or not, based on information that the user has already specified as unknown, and, in some domains, based on the temporal context defined by the particular problem. This is par- ticularly important in most diagnostic domains. In the domain of skeletal dysplasias diagnosis, for instance, a feature may be unknown because the relevant radio- graphs are unavailable, or because the patient is too young for certain features to be observable. The system should also show that it understands dependencies between items of input information by removing redundant information in the light of new input. When new information volunteered by the user subsumes previously given information, possibly by being more specific, the system should make it evident to Knowledge-Based Systems Volume 6 Number 3 September 1993 143 Towards competent information acquisition interactions: E T the user that it is aware o f the r e d u n d a n c y and has eliminated it; this should inspire confidence and respect. Since the information volunteered by the user is likely to be c o m m u n i c a t e d back to the user by the system in different contexts (for example in run-time explanations, or in summing up final recommendations), there should be ample scope for the system to d e m o n s t r a t e to the user that it has this capability. Summary The main requirements for c o m p e t e n t information acquisition by an expert system are mixed-initiative interactions, flexibility in the entry o f user-volunteered information, a n d c o m p e t e n c e o n the p a r t o f the system in querying the user. These requirements can be achieved if the system has an explicit and flexible d a t a model which is used b y an intelligent reasoner to d r a w inferences from an actual set o f data. W e have implemented such a d a t a model and a reasoner, the decide-status function. D E C I D E - S T A T U S F U N C T I O N : A N I N T E L L I G E N T D A T A R E A S O N E R The decide-status function is an intelligent reasoner a b o u t problem-specific d a t a o r findings. It provides the platform for achieving the seemingly diverse goals dis- cussed above. Earlier w o r k on an a t e m p o r a l decide-sta- tus function has been reported in Reference 10. Decide-status operation The decide-status operation is first illustrated through an example from the d o m a i n o f skeletal dysplasias. Consider the following p r o b l e m definition. Domain-Model: synonyms((metaphyses flared) (long- bones dumbbell-shaped)) (carpal-centres poor-ossification) =:, (carpal-centres small) (carpal-centres small) =~ (epiphyses small) Problem findings: (long-bones dumbbell-shaped from-birth) (carpal-centres poor-ossification from- birth) (metaphyses irregular from-birth) Query finding." (and (metaphyses flared irregular [ta tb]) (epiphyses small [ta tb])) The d o m a i n m o d e l consists o f t w o s y n o n y m o u s findings and two implications. Each o f the p r o b l e m findings has the qualitative-temporal aspect 'from-birth' which trans- lates to the time interval [0 now]. The query finding is a c o m p o u n d finding consisting o f the t w o simple findings • metaphyses being b o t h flared a n d irregular during the time-interval [ta tb], Keravnou et a l • epiphyses being small during the same time interval. M e t a p h y s e s and epiphyses constitute the subjects o f the respective findings (finding subjects). The decide-status function reasons b a c k w a r d s f r o m the q u e r y finding. A goal tree, having the q u e r y finding at its r o o t is being (implicitly) constructed (see Figure 1). The nodes o f the goal-tree n a m e findings and the branches strategy applications. The leaf n o d e o f a 'suc- cessful' (i.e. n o t pruned) branch names a problem finding. The decide-status o p e r a t i o n is simply described as deciding the truth status o f a finding as true, false, or unknown, given a g r o u p o f findings which collectively hold in some context. A firm answer (true or false) returned b y the decide-status function for a finding quer- ied against the user observations can be directly corro- borated, e.g. through observation. The decide-status function operates on an explicit data model. Formalisation o f data model The d a t a model has a representational aspect, i.e. it provides a formal language for representing findings (data), and an inferential aspect, i.e. it provides a formal set o f relations between findings which form the channels through which intelligent inferences a b o u t an actual set o f findings can be drawn. The representational and infer- ential aspects o f the decide-status d a t a model are as follows. Representational aspect: grammar for expressing findings: (finding):: = (compound-finding)l(simple-finding)1 ( a t o m i c - f i n d i n g ) ( c o m p o u n d - f i n d i n g ) : : = (and (findings))l(or (findings)) ( f i n d i n g s ) : : = < f i n d i n g ) ( f i n d i n g s ) t ( f i n d i n g ) (simple-finding):: = ( ( f i n d i n g - s u b j e c t ) ( a t t r i b u t e - v a l u e s ) ( t i m e - i n t e r v a l ) ) ( a t t r i b u t e - v a l u e s ) : : = n i l t ( a t t r i b u t e - v a l u e ) ( a t t r i b u t e - v a l u e s ) ( f i n d i n g - s u b j e c t ) : : = symbol ( a t t r i b u t e - v a l u e ) : : = (locality-value)l (non-locality- v a l u e ) (locality-value):: = symbol (non-locality-value):: = symbol ( t i m e - i n t e r v a l ) : : = nil l(closed-interval)l(open-interval)l (open-from-left)1 ( o p e n - f r o m - r i g h t ) (closed-interval):: = [ ( b a s e ) (limit)] (open-interval):: = ( ( b a s e ) ( l i m i t ) ) ( o p e n - f r o m - l e f t ) : : = ( ( b a s e ) ( l i m i t ) ) ( o p e n - f r o m - r i g h t ) : : = [ ( b a s e ) (limit)] (ba se):: = integer (limit):: = integer ( a t o m i c - f i n d i n g ) : : = ((finding-subject) (simp-att- v a l s ) ( t i m e - i n t e r v a l ) ) (simp-att-vals):: =nill ( l o c a l i t y - v a l u e ) (non-locality- value )1 ( locality- value )1 ( no n-locality-value ) ( a t e m p o r a l - f i n d i n g ) is a ( f i n d i n g ) w i t h o u t temporal- aspect(s) 144 Knowledge-Based Systems Volume 6 Number 3 September 1993 Towards competent information acquisition interactions: E T Keravnou et al. (and (metaphyses flared irregular [ta tb]) (epiphyses small [ta tb]))? (metaphyses flared irregular (epiphyses small [ta tb]) [ta tb]) ~ COMPOSITION (metaphyses flared [ta tb]) MATCHER INFERENCER (carpal-centres small [ta tb]) l INFERENCER (metaphyses irregular [ta tb]) I MATCHER (metaphyses irregular [0 nowl) (carpal-centres poor-ossification [ta tb]) MATCHER ]1 (carpal-centres poor-ossification [0 now]) (long-bones dumbbell-shaped [ta tb]) I MATCHER (long-bones dumbbell-shaped [0 nowl) ]1 key: problem-f'mding *matches if [ta tb] is in [0 now] Figure 1 Decide-Status Knowledge-Based Systems Volume 6 Number 3 September 1993 145 Towards competent information acquisition interactions: E T Keravnou e t al. finding subject I ~ f'mdings for subject F y location and any time ) . . . . . . . . . . . . . . . . . . . . . . . s / / I ,'locality t t I t time Figure 2 Space of findings is three-dimensional space of find- ing subject, time and locality Inferential aspect: model entity relationships: is-a((finding-subject ) (finding-subject)) part-of((finding-subject) (finding-subject)) syn-subjects((finding-subject)(finding-subject)); symmetric relation negative-value((finding-subject) (attribute-value)) status-value((finding-subject)(attribute-value)) normal-subject((finding-subject)) abnormai-subjeet((finding-subject )) negative-finding((finding)) positive-finding( (finding) ) most-generai-positive-finding((finding)) status-finding((finding)) prompting-query((atemporal-finding) (atemporal- finding)) elarifying-query((atemporal-finding) (atemporal- finding)) synonyms-value((attribute-value) (attribute-value)); symmetric relation opposites-value((attribute-value)(attribute-values)); symmetric relation synonyms((atemporal-finding) (atemporal-finding)); symmetric relation impHes((atemporal-finding) (atemporal-finding)) alternatively expressed as: (atemporal-finding) => (atemporal-finding) when-to-ask((atemporal-finding) (earliest-time)) primary-trigger((finding) (hypothesis)) secondary-trigger((finding) (context-hypothesis) (alternative-hypothesis)) Decide-status operation The entire space of findings is partitioned along three dimensions, finding subject, time and locality, as shown in Figure 2. The decide-status operation is formally des- cribed as follows. Let Q be a query for some problem, expressed as holds(F)?, where F is a finding. The decide-status func- tion Decide-Status determines whether Q is decidable from the theory P constituting the problem, i.e. whether P ~ Q or P ~ ~ Q If Q is undecidable