An Expert System Using Nonmonotonic Techniques

for Benets Inquiry in the Insurance Industry

Leora Morgenstern Moninder Singh
IBM T.J. Watson Research Center Department of Computer & Information Science
30 Saw Mill River Road University of Pennsylvania
Hawthorne, NY 10532 Philadelphia, PA 19104-6389
[email protected] [email protected]
Abstract 2 Existing Technology
Benets inquiry has traditionally been performed by cus-
This paper describes BenInq, an expert system tomer service representatives (CSRs) who rely on basic
for benets inquiry in the insurance industry. charts, detailed manuals, and implicit knowledge gained
BenInq is designed to be used by both customer through common sense and job experience. As insurance
service representatives, who answer customers' products have proliferated, there have been attempts to
questions in real time, and policy modiers, harness computing power to help with the benets in-
who update insurance products. The main quiry task. Often, such eorts are limited to placing
challenges were the design of the KR structure charts and manuals on-line. Other eorts include text-
| in particular, representing a huge knowledge based and code-based systems, discussed below.
base at varying levels of granularity, and the de- Text-based Systems: Text-based systems allow for
velopment of the reasoning methods | in par- some search and indexing of subject areas. The medical
ticular, developing a method that determines insurance company for which we consulted had a rudi-
which insurance regulations apply to an insur- mentary text-based system. Information is divided into
ance service. This required nonmonotonic rea- chunks or subject areas. A piece of text is associated
soning, which we modeled using an extension of with each subject area. Subject areas might include pre-
inheritance methods. BenInq represents one of ventive care, immunizations, and maternity. The text
the few large scale industrial applications that associated with preventive care lists the dierent types
explicitly uses formal nonmonotonic reasoning of preventive care available, such as routine physicals
techniques. and standard immunizations, as well as coverage rates
and allowed frequency of services. Another screen may
deal with one of the topics described in one screen; e.g.,
there may be a screen devoted to immunizations, a topic
1 Introduction described in the preventive care screen. The CSR uses
this system by pulling down a menu of topics, clicking on
Benets inquiry, the process of querying an insurance a topic, and reading the information on the associated
company to determine one's benets, is becoming in- screen which comes up.
creasingly complex. Whereas years ago, a medical in- The advantages to this system over on-line manuals
surance company would oer only a small number of are rst, the system allows rudimentary search, and sec-
products, today companies have thousands of dierent ond, the information is partly organized in a modular
insurance products, each of which contains a myriad of fashion. But there is no reasoning; the system has merely
services and regulations. Furthermore, these products pruned the amount of information to be read. Second,
are always changing. The amount of information is di- the system does not make explicit the many intercon-
cult to keep up with; thus, the need for an expert system. nections between subject areas. For example, nothing in
The expert system described in this paper, BenInq, the system indicates a connection between the screens
was created for the application of benets inquiry in the on immunization and preventive care. The CSR must
medical insurance domain. The paper is structured as reason, e.g., that the schedule rate for preventive care
follows. We rst describe the existing technology for ben- most likely applies to routine immunizations. The lack
ets inquiry. We next list the desiderata for our expert of connection between related screens makes updating
system. Subsequently, we discuss the knowledge repre- the system dicult. If both the preventive care and im-
sentation and reasoning issues, focusing on the nonmo- munization screens do have schedule rate information
notonic techniques used. We then present BenInq, eval- and this schedule rate changes, the individual modifying
uate its usefulness, and discuss how the techniques can the system must make changes on both screens. Third,
be generalized. there are a small number of screens relative to the num-
ber of types of questions a CSR may have to answer. 3 Knowledge Rep and Reasoning
Thus, the CSR may not get the level of information that
she needs. This is not an artifact particular to this appli- Insurance employees such as CSRs and PMs must rea-
cation but is due to the logistical diculties of creating son about services, benets, coverage information, and
and continually modifying a large number of screens. business rules. A service represents some class of med-
Making inquiries via codes: Many insurance com- ical services, such as Surgery or X-rays. A benet is
panies have a code-based scheme for answering cus- some component of an insurance product which covers
tomers' questions. Whether automated or not, such or excludes a particular service. For example, the Drug
schemes operate using a table-lookup methodology. Benet may generally cover Prescription Drugs, though
When a customer calls to inquire if a particular service is it may exclude certain classes of prescription drugs, such