the reply unknown is returned. The problem theory P = domain-model U problem-findings, where the domain model is the instantiation o f the infer- ential aspect of the data model for the particular domain, and the problem fndings are the instantiations of the representational aspect o f the data model, i.e. assertions about which findings hold for the specific problem. The domain model therefore defines the finding subjects, attribute values and temporal aspects in the domain, as well as the relationships between these entities (e.g. status and negative values for finding subjects, synonymous finding subjects, synonymous findings, opposite attribute values). The meaning o f these relationships (given above) will be clarified when the decide-status operation is detailed. Decide-Status determines whether Q is true, false or unknown from the set of problem findings given the domain model, by attempting to construct a p r o o f for either Q or ~ Q. If this is not possible, Q is unknown. The p r o o f is constructed in a 'backwards-reasoning' or 'goal-driven' fashion where Q forms the goal, yielding a focused search strategy. Q ( ~ Q) follows from a set of problem findings if it is directly subsumed or implied by a subset o f the problem findings. The domain model is used to explicate relationships between the query finding F and the problem findings. Starting with the query Q at the root of the p r o o f tree, the goal-decomposition stra- tegy is applied until the goal is decomposed into simple subgoals, corresponding to simple queries (i.e. F is decomposed to simple findings). The domain model is used to generate subgoals from a given goal when that goal cannot be further decomposed. Subgoals generated from the domain model may be compound (e.g. the antecedents o f implications may be compound findings) which then need to be decomposed. Subgoals are themselves query findings. A tree branch terminates successfully if the query finding constituting its leaf node is directly determined from the problem findings, i.e. it directly matches, or conflicts, with problem findings. On the other hand, a tree branch is pruned if the simple, undetermined query finding corres- ponding to its leaf node cannot be mapped further through the domain model. A simple goal is determined if either itself or at least one o f the subgoals (which the goal can be mapped onto using the domain model) is directly determined from the problem findings, i.e. the subgoal leads to a successful tree branch. For a conjunc- tive compound finding to be determined as true all its component findings must be determined as true. For it to be determined as unknown at least one o f its component findings must be undetermined whilst the remaining component findings are determined as true. Lastly, deter- mining one o f the component findings as false suffices to 146 Knowledge-Based Systems Volume 6 Number 3 September 1993 Towards competent information acquisition interactions: E T Keravnou e t a/. determine the c o m p o u n d finding as false. A disjunctive c o m p o u n d finding is determined as true, false or u n k n o w n if at least one o f its c o m p o n e n t findings is determined as true, all o f its c o m p o n e n t findings are false, or at least one o f its c o m p o n e n t findings is u n k n o w n whilst the remaining are false, respectively. Thus if the current subgoal o f some simple goal is undetermined, the d o m a i n model is used to generate an alternative subgoal for t h a t goal. A t the top level there are two generic strategies corresponding to two alterna- tive derivation methods: a matcher a n d an inferencer. The m a t c h e r a n d inferencer strategies are described in detail in the third a n d f o u r t h sections, respectively. M A T C H E R The matcher strategy is invoked for a particular finding F queried relative to a group o f findings G, if F has relevant findings in G or F has relevant s y n o n y m s in the d o m a i n model. finding subject I I t I / , I I I I ) - - t s ~ _ y _ _ - e , , " I X ,locality I ) J S • tl i i I i s S ' s - j J s t t a time / I I I I ,'I / 1 • I • I I I I I t - S " J . - ) J - - J tb R e l e v a n t f i n d i n g s The relevant findings o f a finding F from a group G o f findings is the subset o f G defined as follows: relevant-finding s( F, G) = {P c G I subject(P) = subject(F) and overlapping(time-interval(P), time-interval(F)) a n d overlapping(locality(P), locality(F)} i.e. those elements o f G which have the same subject as F and whose time-interval a n d locality attributes overlap with the corresponding time-interval and locality attri- bute o f F (if the time or locality are n o t given explicitly, the defaults 'now' and 'some-relevant-locality' are assumed, respectively). Finding subjects are discrete, time intervals specify a c o n t i n u o u s period o f time, a n d locality attributes can specify a collection o f discrete, (physically) disjoint localities, or a c o n t i n u o u s space o f adjoining localities. A finding, therefore, defines a subs- pace on the entire space o f findings given in Figure 2. The subspace could be a single point if the finding's time interval refers to a time point a n d its locality to a single discrete locality. The relevant findings o f a finding, from a group G o f findings, are those findings whose subspaces overlap with (e.g. are contained in) the subspace o f the given finding (see Figure 3). L o c a l i t y attribute Consider the following findings f r o m the d o m a i n o f skel- etal dysplasias, expressed in natural language: 'all the vertebrae [of the spine] are irregular'*, 'some o f the verte- brae are fiat, i.e. they have the condition platyspondyly', 'the femoral capital epiphyses are small', and 'the middle o f the face is flat'. In each o f these findings an abnormal- ity is described which occurs in a specific region (local- *A finding will be denoted in quotes, e.g. 'platyspondyly at birth', in natural language or when referring to the user interface, and as a parenthesised expression, e.g. (platyspondyly (atmonths (0 0))) when referring to the internal representation. Key / - - 7 subspace representing finding (S ... [Ix ly] [ta tb]) subspace of a relevant-finding (contained-within or overlapping-with the finding's space) Figure 3 Finding and its relevant findings as subspaces on findings space ity). The locality specifies a generic region, or a compo- nent, or a subtype, o f the subject. The above findings can be expressed in the f o r m a t (subject) (locality) (attri- bute-values) as follows: 'spine all-vertebrae irregular', 'spine some-vertebrae flat', 'epiphyses femoral-capital small', and 'face middle flat'. In addition, generic terms can describe localities rele- vant to m a n y subjects. The above examples use 'all-ver- tebrae' and 'some-vertebrae'. It is also possible to talk a b o u t 'all-long-bones', 'all-epiphyses', 'all-digits' etc., a n d similarly a b o u t 'some-long-bones', 'some-epiphyses' a n d 'some-digits'. All these can be expressed in terms o f the two generic locality descriptions: localised a n d gener- alised. These two localities are opposing (conflicting) localities. Other generic locality descriptions could be 'upper', 'middle' and 'lower', 'left' a n d 'right', 'laterally' a n d 'medially', 'anteriorly' a n d 'posteriorly' etc. These are n o t opposing localities since they specify disjoint regions. Since the d a t a model allows for the explicit represen- tation o f taxonomic relationships between finding sub- jects, the above findings can therefore be rewritten as 'vertebrae irregular generalised', 'vertebrae flat localised' (or 'platyspondyly localised'), 'femoral-capital-epiphyses small' a n d 'face middle flat'. Locality attributes are spe- cial attributes as illustrated by the following example. The finding 'vertebrae flat irregular localised from-birth' can be converted into the two atomic findings 'vertebrae Knowledge-Based Systems Volume 6 Number 3 September 1993 147 T o w a r d s competent information acquisition interactions: E T K e r a v n o u e t al. flat localised f r o m - b i r t h ' and 'vertebrae irregular loca- lised from-birth', which say t h a t some vertebrae are irre- gular f r o m birth and some are flat f r o m birth*. T h e scope o f a locality attribute is a finding subject if the finding subject describes an a b n o r m a l i t y (e.g. 