covered, he is asked to provide the CSR with the proce- as Experimental Drugs. Coverage information indicates
dural (CPT) code, which represents the service which is which services are covered or excluded by which ben-
to be performed, and the diagnostic (ICD-9) code, which et, as in the above example. Business rules limit or
represents the diagnosis that is the reason for the service modify the ways in which services are covered by ben-
that is to be performed. The CSR then feeds this pair ets. Rules include cost-share rules, which specify the
of codes into the system, which responds with the infor- deductible and co-pay of a covered service, access rules,
mation that the service is or is not covered, along with specifying whether it is required to go to a network
the appropriate cost of the service. provider, and administrative and medical rules specify-
The advantage of code-based inquiry is that often ing other restrictions. For example, there is a cost-share
customers' question can be answered quickly and un- rule specifying a 20% co-pay for surgical services, and
ambiguously. The disadvantages are: First, customers an administrative rule stating that patients in drug re-
must know complete code information before they can habilitation programs lose all rehab benets for a year if
call their CSRs; this slows down the inquiry process. they are non-compliant.
Second, the code-based scheme only allows questions CSRs and PMs must deal with a large amount of in-
at a detailed level of granularity. But customers of- formation. There are tens of thousands of services of
ten wish to ask general questions, such as Are routine interest. There may be several thousand business rules
immunizations covered?. Third, code-based systems
for an insurance product; many of these rules (such as
do not allow questions that rely on information in ad- cost-share) will apply to a great number of services.
dition to codes such as If a 44-year-old woman had In developing BenInq, we needed to focus on the ex-
a mammogram two years ago, will she be covered ternal representation|the representation that the CSRs
for a routine mammogram now? , in which the treat- and the PMs see, as well as the internal representation,
ment history of the patient is relevant. Fourth, updating which the inner workings of the system use. Since Ben-
code-based systems is an exceedingly dicult process. Inq would be used primarily to answer questions about
There are tens of thousands of CPT codes and tens of service coverage and applicable business rules, the in-
thousands of ICD-9 codes, and a large subset of the pos- ternal representation had to facilitate ecient reasoning
sible pairings of these codes must be considered.1 about these issues. In addition, to ease use by CSRs and
Desiderata | Toward Replacing Existing Tech- PMs, the external representation had to be perspicuous
nology: We aimed to develop an expert system that to employees of the insurance company. Moreover, PMs
supports benets inquiry but avoids the problems of needed to directly manipulate knowledge in the system
both the text-based and code-based systems. In par- in order to modify products. The easiest way to achieve
ticular, we wished to develop a system that this was to have a direct mapping between the inter-
 allows questions at varying levels of granularity nal and external knowledge representations. A complete
 gives clear, unambiguous answers to commonly asked direct mapping is not always possible.
questions The choice of representation resulted from the obser-
 allows representation of very large amounts of material vation that many services can be organized in a taxo-
and navigation around a large information space nomic manner (see Figure 1). For example, the services
 supports connections among related topics of PET Scans, Lab Tests, EKG/EEG, and Genetic Test-
 supports easy updates and modications ing are all subtypes of Diagnostic Services. The taxon-
The ability to modify is important because products omy generally is at least several layers deep. Benets
change so frequently; an outdated benets inquiry sys- likewise can be arranged taxonomically (see Figure 2).
tem is useless. Thus, the system had to be usable not The structure, however, is not purely taxonomic. Cer-
only by CSRs but also by policy modiers (PMs), the tain services have multiple supertypes. For example, Ge-
insurance company employees responsible for making netic Testing is both a subtype of Diagnostic Services
changes within a particular insurance product. and of Family Planning Services. Thus, the structure is
a dag rather than a tree.
1
The nightmarish prospect of updating millions of code The quasi-taxonomic structure led us to consider a se-
pairs is somewhat mitigated by the fact that CPT and ICD-9 mantic network|in particular, an inheritance network
codes have their own structure and support some rudimen- with exceptions|as our KR structure. There are e-
tary abstraction. cient algorithms for reasoning with a semantic network.
Moreover, such networks are natural to laypersons such in well-understood forms such as inheritance with excep-
as CSRs and PMs; this naturalness guarantees a di- tions. Nonmonotonicity is also present in this applica-
rect mapping between the external and internal knowl- tion in the regulations that apply to a node; we discuss
edge representations. We needed to determine, however, this further in section 3.4.
whether a semantic network was capable of encoding all
the knowledge that was needed for the benets inquiry 3.2 Representing Rules
system, specically, (1) coverage or exclusion of medical Certain rules lend themselves to representation within a
services and (2) rules limiting or modifying the ways in semantic network. Consider the rule:
which a service was covered. There is a co-pay of 20% for diagnostic services
3.1 Representing Coverage Information To represent this rule using the standard inheritance
network model, one could have a node representing the
Coverage and exclusion information can easily be en- services which have a 20% co-pay, and a subtype link
coded in the semantic network. In addition to the sub- between the Diagnostic Services node and this node.
type link, one introduces covers and excludes link which On the other hand, a more complex rule such as
connect benet and service nodes. Thus, to represent Patients in Drug Rehabilitation programs lose all rehab
the information that Diagnostic Services are covered by benets for a year if they are non-compliant
the Diagnostic Benet, one places a covers link between cannot be so easily represented. One could posit a node
Diagnostic Benet and Diagnostic Services; to represent that represents the services which have the property that
the information that PET Scans are excluded by the Di- if patients are non-compliant wrt that service, then they
agnostic Benet, one places an excludes link between lose all rehab benets for a year, and then have a sub-
Diagnostic Benet and PET Scans. The covers and type link between the Drug Rehab Services node and
excludes links propagate along the taxonomy, so that, this node. But such a node appears quite articial, and
e.g., we can reason that Diagnostic Benet covers Blood well outside of the spirit of a semantic network, where
Work, since Blood Work is a subtype of Diagnostic Ser- nodes are supposed to represent easily understood and
vices. The propagation is relative to specicity consid- expressible concepts.
erations, so we do not use propagation to conclude that Instead, we choose to represent rules as well-formed
PET Scans are covered by the Diagnostic Benet: the formulae in some classical logic. For example, the rule
path from PET Scans to Diagnostic Benet along the ex- above could be translated as
cludes link is more direct than the path from Pet Scans Drug-rehab(p) ^ enrolled(x,p,i) ^ 9 j sub(j,i) ^ non-
to Diagnostic Benet along the covers link. compliant(x,p,j) ) ((Drug-rehab(p2) ^ enrolled(x,p2,i2) ^
It is obvious that there is a close connection between time(start(i2)) - time(end(j))  1year) ) : covered(x,p2)).
this structure and a standard inheritance network with The w translation of business rules exists in the internal
exceptions [Horty et al., 1990], and in fact we can make representation.
this connection (an example of a direct mapping) precise These rules seem to be attached to specic nodes in the
in the following way: Replace each benet node by a ser- net. For example, the rule specifying that the diagnostic
vice node which represents all services covered by that co-pay is 20% ought to be attached to the Diagnostic Ser-
particular benet (so that, e.g., the Diagnostic Benet vices node in the network. This observation prompted
node would be replaced by the node Services Covered the development of FANs (formula-augmented semantic
by Diagnostic Benet). Replace each covers link by a networks). A FAN is a structure which allows attaching
subtype (isa) link, and each excludes link by a cancels ws to nodes in a semantic network. Formally, a FAN is
link. One then uses a standard traversal algorithm for a tuple consisting of a set of nodes, a set of ws, a set
inheritance networks with exceptions [Horty et al., 1990; of links (subtypes) on nodes, a (possibly empty) set of
Stein, 1992] to determine whether or not services are cov- partial orderings on these links, allowing the prioritiza-
ered. Both representations are useful: the standard in- tion of links in cases of multiple supertyping, and links
heritance network with exceptions representation allows connecting nodes and sets of ws. Details are given in
us to use well-documented algorithms without modica- [Morgenstern, 96].
tion, while the explicit use of covers and excludes links The semantics of a rule attached to a node is roughly:
is more accessible to CSRs and PMs. the rule attached to a node usually applies to the ser-
The importance of the cancels links in the standard vices represented by the node. Nonmonotonicity is again
inheritance representation points out that much of the present; the rule may not apply to all subtypes of the
relevant information for benets inquiry is nonmonot- node. Thus, a natural question is: are rules inherited
onic. It is rarely the case that a class of services is along taxonomic lines? We discuss this issue in sec. 3.4.
covered or excluded by some benet; there are almost
always exceptions. This nonmonotonicity is hardly sur- 3.3 Categorizing Nodes
prising; proponents of nonmonotonic logic [Reiter, 1980]
have argued since its inception that nonmonotonic rea- Benet nodes correspond to the types of benets pro-
soning is common in everyday reasoning (including the vided by a particular insurance product. There are
business world). What is surprising is how rarely non- about twenty to thirty benet nodes in a semantic net-
monotonicity has entered commercial applications, even work representing a typical insurance product.
 In contrast to benet nodes, which are few in number, Benefit Inquiry System