'platyspondyly'); otherwise, in the case o f ' n o r m a l ' (i.e. anatomical c o m p o - nent) subjects like vertebrae, its scope is a (finding-sub- ject attribute-value) pair, where the attribute value is a non-locality value. This is because it is meaningful to say ' p l a t y s p o n d y l y localised' but it is not meaningful to say 'vertebrae localised'. Suppose that the p r o b l e m findings include 'verteb rae irregular localised' and that the query finding is 'verte- brae irregular generalised'. T h e query finding is estab- lished as false because the localities generalised and loca- lised are opposing; in this respect, these two generic localities are treated as any o t h e r attribute value f r o m the perspective o f the matcher. Suppose, further, that the problem findings include 'long-bones short' and that the q u e r y finding is 'ulnae short' (ulnae being two o f the long bones). T h e query finding is established to be true as follows: 'long-bones short' = 'long-bones all-long-bones sh o rt ' 'ulnae short' = 'long-bones ulnae short' {ulna} c all-long-bones I f the query finding were 'long-bones short' and the problem finding were 'ulnae short' the query finding would be u n k n o w n . T h e relevant findings o f a finding could therefore include findings whose subjects are taxonomically related to the subject o f the finding. T a x o n o m i c relationships between finding subjects, however, are better manipu- lated by the inferencer (see the f o u r t h section). T i m e i n t e r v a l s Findings have temporal aspects which default to 'now'. The t e m p o r a l reasoner which supports the decide-status function is discussed in Reference 11. T h e t e m p o r a l aspects o f findings are expressed qualitatively, e.g. 'at- birth', ' f r o m - b i r t h ' , 'under-yrs', ' = + y r s ' , 'infancy', 'mid- childhood' etc. in the S D D system. These qualitative expressions are translated internally into time intervals. Because o f the u n c e r t a i n t y inherent in some o f the quali- tative t e m p o r a l aspects, e.g. ' f r o m a b o u t a certain time' or 'up to a b o u t a certain time', the time intervals can be 'open'. Since an open time interval essentially delineates a smaller closed interval, f o r the purposes o f this p a p e r it is assumed that there is no time uncertainty and hence time intervals are closed intervals. Recently, Console et al? 2 have r e p o r t e d o n a causal t e m p o r a l f r a m e w o r k allowing for 'variable' time intervals. In general m u c h a t t e n t i o n has been given to t e m p o r a l reasoning by the AI c o m m u n i t y in the recent years, b o t h at the theoretical and applicative levels (e.g. References 13-20). R e l e v a n t s y n o n y m s Th e c o n c e p t o f relevant synonyms is illustrated t h r o u g h an example f r o m the d o m a i n o f skeletal dysplasias. T h e S D D d o m a i n m o d el includes the following relations: syn-subjects (vertebrae vertebral-bodies) synonyms ((platyspondyly) (vertebral-bodies flat)) Let the q u ery finding F be (vertebral-bodies flat). T h e relevant s y n o n y m s o f F are (platyspondyly) and (verte- brae fiat). T h e first s y n o n y m is a direct synonym o b tained fro m the synonyms relation in the d o m a i n model. T h e second s y n o n y m is a direct subject synonym o b t a i n e d by directly substituting a s y n o n y m o u s subject fo r the query- finding subject. Again the s y n o n y m o u s subjects are dir- ectly given in the d o m a i n model t h r o u g h the syn-subjects relation. This example illustrates the two direct routes t h r o u g h which relevant sy n o n y m s m a y be derived. In addition there are two indirect routes. I f the subject p l a t y s p o n d y l y h ad an y s y n o n y m o u s subjects, o r if find- ing (vertebrae flat) h ad an y direct synonyms, t h en m o r e relevant sy n o n y m s would have been obtained. T h u s the first indirect ro u t e is to substitute s y n o n y m o u s subjects for the direct s y n o n y m s o f F an d the second indirect ro u t e is to determine direct sy n o n y m s fo r the direct sub- ject s y n o n y m s o f F. T h e relevant s y n o n y m s o f a q u ery finding F are there- fore defined as follows: relevant-synonyms(F) = direct-syn(F) U direct-subj-syn(F) O indirect-subj-syn( F) U indirect-syn( F) direct-syn(F) = {O I synonyms(Oa ~a) an d directly-matches(F, ~ ) an d time-interval(O) = time-interval(F)} where Xa is the a t e m p o r a l finding o f finding X. direct-subj-syn( F) = { O L syn-subjects( subject( O ) subject (F)) an d attribute-values(O) = attribute-values(F) an d time-interval(O) = time-interval(F)} indirect-subj-syn(F) = U direct-syn(03 i where Oi c direct-subj-syn(F) indirect-syn(F) = U direct-subj-syn(Oi) i where Oi e direct-syn(F) U indirect-subj-syn(F) *But not necessarily the same vertebrae which is a limitation of the representation if this is what is required to be expressed. T h e direct sy n o n y m s o f a finding are o b t a i n e d f r o m the synonyms relation in the d o m a i n model. T h e findings in the instances o f this relation are a t e m p o r a l an d hence the 148 Knowledge-Based Systems Volume 6 Number 3 September 1993 Towards competent information acquisition interactions: E T Keravnou et al. time interval o f the q u e r y finding is removed. T h e result- ing a t e m p o r a l finding is m a t c h e d against the a r g u m e n t s o f the instances o f the synonyms relation. I f a direct m a t c h occurs with one o f the two a r g u m e n t findings, the o t h e r finding a u g m e n t e d with the relevant time interval is a direct s y n o n y m o f the q u e r y finding. An a t e m p o r a l finding directly matches a n o t h e r a t e m p o r a l finding if they have identical subjects a n d the attribute values o f the f o r m e r are subsumed by the a t t r i b u t e values o f the latter. F o r example (metaphyses flared) directly matches against (metaphyses wide irregular), 'flared' and 'wide' being s y n o n y m o u s values (in the c o n t e x t o f the subject 'metaphyses'). F o r efficiency reasons the synonyms rela- tion is implemented as tuples (equivalence classes) o f s y n o n y m o u s findings, indexed by the relevant subjects. Similarly, the syn-subjects relation is i m p l e m e n t e d as tuples o f s y n o n y m o u s subjects. Completeness of relevant-synonyms definition T h e completeness o f the relevant-synonyms definition is p r o v e n below, starting with the following assumptions: • Canonical subject assumption: One subject f r o m each equivalence class o f s y n o n y m o u s subjects is desig- n a t e d as the canonical subject. • Minimality assumption: T h e equivalence classes o f s y n o n y m o u s findings are expressed in a minimal way, i.e. o n l y the canonical subjects are used. • Completeness assumption: T h e knowledge base is complete with respect to all s y n o n y m relations, i.e. the equivalence classes o f s y n o n y m o u s subjects, the minimal equivalence classes o f s y n o n y m o u s findings, and the equivalence classes o f s y n o n y m o u s attrib u t e values are complete. Suppose t h a t the relevant-synonyms definition is incom- plete, m e a n i n g t h a t one o f the relevant s y n o n y m s o f some finding, which is derivable f r o m the knowledge base, is n o t covered by the definition. T h e a b o v e p r o c e d u r a l * definition o f relevant syno- nyms (which facilitates i m p l e m e n t a t i o n ) is 'equivalent't to the following declarative definition: • P is a relevant synonym o f F if syn-subjects( subject( P), subject(F)) a n d attribute-values(P) = attribute-values(F) a n d time-interval(P) = time-interval(F) • P is a relevant synonym o f F if *Procedural because it imposes a specific sequence in the operations involved. t'Equivalent' because all findings returned by the procedural definition are also returned by the declarative definition and any finding returned by the declarative definition which is not explicitly returned by the procedural definition is directly related to one returned by the proce- dural definition through their attribute values. These findings do not need to be returned explicitly since the matcher can dynamically estab- lish whether such a link exists between two findings. synonyms(X, Y) a n d imatches(F,, X) a n d imatches(P., Y) a n d time-interval(P) = time-interval(F) where X an d Y are a t e m p o r a l findings; F,(P,) is the a t e m p o r a l finding o f F(P), a n d imatches(U., V,) if {subject(U,) = subject(V,) o r syn-subjects(subject( U,), subject(V,))} a n d (attribute-values(V,) => attribute-values(U,)) i.e. imatches(Ua, V,) i f V, => U. (V, subsumes U, where s u b s u m p t i o n includes identity). A relevant s y n o n y m is t h erefo re n o t co v ered i f the syn-subjects relation in the first clause is not satisfied, either the synonyms o r o n e o f the imatches relations in the second clause is n o t satisfied; i f an imatches relation is n o t satisfied this m ean s t h a t either the syn- subjects relation is n o t satisfied o r the => relation is n o t satisfied; the => relation is n o t satisfied i f syno- n y m -v al u e relations are n o t satisfied. In each o f the a b o v e cases there is a violation o f the completeness assu m p t i o n (the s y n o n y m o u s subjects are incomplete, o r the minimal s y n o n y m o u s findings are incomplete o r the s y n o n y m o u s attribute values are incomplete). H e n c e the a b o v e definition o f relevant syno- n y m s is complete. Using non-redundant internal representation T h e ability o f a system to u n d e r s t a n d s y n o n y m o u s expressions ( t h r o u g h s y n o n y m o u s findings, s y n o n y m o u s finding subjects a n d s y n o n y m o u s attribute values) is con- sidered b y the d o m a i n experts o f the S D D system as an i m p o r t a n t req u i rem en t because it gives m u c h flexibility to the users o f the system in entering i n f o r m a t i o n dir- ectly. H o w e v e r the question t h a t justifiably arises is w h et h er this r e q u i r e m e n t should be catered f o r by the user interface o r w h e t h e r the system should s u p p o r t an internal r e d u n d a n t representation, requiring the d y n a m - ic c o r r e l a t i o n o f s y n o n y m o u s expressions. I f the first o p t i o n is t ak en t h en a f r o n t - e n d processor w o u l d trans- late every user finding to its internal canonical (standard) representation; o n o u t p u t , findings w o u l d be translated b ack to the expressions used b y the user. T h e advantages o f this o p t i o n are t h a t it reduces the ru n -t i m e processing and, to a large extent, c o m m u n i c a t e s with users in their o w n terms (with the ad d ed possibility o f expressing their findings in the system's terms). T h e benefit o f the second a p p r o a c h is t h a t the infor- m a t i o n is c o m m u n i c a t e d back to the user as entered w i t h o u t incurring an y overheads, whilst with a s t a n d a r d internal rep resen t at i o n it m a y be difficult t o translate some item o f i n f o r m a t i o n b ack to its original form. T h e canonical f o r m f o r some finding, chosen o n the basis o f u n i f o r m i t y an d o t h e r criteria which aim to eliminate Knowledge-Based Systems Volume 6 Number 3 September 1993 149 Towards competent information acquisition interactions: E T Keravnou e t al. ambiguity, m a y not in fact be the preferred form from the user's point o f view. holds(S). I f 0 [--- Y then return true. I f ® ~ {E and • then return false. Step 3: • is unknown. Synonymous attribute values Attribute values m a y be s y n o n y m o u s in the context o f specific subjects. F o r example (metaphyses flared) m a y be taken as s y n o n y m o u s with (metaphyses wide), and (thorax broad) is s y n o n y m o u s with (thorax wide). How- ever metaphyses are n o t described as broad nor is thorax described as flared. Nonetheless the three terms {broad, wide, flared} are considered s y n o n y m o u s in the SDD system. The tradeoff here is between the simplicity o f this representation, which m a y result in some unnecessary processing, and a more complex representation structure which makes explicit the context under which attribute values are synonymous. Findings which are s y n o n y m o u s through their attri- bute values are dynamically determined by the matcher. Negative, positive and status findings The negative finding for a subject at a given time and locality is an atomic finding whose attribute value is the negative value for the subject (see the decide-status d a t a model above). A negative finding expresses a normal situation and thus excludes the relevant positive findings. Examples o f negative findings are (platyspondyly absent) and (hands normal). A positive finding therefore expresses an abnormality. F o r each subject there is a single negative finding for a given time and locality, a most general positive finding and a number o f more speci- fic positive findings. The most general positive finding for an a b n o r m a l subject simply expresses the presence o f the given subject, and for a normal subject it expresses the fact that the given subject is abnormal, e.g. (platyspon- dyly) and (hands abnormal). Some anatomical-component subjects have a status finding which expresses the absence o f the given finding subject (at a particular time and locality), e.g. (humeri absent) or (skull-vault absent-ossification). The status finding is an atomic finding whose attribute value is the status value for the subject. A status finding is a positive finding since it expresses an abnormality. However it is a special positive finding because it excludes all other find- ings for that subject (for the given time and locality). The matcher operation is as follows Let • be u {holds(E)} i where F i e relevant-findings (Q, G r o u p ) and • be holds- (Q). • Step 1: I f ~ [-- • then return true. I f ~ ~ - - ~ then return false. • Step 2: Let S e relevant-synonyms (Q, Group), ® be I f Step 1 fails, Step 2 is tried, and, if that fails as well, the query finding Q is determined as u n k n o w n by the matcher. In order to decide whether a finding matches, or is refuted, against a group o f findings, the matcher uses the following axioms. • A x i o m 1: negative-finding (N) and positive-fnding (P) and subject ( N) = subject(P) and overlapping(locali- ty(N), locality(P)) and overlapping (time-interval(N), time-interval(P)) and holds(P) ~ ~ holds(N). • A x i o m 2: negative-finding(N) and positive-finding(P) and subject( N) = subject( P) a n d overlapping(locali- ty(N), locality(P)) and overlapping(time-interval(N), time-interval(P)) and holds( N) ~ .,~ holds( P). • A x i o m 3: negative-finding(NO and negative-find- ing(N2) and subject(N=)= subject(N2) and includes(lo- cality(NO, locality(N2)) and includes(time-inter- val(NO, time-interval(N2)) and holds( N O--* holds( N2). • A x i o m 4: positive-finding(P) and most-general-posit- ive-finding(M) and subject(P)= subject(M) a n d inclu- des(locality(P), locality(M)) and includes(time-inter- val(P), time-interval(M)) a n d h o l d s ( P ) ~ h o l d s ( M ) . • A x i o m 5: non-status-finding(N) and status-finding(S) and subject(N) = subject(S) and overlapping(local- ity(N), locality(S)) and overlapping(time-interval(N), time-interval(S)) and holds( N ) ~ ,,~ holds(S). • A x i o m 6." non-status-finding(N) and status-finding(S) and subject(N)=subject(S) and overlapping(local- ity(N), locality(S)) and overlapping(time-interval(N), time-interval(S)) and holds( S ) ~ ~ holds( N). • A x i o m 7: status-finding(SO and status-finding(S2) and subject(SO = subject(S2) and includes(locality(SO, locality(S2)) and includes(time-interval(Si), time- interval( Sz) ) and holds(SO--holds(S2). • A x i o m 8: positive-finding(P) and V V e values (P) q F such that positive-finding(F) and match-value(P, F, lO and holds( F ) ~ holds( P). • A x i o m 9: positive-finding(P) and value (P, V) and positive-finding(F) and refute-value(P, F, V) and holds(F) ~ ~ holds(P). • A x i o m 10: positive-finding(P1) and positive-find- ing(P2) and subject( P O = subject( P2) and includes(lo- cality(P2), locality( P O ) and includes(time-inter- val(P2), time-interval(PO) and value(Pb Vl) and value(P2, V2) and identical-or-synonymous( V1, V2)~match-value(P1, P2, VI). • A x i o m 11: positive-finding(PO and positive-find- ing(P2) and subject(PO=subject(P2) and overlapp- ing(locality(P2), locality(Pl)) and overlapping(time- interval(P2), time-interval(Pl)) a n d value(Pl, Vl) and value(P2, V2) and opposite-values(V1, I12) --*refute- value(Pi, P2, VO. U {holds (F0} i where F i e relevant-findings (S, Group), and E be (includes (X, Y) means that X includes Y. subject, local- ity, time-interval and values are selector functions which apply to findings.) 