there are many service nodes. Medical care providers File Benefits General Rules Services Options Modify Help

speak of tens of thousands of medical services. Even if Lab Test

we do not represent all of these, we need to have some

Diagnostic x
PET Scan
Covered By: Diagnostic Benefit

broad classication of services. A rst cut at classica- EKG/EEG Associations

tion is provided by the insurance industry, which cat- Lab Test Blood Work Rules:
Detail

egorizes services according to (professional) procedure,

Summary
AAAAAAAA
AAAAAAAA AAAA
AAAAAAAA
AAAARoutine
AAAAAAAA
AAAAAAAA AAAA
AAAAAAAA
PAP AAAAPAP Cost Admin

condition or setting. Examples of professional procedure

AAAAAAAAAAAA AAAA
AAAAAAAA
AAAAAAAA
AAAAAAAAAAAAAAAA
AAAAAAAA
AAAAAAAA AAAA
AAAAAAAA
AAAA Medical Access

services, which are categorized according to the profes-

Genetic
AAAA AAAA
AAAAAAAAAAAA Pregnancy Sickness
AAAA AAAAAAAA
Testing
AAAA AAAAAAAAAAAA PAP
AAAAAAAAAAAAAAAA

sional who provides the services, are Surgical services,

Combined
Diagnostic Major Diag
Surgery Invasive

Therapy, and Diagnostic services. Examples of condi-

Combined Rules
CACS & CAA Rules
Text: Only network Labs

tion services, which are categorized according to a par-

Endoscopic Routine x
Endoscopic covered
Covered as Ancillary: Logical Form: ....

ticular class of conditions, include Maternity and Mental

Sickness
Inpatient Hospital Endoscopic Text: Copay 20%
Hospice Logical Form: ...

Health. Examples of setting services are Hospital Ancil- OK OK

lary services or Inpatient Hospice Room and Board.

Each of these classes of services itself contains a few
major classications, roughly corresponding to the way Figure 1: CSR's view of BenInq | Diagnostic Services
laymen speak of medical services. Formally, a classica-
tion is considered major if it is a root node in the service triple, if common enough, is given its own node. Thus
taxonomy. Examples are Surgical, Diagnostic, Preven- we have Routine Endoscopic and Sickness Endoscopic. 2
tive, and Hospital Services. There are between 30 and Second, each node is associated with a (possibly empty)
40 major classications for a typical insurance product. list of settings which change the coverage or regulations
If we decide to explicitly represent tens of thousands of that node. For example, the node Lab Tests would
of services in the semantic network, each major classi- have associated with it a list of settings including Hospi-
cation will have a very large number of nodes, much tal and Hospice, since payment for Lab Tests is dierent
too large to display or for a CSR to comprehend. We in those settings. We discuss this further in section 4.1.
are faced with a trade-o of ease of display, navigation,
and comprehensibility, versus granularity and compre- 3.4 Reasoning Methods
hensiveness. Achieving arbitrarily ne levels of granu-
larity for all types of services will result in a system that Two types of reasoning are needed for the semantic net-
may be too unwieldy for a CSR to use. work:
Our approach is to generate nodes in the (external 1. Determining if a node is covered or excluded by
representation of the) semantic network according to the some benet.
following principle. A subtype is explicitly given only if 2. Determining which rules apply to a node.
there is a reason: for example, a particular subtype may The rst problem has been studied extensively in the
not be covered, or dierent rules may attach to the sub- literature [Horty et al., 90; Stein 92]. BenInq implements
type. Thus, Endoscopic Surgery is given as a subtype of Stein's (92) algorithm to solve the problem.
Surgery because there are dierent rules for Endoscopic The second problem has been studied extensively by
Surgery than for standard Surgery; Routine Endoscopic [Morgenstern, 96]. A short summary follows. The rules
Surgery is given as a subtype of Endoscopic Surgery be- that apply to a node are not merely the rules that are
cause Routine Endoscopic Surgery is not covered. On attached to a node. For example, the rules that are at-
the other hand, there are no explicit nodes for appendec- tached to Diagnostic Services should also apply to the
tomy or gall bladder surgery, because these are standard node Lab Tests. It might seem that given a node of in-
surgeries; one can always just look at the Surgery node terest (which we call the focus node), we can collect all
to determine coverage and the rules. rules attached to supertypes of that node3 . But such an
Indeed, a fully comprehensive network would have approach could lead to rampant inconsistency; it is very
considerably more than tens of thousands of nodes. As possible that one of the rules at Lab Tests is inconsistent
indicated above, insurance companies often speak of with the rules at Diagnostic Services.
procedure/condition or procedure/setting pairs or even The issue of which rules apply to a node in a FAN is
procedure/condition/setting triples. For example, Lab known as the w-inheritance problem. The solution is
Tests performed in a hospital are generally paid in a dif- to compute a maximally consistent subset (mcs) of all
ferent manner (dierent schedule of payment and dier- the rules attached to nodes to which there is a positive
ent regulations) than Lab Tests performed in a doctor's undefeated path from the focus node. Since there may be
oce or home. Expanding the network to include these 2
Of course, these are really sets of pairs; Sickness Endo-
pairs and triples would mean that the network would scopic is the set of pairs (Sickness Endoscopic, x) where x
have millions of nodes. This would not only yield an un- stands for one of a large number of conditions for which en-
wieldy network, but would also be extremely wasteful: doscopic surgery is considered appropriate.
most pairs and triples are not of interest. We deal with 3 Or more precisely to nodes to which there is a positive
this problem in several ways. First, sometimes a pair or undefeated path [Horty et al., 90] from the focus node.
many such maximally consistent subsets, one prefers the Benefit Inquiry System