150 Knowledge-Based Systems Volume 6 Number 3 September 1993 Towards competent information acquisition interactions: E T Keravnou et aL INFERENCER The inferencer is invoked for some finding Q queried relative to a group of findings if Q has relevant impli- cants in the domain model. Relevant implicants The relevant implicants of a finding are obtained from the implies relation in the domain model: implies( Fl, F2) i.e. F, =~F2, where F, and F2 are atemporal findings, relevant-implicants( F) = positive-implicants( F) U negative-implicants(F) where positive-implicants(F) = {Pt [ implies(P, C) and matches(Fa, C)} negative-implicants(F) = {Pt I implies(P, C) and m a t c h e s ( ~ F~, C)} Pt is the temporal finding obtained from P and the time interval of F, Fa is the atemporal finding of F and mat- ches(X, Y) is true if the matcher returns 'true' when X is queried against the singleton group of findings consisting of Y. The implies relation collectively covers dependency, definitional, or causality relationships between findings. Generalisations and restrictions The is-a and p a r t - o f domain relations are also used to obtain more relevant implicants according to the con- text-free axioms given below. • A x i o m 1: sholds(P, S) and is-a(s, S ) ~ s h o l d s ( P , s). sholds(P, S) means that proposition (i.e. finding) P holds for subject S, e.g. (long-bones short)~(ulnae short). • A x i o m 2: V s such that is-a(s, S) sholds(P, s ) ~ sholds(P, S), e.g. {(ulnae short) and (radii short)} ~long-bones short). • A x i o m 3: V s such that part-of(s, S) sholds(P, s)--* sholds(P, S), e.g. {(cervical-spine normal) and (thor- aco-lumbar-spine normal)}--* (spine normal). • A x i o m 4: ~ s such that part-otis, S) sholds(abnor- real(s), s) ~ sholds(abnormal(S), S), e.g. (mid-face abnormal)--* (face abnormal). • ? A x i o m 5: sholds(P, S) and part-of(s, S) ~ sholds (P, s). The fourth axiom deals with the most general positive findings. This is because if a part of some subject is abnormal in some specific way, it does not necessarily mean that the whole subject is similarly abnormal, e.g. the trunk and the limbs may be loosely described as parts of one's stature; if the trunk is short it is not necessarily the case that the stature is also short (although the opposite would be unusual). Similarly, Axiom 5 is not really used because it is not truly context-free (except when the property is normality). Axiom 5 may apply under some interpretations only. For example if the spine is abnormal it does not mean that all the vertebrae on each of the regions of the spine (cervical-spine, lum- bar-spine, thoracic-spine) are abnormal. On the other hand, if the spine is irregularly ossified it can be inferred that all the vertebrae are irregularly ossified. In other words if we are dealing with a general abnormality (e.g. 'spine abnormal') it is not possible to infer that all the subcomponents are abnormal. With some specific abnor- malities, however (e.g. 'spine irregularly-ossified'), it can be inferred that the given abnormality applies to all the subcomponents. The scope of this axiom therefore depends on domain specific knowledge. Suppose that the query finding is (ulnae severely- short) and that the given finding is (short-limbed-dwar- fism). The latter finding is synonymous with (long-bones severely-short) which is a generalisation of the query finding. Hence the inferencer, by dynamically generating implications based on taxonomic relations (is-a and part- of), enables 'linking' between the query finding and a given finding which is synonymous with generalisations/ restrictions of the query finding. This example illustrates the co-operation between the matcher and inferencer. Inferencer operation The inferencer operation is simple to express since the inferencer calls Decide-Status recursively to decide the truth status of the antecedents of the relevant implicants of the query finding. I f a positive implicant is entailed to hold then the query finding holds (the inferencer returns true) and if a negative implicant is entailed to hold the query finding does not hold (the inferencer returns false). Otherwise the answer to the query is decided to be unknown by the inferencer. The inferencer axioms are given below. Let F be the assertions about which findings hold for the given problem • A x i o m 1: L e t P c positive-implicants(F). If F ]-- holds(P) then F ~ holds(F). • A x i o m 2: Let N ~ negative-implicants(F). If F [-- holds(N) then F ~-- ,-~ holds(F). If the problem theory is consistent, i.e. both the domain model and problem findings are consistent, then the two axioms are never simultaneously applicable for the same query finding. The chain of implications which is dyna- mically constructed through the recursive calls between Decide-Status and the inferencer is recorded so that a loop can be immediately detected and the relevant infer- ence chain terminated. The matcher and inferencer strategies both compete and co-operate. If the matcher cannot return a firm answer, the inferencer is invoked; however it is the matcher that terminates a successful inference chain generated by the inferencer by determining the relevant terminal antecedent finding as true. Decide-Status is implemented in Franz LISP on a Sun Knowledge-Based Systems Volume 6 Number 3 September 1993 151 Towards competent information acquisition interactions: E T Keravnou e t al. 3/60 running Unix (declarative definitions for the matcher, inferencer and Decide-Status functions are given in Reference 6). The algorithms for the Decide- Status and the inferencer, illustrating the chain recursion between them are as follows. The definition below assumes that the a r g u m e n t Find- ing is a simple finding; the generalisation to include com- p o u n d findings is given in Reference 6. Finding is queried against a group o f findings, G r o u p , which are assumed to hold. Impl-Chain is a chain o f implications, initially empty, which is constructed dynamically t h r o u g h the recursive calls between Decide-Status a n d the inferencer. Decide-Status (Finding, Group, Impl-Chain): let Truth-Value = Matcher(Finding, G r o u p ) if firm(Truth-Value) then return(Truth-Value) else let Rel-Impl = relevant-implicants(Finding) New-Chain = cons(Finding, Impl-Chain) return(Inferencer(Rel-Impl, G r o u p , New-Chain)) A firm truth value means 'true' or 'false', but not ' u n k n o w n ' . The relevant implicants o f Finding are those findings which could be used to infer Finding. Inferencer (Rel-Impl, Group, Impl-Chain): if null(Rel-Impl) return ('unknown') else if member(first(Rel-Impl), Impl-Chain) then return (Inferencer(rest(Rel-Impl), Group, Impl-Chain)) else if Decide-Status(first(Rel-Impl), Group, Impl- Chain) = 'true' then if positive (first (Rel-Impl)) then return ('true') else return ('false') else return(Inferencer(rest(Rel-Impl), Group, Impl-Chain)) The inferencer determines whether any o f the relevant implicants can be determined from the group o f findings. No formal analysis o f the complexity o f the decide-status function has been done. However it is a computationally intensive operation; much o f this processing is attributed to the inferencer operation since the inferencer m a y insti- gate a number o f alternative inference chains which m a y all prove to be unfruitful. The matcher is always tried before the inferencer. As Decide-Status is invoked in m a n y different reasoning contexts in the expert system (see below), its optimisation would be beneficial. Finally Decide-Status is a m o d u l a r piece o f software which faci- litates extensions. New strategies or axioms can be easily added. Such extensions m a y be accompanied by exten- sions to the d a t a model. R E A L I S I N G I N F O R M A T I O N - A C Q U I S I T I O N G O A L S T H R O U G H D E C I D E - S T A T U S Decide-Status operates on an explicit d a t a model. Its operation is conceptually simple ('given a d o m a i n model and a set o f findings which are assumed to hold, does a particular finding follow or not?'), but powerful enough to either directly implement or support the implemen- tation o f the information-acquisition interaction objec- tives discussed in the first section. The d a t a model makes explicit the structure o f the d o m a i n o f findings. Findings are n o t indivisible, atomic entities, i.e. strings o f characters; they have an internal meaningful structure, enabling the decomposition o f individual findings and the correlation between find- ings*. Mixed-initiative interaction The dependencies between findings (synonyms, generali- sations, restrictions, implications) are explicit; Decide- Status, through these dependencies, can correlate expres- sions which mean literally or essentially the same thing, and can detect r e d u n d a n c y and inconsistency in the user- volunteered information. As a result o f this, the user has the option to volunteer information directly rather than through a sequence o f menus. It is still possible that some information volunteered by the user will not be under- stood by the system in which case the system should inform the user. As explained in the first section, the overall sequence o f questions raised by the system can be viewed at two levels o f abstraction. The top level gives the sequence o f strategy applications generated by the workings o f the problem solver and the b o t t o m level gives the question subsequences corresponding to the different strategy applications 2. Thus some strategy application determines that a given set o f information items should be elicited from the user. However it is the d a t a model and the associated relevant reasoning that organises the required information into an intelligible, coherent sequence o f questions. Questions for the same subject are grouped together and preceded by the relevant general questions (most general positive finding and, if applicable, status finding). F o r example if the system wants to ask whether the skull vault is small it will first try to establish whether the skull vault is abnormal and whether it is ossified (i.e. present) in this order. The context for raising specific questions in a meaningful way is established by asking general questions first. I f this were n o t so then an answer could be ambiguous; for example if it is n o t known whether the patient exhibits platyspondyly and the system raises the question 'platyspondyly severe?' a nega- tive answer would either mean that 'there is no platy- spondyly' or that 'there is platyspondyly but it is mild'. In addition to general questions, there m a y be other contextual questions, either associated with the subject or specific findings o f the subject, which need to be raised prior to asking some particular question. At present the d a t a model does n o t include such contextual associa- tions, but their inclusion is straightforward. The ordering o f the groups o f questions for different subjects is n o t necessarily arbitrary. Again contextual associations between different subjects, if any, can indi- *In m a n y s y s t e m s findings h a v e n o s t r u c t u r e a n d internally t h e y c a n be represented b y a n o n y m o u s codes, w h i c h o f c o u r s e speeds u p t h e p r o - cessing; this is o f t e n t h e case w i t h m e n u - d r i v e n interfaces where strings are displayed o n t h e screen b u t u s e r replies a r e t r a n s l a t e d i n t o internal, n o n - d e c o m p o s a b l e codes. 152 Knowledge-Based Systems Volume 6 Number 3 September 1993 Towards competent information acquisition interactions: E T Keravnou e t al. cate a p a r t i c u l a r sequence. In the S D D system the spatial p r o x i m i t y o f a n a t o m i c a l c o m p o n e n t s is used to o r d e r questions for adjoining a n a t o m i c a l c o m p o n e n t s consecu- tively, thus saving the user f r o m having to keep swapping between different X - r a y films; hence questions o n the ribs a n d the thoracic spine will be g r o u p e d together. A user reply to a system question is immediately t a k e n into consideration, which is reflected in subsequent questions being screened out. T h e decide-status funct i o n is used p r i o r to asking every question to see w h e t h e r the answer to the question can already be determined. In some expert systems the questions raised by the system are m o r e general t h a n the c u r r e n t l y p u r s u e d goal requires, where the generality is n o t for establishing a meaningful c o n t e x t for raising specific questions b u t fo r acquiring all relevant i n f o r m a t i o n simultaneously (e.g. instead o f asking 'does p a r a m e t e r x o f object y have value z?' it will ask ' w h a t is the value o f p a r a m e t e r x fo r object y?'). This is absolutely necessary if the system does n o t allow the user to v o l u n t e e r i n f o r m a t i o n . In the pro - posed f r a m e w o r k this is n o t so. I m m e d i a t e l y after a questioning r o u n d by the system, the user is given the o p p o r t u n i t y to v o l u n t e e r i n f o r m a t i o n . T h e main objective o f expert systems t e c h n o l o g y is to disseminate (and possibly augment) h u m a n expertise 21. Thus, the m a j o r i t y o f users o f an expert system, a l t h o u g h knowledgeable in the given d o m a i n , will n o t be experts themselves. In this respect, it is possible that i n f o r m a t i o n volunteered by the user will be e r r o n e o u s a n d t h a t the user will n o t u n d e r s t a n d a question raised by the system. Requiring the system to t r a p e r r o n e o u s user input and to give a p p r o p r i a t e guidance to the user when respond i n g to system questions 7,22,23 puts a n o t h e r level o f complex i t y on the system i n f o r m a t i o n - a c q u i s i t i o n initiatives. E r r o n - eous input which conflicts with o t h e r i n f o r m a t i o n can be trapped. Clarifying questions (see below) d o n o t aim to trap e r r o n e o u s i n p u t as such, b u t r a t h e r to t r a p c o m m o n inaccuracies in the interpretations o f actual situations. Serious inaccuracies result in e r r o n e o u s input. In the S D D system an on-line d a t a b a s e o f X - r a y images will be used in the validation o f user input and in provid i n g guidance to the user in answering a system question. System questions m a y be illustrated by displaying a rele- vant image and, similarly, user input m a y be validated b y displaying images which illustrate w h a t the user has said. I f the displayed image does n o t m a t c h the actual case image then it is likely t h a t the user has misinterpreted the particular situation*. Flexibility in user-volunteered information T h e instantiation o f the d a t a m o d e l for a specific d o m a i n aims to c a p t u r e the entire v o c a b u l a r y regarding the expression o f findings, in o r d e r to give the user the required flexibility. In a d d i t i o n the user m a y wish to revoke i n f o r m a t i o n . Solutions to t r u t h m a i n t e n a n c e are n o t cheap since assumptions a n d inference dependencies m u s t be expli- citly represented. On the basis o f w h a t i n f o r m a t i o n is assumed to h o l d at some stage in the consultation, par- tial solutions are g en erat ed a n d pursued. These alterna- tive solutions are evaluated against each other. I f some i n f o r m a t i o n is r e v o k e d t h en its rev o cat i o n will f a v o u r the solutions conflicting with this i n f o r m a t i o n a n d will reduce the promise o f those solutions s u p p o r t e d by the r e v o k e d i n f o r m a t i o n . I f the r e v o k e d i n f o r m a t i o n is criti- cal f o r a p art i cu l ar solution t h en t h a t solution m a y be excluded. Decide-Status does n o t directly enable infor- m a t i o n r e v o c a t i o n b u t it su p p o rt s it to some extent. In a diagnostic c o n t e x t the promises o f the various co m peting hypotheses are dynamically r e c o m p u t e d d u ri n g every diagnostic cycle, always using the i n f o r m a t i o n which is believed to h o l d at t h a t stage (Decide-Status is called to determine w h et h er hypotheses' expectations are satisfied o r refuted a n d w h et h er case findings are a c c o u n t e d by, or are in conflict with, hypotheses). H e n c e revocations will be reflected in the new 'p ro m i se' f o r an active hypothesis. In practice, however, because Decide-Status is a c o m p u - tationally expensive process, some o f its derivations are stored fo r fu t u re use. I f a r e v o c a t i o n results in the refu- t at i o n o f a previous derivation t h en inconsistencies will occur. A n u m b e r o f d o m a i n - i n d e p e n d e n t t ru t h -m ainten- ance f r a m e w o r k s have been p r o p o s e d in the literature 26-28 which present theoretically interesting a p p r o a c h e s b u t which m a y fail in practice because o f the associated c o m p u t a t i o n a l o r o t h e r overheads. In practice, for rea- sons o f viability, a t r u t h - m a i n t e n a n c e f r a m e w o r k will be tailored to the specifics o f a p art i cu l ar p r o b l e m solver, o r m a y even be 'h ard -w i red ' in the workings o f the p r o b l e m solver. In the S D D system the case findings are divided into ' h a r d ' findings a n d 'soft' findings. T h e distinction between h a r d an d soft findings is m a d e in m a n y medical expert systems, often to distinguish between l a b o r a t o r y findings an d co-incidental (circumstantial) findings. S D D ' s m ean i n g an d usage o f these terms m a y be differ- ent f r o m o t h e r systems, a l t h o u g h the c o n c e p t is essen- tially the same: some findings are m o r e i m p o r t a n t t h a n others in a diagnostic contextt. H a r d findings c o r r e s p o n d to diagnostically significant observations o f a b n o r m a - lites. Soft findings describe abnormalities which m a y be a t t r i b u t e d to n a t u r a l causes. T o be acceptable, any hypothesis m u st a c c o u n t fo r a reasonable p r o p o r t i o n o f h a r d findings, while soft findings m a y conceivably be ignored. Obviously, revoking a h a r d finding will affect the solution space b u t the rev o cat i o n o f a soft finding will n o t necessarily affect the solution space. Similarly, some findings are r a t h e r critical fo r certain hypotheses, i.e. the