preferred maximally consistent subsets (pmcs's) based on File Benefits General Rules Services Options Modify Help

criteria of specicity and path preference. For example,

if there are two mcs's at a focus node, and they dier in
Modify Node Attributes

Generic

that the rst has a w from a more specic node than

Node Name Endoscopic Surgical Benefit
Benefit

the second, than the rst is preferred to the second.

Description Benefit covering routine endoscopy
Surgical Surgical
Benefit

The algorithm to determine a pmcs at a focus node Node Type Benefit Catagory
Aesthetic

is given in [Morgenstern, 96]. It calls for rst prepro-

Cosmetic

cessing the semantic network to remove con icted and

PCA

pre-empted edges. Then the network is traversed up- SuperTypes

ward, starting at the focus node. As each node is vis- Diagnostic

Benefit
Hypnosis\
Acupuncture
SubTypes SubEvents

ited, the node is marked, and a pmcs of the collected ws SummRules

along with the new ws is computed. At each upward

Supplies
Benefit Detail Rules
Covers

branching point, the branches are ordered according to Excludes OK Cancel

the given prioritization, and the branches are recursively

Subtype

traversed according to that order.

The most computationally intensive part of this Figure 2: Surgical Benet before updating
process is computing the pmcs, and in particular, de-
termining if a set of ws is consistent.4 Our system uses attribute-inheritance component recomputes the cover-
the method of Selman and Kautz (1993) to randomly age or exclusion of the relevant nodes. If a w is added,
produce a valid truth assignment to a set of ws. If it is deleted, or modied at a node, the w-inheritance com-
consistent, this procedure generally works very rapidly. ponent recomputes the ws that apply to that node and
Thus, if after a short time no valid assignment has been the nodes beneath it in the dag.
produced, the system tries to show inconsistency in a The inquiry tool uses the attribute-inheritance compo-
more traditional manner. nent of the reasoning engine in addition to the graphical
interface. Since the w-inheritance component is by far
4 System Description the most computation-intensive portion of BenInq, we
BenInq has been implemented in VisualAge Smalltalk on try to restrict its use within the system as much as pos-
an OS/2 platform. The system incorporates two tools: sible. It is therefore not used by the inquiry tool. Thus,
the inquiry tool which is used by CSRs to answer cus- the set of ws that apply to a node are stored during the
tomers' questions, and the authoring tool which is used authoring process.
by PMs to modify products. BenInq has two main com- Below, we brie y discuss how BenInq is used by its
ponents, a graphical user interface and a reasoning en- two classes of users, CSRs and PMs.
gine, which are used by both tools. The graphical user 4.1 A CSR's view of BenInq
interface is used to navigate through the large number of
nodes and to display the dags that are rooted at specic BenInq gives a CSR the ability to view various benets
benet or service nodes. It consists mainly of functions and services by either selecting the appropriate item in
to search through the network, draw nodes in the net- one of several pull-down menus (organized by node type)
work structure, and display dialog boxes. We exploited or by using a node search mechanism. This allows the
the power of VisualAge Smalltalk, which comes with a li- CSR to view all the subtypes of a chosen service or ben-
brary of drawing tools (particularly for menus and dialog et, and to view all top-level services that are either cov-
boxes) and facilitates the creation of graphical interfaces. ered or excluded by a chosen benet. The system also
The reasoning engine consists of two components: gives the CSR the ability to determine which services
one for performing standard attribute inheritance and are covered or excluded, and by which benet. More-
one for performing w inheritance. The attribute- over, the CSR can easily determine all the rules that
inheritance component uses the algorithm described in apply to a particular service, including those inherited
[Stein, 1992]. The w inheritance component uses the from its ancestors. The CSR can choose to view rules of