presence o f some finding results in concluding the hypothesis, o r the absence o f some finding results in refuting the hypothesis. Again a n y revocations directly o r indirectly resulting in refuting o r establishing the presence o f such findings m u st be t ak en into consider- at i o n when some i n f o r m a t i o n is revoked. Decide-Status can be used to determine the t r u t h status o f such critical findings, which h av e resulted in considerably enhancing o r reducing the promise o f p o t en t i al alternative solu- tions; the reversal o f potentially irreversible decisions o n the p a r t o f the p r o b l e m solver m a y thus be possible. *The use of images in computer-aided diagnosis in radiology is dis- cussed in References 24 and 25. *Again Decide-Status is used to decide whether a case finding is soft or hard; this reasoning context is outside the scope of the paper. Knowledge-Based Systems Volume 6 Number 3 September 1993 153 Towards competent information acquisition interactions: E T Keravnou e t al. Similarly, conflicts in the user-volunteered infor- mation can be detected by Decide-Status. I f a new obser- vation is determined as false by the previous obser- vations the p r o o f tree constructed by Decide-Status can also be used to explain the conflict (see Reference 10). Decide-Status can be used to determine dependencies between the items o f information volunteered by the user. I f some observation is deducible from other obser- vations then it is redundant. F o r example if the user says that the stature is abnormal and subsequently says that the stature is short, the first observation is eliminated. In a diagnostic context pieces o f evidence should be inde- pendent. Equally, hypotheses' expectations must be non- redundant; again Decide-Status is used to eliminate poss- ible redundancies. In addition the d a t a model (but not Decide-Status per se) is used to temporally screen a hypothesis profile, thus eliminating expectations whose (relative) temporal aspects refer to the future with respect to the patient (see Reference 11). System competence in querying user The system maintains two lists o f findings: Observations and U n k n o w n Findings. The former includes the find- ings which are currently assumed to hold; some o f these findings are directly volunteered by the user whilst others are elicited from the user through questions. Revoking a user observation results in deleting the particular obser- vation from Observations. The U n k n o w n Findings list includes all findings which have been specified by the user, in response to a system question, as unknown. The previously u n k n o w n answer to a question m a y subse- quently become known to the user, who will have the o p p o r t u n i t y to volunteer this information. W h e n new information is volunteered by the user, the U n k n o w n Findings must be revised, since the new information m a y result in determining a firm truth status for a previously u n k n o w n finding. Decide-Status is thus invoked to deter- mine the truth status o f U n k n o w n Findings against Observations. Prior to asking a question, the system first checks whether the truth status o f the query finding can be determined from the Observations list. I f Decide-Status (query-finding, Observations, nil) returns ' u n k n o w n ' , the system checks whether the user has already said that the given item o f information is u n k n o w n , i.e. matcher (query-finding, Unknown-Findings) returns a firm answer (true or false) or if the query finding is more specific t h a n one o f the findings in the U n k n o w n Find- ings list; specificity here is determined using is-a and part- o f relations. F o r example suppose t h a t the user was asked the question 'skull abnormal?' to which the reply ' u n k n o w n ' was given. ( U n k n o w n replies are fairly c o m m o n in the d o m a i n o f SDD since Often the X-ray images for the patient do n o t give a complete skeletal survey and few clinical findings are usually available). I f subsequently the system wants to find whether the skull is small it will n o t ask the question because it knows that the status o f the skull is unknown. Similarly, if the user does not know if the spine is abnormal, the system should n o t ask whether the cervical spine is abnormal. However if the user does not know a b o u t the status o f the cervical spine it is still intelligible on the part o f the system to ask a b o u t the status o f the spine. Since specific questions are always preceded by the relevant general questions (see above), if the U n k n o w n Findings list includes 'platyspondyly severe' then it can be inferred that the presence o f platyspondyly has already been established; otherwise if platyspondyly were absent, 'platyspondyly severe' would be false, not unknown. Finally the d a t a model associates prompting and clari- fying questions, and temporal information (when-to-ask relation) with specific findings. These associations contri- bute much to the naturalness o f the information-acqui- sition interaction. Prompting and clarifying questions are raised in response to information volunteered by the user whilst temporal information is used to stop the system from asking unintelligent questions such as 'is the gait o f y o u r 2 m o n t h old baby waddling?'. When-to-ask relations are used in the context o f temporal screening. The three relations are as follows: • prompting~-query(X, Y,) e.g. prompting-query ((platy- spondyly) (platyspondyly throughout)), • clarifying-query(XaY~) e.g. clarifying-query ((face fiat) (mid-face hypoplastic)), • when-to-ask(W~ T) e.g. when-to-ask ((kyphoscoliosis severe) ( = + yrs 1)). where Xa, Ya and Wa are atemporal findings and T is a relative time point. Suppose that the user volunteers finding F. If mat- cher(X~, [Fa]) returns true, i.e. if Xa holds given Fa, then the temporal version o f the corresponding atemporal prompting query will be generated. Thus if the user volunteers 'platyspondyly' then the question 'platyspon- dyly throughout?' is prompted. I f the user volunteers 'platyspondyly mild' again the same question will be prompted. However if the user volunteers 'platyspondyly t h r o u g h o u t ' no prompting will be given; the prompting query is generated but it is immediately quashed because Decide-Status indicates that it is a r e d u n d a n t question. Clarifying queries are dealt with in the same way. Thus if the user volunteers 'face flat', the clarifying question 'do y o u mean mid-face hypoplastic?' will be raised. A future SDD extension is for the system to infer some prompting and clarifying questions. Regarding when-to-ask associations, suppose that Fa is an atemporal feature o f some hypothesis. I f matcher (F~, [Wa]) returns true, then, if T refers to a future point in time relative to the patient, the particular expectation o f the hypothesis is screened out. Thus if the patient is neonatal then the expectation o f 'severe kyphoscoliosis' for any hypothesis will be screened out. However the more general expectation o f 'kyphoscoliosis' will n o t be screened out, since mild forms o f this abnormality are observable from birth. The d a t a model and the decide-status reasoner can therefore directly support the realisation o f most o f the information-acquisition interactions objectives men- tioned in the first section. 154 Knowledge-Based Systems Volume 6 Number 3 September 1993 Towards competent information acquisition interactions: E T Keravnou e t al. C O N C L U S I O N S T h e n e e d f o r c o m p e t e n t c o n v e r s a t i o n a l s t r u c t u r e s b e t w e e n e x p e r t s y s t e m s a n d their users w a s identified e a r l y in t h e life o f e x p e r t s y s t e m s t e c h n o l o g y ; o v e r t h e y e a r s , as t h e r o l e o f e x p e r t s y s t e m s as intelligent a d v i s e r s h a s a c q u i r e d m o r e p r o m i n e n c e , so h a s t h e n e e d f o r a d e q u a t e i n t e r a c t i o n w i t h users. T h e c o n v e r s a t i o n a l s t r u c t u r e e n c o m p a s s e s t h e e n t i r e i n t e r a c t i o n b e t w e e n t h e s y s t e m a n d t h e user. T h i s i n c l u d e s t h e q u e s t i o n s r a i s e d b y t h e s y s t e m , b o t h i n d i v i d u a l l y as well as a n o r d e r e d s e q u e n c e , the i n s t r u c t i o n s a n d g u i d a n c e given t o t h e u s e r b y t h e s y s t e m f o r p e r f o r m i n g a c t i o n s a n d a n s w e r i n g q u e s t i o n s , t h e s y s t e m ' s e x p l a n a t i o n s r e g a r d i n g its reas- o n i n g a n d / o r s u g g e s t i o n s , a n d t h e k i n d s o f initiatives a n d c h o i c e s w h i c h the u s e r is a l l o w e d t o t a k e , e.g. v o l u n t e e r d a t a , v o l u n t e e r s u g g e s t i o n s , a n d r e t r a c t d a t a o r sugges- tions. T h e issue o f a d e q u a t e e x p l a n a t i o n s h a s a t t r a c t e d m o r e a t t e n t i o n t h a n t h e o t h e r a s p e c t s o f a c o n v e r s a t i o n a l s t r u c t u r e . T h e s e o t h e r aspects, t h e i n f o r m a t i o n - a c q u i - sition ones, a r e e q u a l l y i m p o r t a n t . T h e f o c u s r e g a r d i n g i n f o r m a t i o n - a c q u i s i t i o n i n t e r a c t i o n a s p e c t s is o n intelli- g e n t f r o n t e n d s f o r m a k i n g a piece o f s o f t w a r e o r d a t a - b a s e m o r e u s a b l e r a t h e r t h a n in t h e c o n t e x t o f e x p e r t c o n s u l t a n t s y s t e m s p e r s e 7. H o w e v e r a t t e n t i o n is s h i f t i n g in this d i r e c t i o n 29. T o be able t o c o n d u c t a n intelligent i n f o r m a t i o n - a c q u i s i t i o n i n t e r a c t i o n , b o t h in t h e sense o f a s k i n g intelli- gent, c o h e r e n t q u e s t i o n s , a n d in t h e sense o f u n d e r s t a n d - ing t h e u s e r ' s i n f o r m a t i o n v o l u n t e e r i n g , a n e x p e r t s y s t e m m u s t h a v e c o m m o n sense in a limited w a y specific t o t h e n e e d s o f t h e p a r t i c u l a r a p p l i c a t i o n . T h i s o b j e c t i v e is m u c h easier t o a c h i e v e t h a n t h e c o n s i d e r a b l y m o r e a m b i - t i o u s o b j e c t i v e o f c a p t u r i n g t h e e n t i r e b o d y o f w o r l d k n o w l e d g e , as in t h e C Y C p r o j e c t 3°. T h e a r g u m e n t p r e s e n t e d in this p a p e r is t h a t a n e x p e r t s y s t e m c a n c o m m u n i c a t e m o r e intelligently i f it h a s a d e e p e r u n d e r s t a n d i n g o f t h e d a t a f o r s o m e p r o b l e m case, i.e. i f it is c a p a b l e o f h a n d l i n g these d a t a intelligently; this is t h e limit o f t h e r e q u i r e d c o m m o n sense. T h e p a p e r is n o t a b o u t intelligent d a t a b a s e s o r intelligent f r o n t ends. T h e d a t a m o d e l is n o t a d a t a b a s e s c h e m a , b u t r a t h e r a s c h e m a f o r c a p t u r i n g t h e k n o w l e d g e a b o u t t h e d a t a in t h e d o m a i n . T h e r e l a t i o n s i n c l u d e d in t h e d a t a m o d e l are l a r g e l y c o m m o n - s e n s e r e l a t i o n s s u c h as p a r t - o f , is-a, trig- gers, a n d i m p l i c a t i o n s , as well as p r o m p t s , c l a r i f y i n g q u e s t i o n s a n d w h e n - t o - a s k . T h e a x i o m s u s e d b y t h e infer- e n c e r a n d t h e m a t c h e r a r e c o m m o n - s e n s e a x i o m s , w h i c h in f a c t e n h a n c e s t h e g e n e r a l i t y o f t h e d a t a m o d e l . T h e p a p e r d o e s n o t p r o v i d e a c o m p l e t e g e n e r i c s o l u - t i o n t o t h e p r o b l e m o f intelligent i n f o r m a t i o n - a c q u i s i t i o n i n t e r a c t i o n s f o r e x p e r t s y s t e m s , b u t d e s c r i b e s a k e r n e l w h i c h p o i n t s t o w a r d s a c o m p l e t e s o l u t i o n . T h e s t a r t i n g p o i n t is t o m a k e explicit t h e k n o w l e d g e a b o u t t h e d o m a i n d a t a t h r o u g h a d a t a m o d e l ; t h e d a t a m o d e l is a t t h e k n o w l e d g e m e t a l e v e l , the i n s t a n t i a t i o n o f it is a t t h e k n o w l e d g e o b j e c t level a n d t h e p r o b l e m d a t a a r e a t t h e f a c t u a l level. A C K N O W L E D G E M E N T S W e a r e g r a t e f u l t o t h e L e v e r h u l m e T r u s t f o r their s u p - p o r t o f t h e S D D p r o j e c t , t o t h e referees o f t h e p a p e r f o r their u s e f u l c o m m e n t s o n a n earlier d r a f t o f t h e p a p e r , a n d t o S o p h i a P r e v e z a n o u f o r h e r h e l p in c l a r i f y i n g m a n y p o i n t s . R E F E R E N C E S 1 Gilbert, N 'Explanation and Dialogue' Knowledge Engineering Review Vol 4 No 3 (1989) pp 235-247 2 Keravnou, E T and Johnson, L Competent Expert Systems: A Case Study in Fault Diagnosis Chapman and Hall (1986) 3 Keravnou, E T, Dams, F, Washbrook, J, Dawood, R, Hall, C and Shaw, D 'Background Knowledge in Diagnosis' Artificial Intelligence in Medicine Vol 4 No 4 (1992) pp 1-17 4 Keravnou, E T, Washbrook, J, Dawood, R M, Hall, C M and Shaw, D 'A Model-Based Diagnostic Expert System for Skeletal Dysplasias' Proc. A I M E "89 Springer-Verlag, Germany (1989) pp 47-56 5 Stenton, S P 'Dialogue Management for Co-operative Know- ledge-Based Systems' Knowledge Engineering Review Vol 2 No 2 (1987) pp 99-122 6 Keravnou, E T, Washbrook, J and Dams, F 'Explicit Data- Modelling in Second-Generation Diagnostic Expert Systems' Proc. l l th International Conference on Expert Systems and their Applications- Vol 2 Avignon, France (1991) pp 185-198 7 Kok, A J 'A Review and Synthesis of User Modelling in Intelli- gent Systems' Knowledge Engineering Review Vol 6 No 1 (1991) pp 21-47 8 Keravnou, E T and Washbrook, J 'Deep and Shallow Models in Medical Expert Systems' Artificial Intelligence in Medicine Vol 1 No 1 (1989) pp 1-28 9 Southwick, R W 'Explaining Reasoning: An Overview of Expla- nation in Knowledge-Based Systems' Knowledge Engineering Review Vol 6 No 1 (1991) pp 1-19 10 Keravnou, E T and Johnson, L 'Intelligent Handling of Data by Integration of Commonsense Reasoning' Knowledge-Based Systems Vol 1 No 1 (1987)pp 32-42 11 Keravnou, E T and Washbrook, J 'A Temporal Reasoning Framework used in the Diagnosis of Skeletal Dysplasias' Artifi- ciallntelligence in Medicine Vol 2 No 5 (1990) pp 239-265 12 Console, L, Janin Rivolin, A and Torasso, P 'Fuzzy Temporal Reasoning on Causal Models' Int. J. Intelligent Systems Vol 6 No 2 (1991) 13 Allen, J F 'Towards a General Theory of Action and Time' Artificial Intelligence Vol 23 (1984) pp 123-154 14 Dean, T L and McDermott, D V 'Temporal Data Base Manage- ment' Artificial Intelligence Vol 32 (1987) pp 1-55 15 Kahn, M G 'Model-Based Interpretation of Time-Ordered Medi- cal Data' PhD Dissertation Medical Information Sciences, University of California, USA (1988) 16 Keravnou, E T (Ed.) 'Medical Temporal Reasoning' Artificial Intelligence in Medicine Vol 3 No 6 (1991) (special issue) 17 Kowalski, R and Sergot, M 'A Logic-Based Calculus of Events' New Generation Computing Vol 4 (1983) pp 67-95 18 Long, D 'A Review of Temporal Logics' Knowledge Engineering Review Vol 4 No 2 (1989) pp 141-162 19 Rosser, B, Washbrook, J, Campbell, J, Keravnou, E T and Long, D 'A Framework for Time Dependent Reasoning Systems' Pro- duct Pl11-1 Esprit Project P2409 Equator (1989) 20 Shoham, Y 'Temporal Logics in AI: Semantical and Ontological Considerations' Artificial Intelligence Vol 33 (1987) pp 89-104 21 Slatter, P E 'Cognitive Emulation in Expert System Design' Knowledge Engineering Review Vol 2 No 1 (1987) pp 27-41 22 Wolstenholme, D 'Saying "I don't know" and Conditional Answers' in Moralee, D S (Ed.) Research and Development in Expert Systems I V Cambridge University Press (1987) pp 115- 125 23 Wolstenholme, D 'External Data in Logic-Based Advice Systems' PhD Thesis Dep. Computing, Imperial College London, UK (1990) 24 Mutalik, P G, Fisher, P R, Weltin, G and Swett, H A 'Expert System Advice as a By-product of Image Acquisition and Reporting: Obstacles to Overcome' Proc. C A R '91 Springer- Verlag (1991) pp 315-320 25 Swett, H A 'Computer-Aided Diagnosis in Radiology' Proc CAR" 91 Springer-Verlag (1991) pp 738-743 Knowledge-Based Systems Volume 6 Number 3 September 1993 155 Towards competent information acquisition interactions: E T Keravnou e t al. 26 de Kleer, J ' A n Assumption-Based T M S ' Artificial Intelligence Vol 28 (1986) pp 127-224 27 Doyle, J A ' A Truth-Maintenance System' Artificial Intelligence Vol 12 (1979) pp 231 272 28 Inoue, K 'Pruning Search Trees in Assumption-Based Reason- ing' Proc. 8th International Workshop on Expert Systems and their Applications (1988) pp 133-151 29 Keravnou, E T and Washbrook, J ' W h a t is a Deep Expert System? An Analysis of the Architectural Requirements of Second-Generation Expert Systems' Knowledge Engineering Review Vol 4 No 3 (1989) pp 205 233 30 Lenat, D B, Ramanthan, V G, Pittman, K, Pratt, D and She- pherd, M 'CYC: Towards Programs with Common Sense' C A C M Vol 33 No 8 (1990) pp 30-49 B I B L I O G R A P H Y Neal, I M 'First Generation Expert Systems: A Review of Knowledge Acquisition Methodologies' Knowledge Engineering Review Vol 3 No 2 (1988) pp 105-145 Wilson, M, Duce, D and Simpson, D 'Life Cycles in Software and Knowledge Engineering: A Comparative Review' Knowledge Engineer- ing Review Vol 4 No 3 (1989) pp 189 204 156 Knowledge-Based Systems Volume 6 Number 3 September 1993