algorithm described in [Morgenstern, 1996] and summa- a particular sort (e.g., cost-share or access) or all rules.
rized in section 3.4. BenInq also allows the CSR to view the associations at
The authoring tool uses both components of the rea- a particular node. Such associations allow the CSR to
soning engine as well as the graphical interface. The see how a procedure's coverage may change depending
graphical interface is used to search for the node or por- on the setting in which the procedure is performed.
tion of the network to be modied. If a node is inserted Figure 1 illustrates how a CSR would use the system
or deleted, or inheritance information is changed, the to answer questions concerning dierent Lab Tests. The
CSR can view the relevant portion of the network either
4
Determining consistency is only semi-decidable for the by rst selecting Diagnostic Services using the pulldown
predicate case; it is decidable but intractable for the propo- menu and then clicking on Lab Tests, or by using the
sitional case. node search mechanism. The screen displayed provides
 Benefit Inquiry System 4.2 A Policy Modier's view of BenInq
File Benefits General Rules Services Options Modify Help
The authoring tool provides the Policy Modier with
several features to eectively modify benets and ser-
Generic
Benefit Endoscopic Routine
vices. The PM can add new benets or services, modify
existing services and benets, and delete existing prod-
Surgical Benefit Endoscopic

ucts. The PM can add a node either before or after

Surgical
Benefit Surgical

Aesthetic
Cosmetic Routine Endoscopic an existing node in the network. She can also spec-
Covered By: Endoscopic Surgical
ify other supertypes, subtypes, and the rules that apply
to the new node. For service nodes, the PM can spec-
PCA
Associations

Rules:
Detail
ify whether the service is covered (or excluded), and by
which benet, or whether coverage/exclusion is inherited
Diagnostic Hypnosis\ Summary
Benefit Acupuncture
Cost Admin

Medical Access from another service node. Similarly, the PM can mod-
ify various attributes of any existing benet or service
Supplies Benefit
Covers

including name, list of supertypes, prioritization of the

Combined
Excludes

Subtype

supertypes, list of subtypes, set of rules that apply to

that node, and for service nodes, the coverage/exclusion
Figure 3: Surgical Benet after updating information and service classication. Finally, the PM
can delete services or benets from the system.
the CSR with much useful information at a single glance. Figures 2 and 3 demonstrate how the authoring tool
First, it is immediately apparent which nodes are not would be used by the PM to add a new benet, Endo-
covered; they have an x in the upper right corner. All scopic Surgical Benet, to explicitly cover the previously
other nodes are covered. There are visual cues indicating excluded Routine Endoscopic Services. Since the new
that a node is not covered by the expected benet (in benet is a form of Surgical Benet, she adds it as a
this example, Diagnostic Benet). Thus, for example, subtype of Surgical Benet. Figure 2 shows a portion of
the CSR nds out that Routine Pap (which is a dierent the dag displaying the Surgical Benet. The PM adds a
color than other nodes), although a lab test, is not cov- new node as a subtype of Surgical Benet (by choosing
ered by the Diagnostic benet. To nd out which benet the appropriate option from the Modify menu), and the
covers a particular node, the CSR clicks on that node. In dialog window pops up. She can then specify the name,
Figure 1, since Lab Tests is currently selected, a dialog type, subtypes, and supertypes as well as the rules that
box shows, among other things, that Lab Tests is covered apply to the new node. Figure 3 shows the situation
by Diagnostic Benet. If the CSR clicks on the Routine in which the new benet, Routine Endoscopic Benet
Pap node, a similar dialog box comes up indicating that has been added as a subtype of Surgical Benet. The
it is covered by the Preventive Benet. Moreover, the service node Routine Endoscopic has also been modied
CSR is given the option to switch views to a tree rooted to change its coverage/exclusion information so that it
at the Preventive benet so that all information relevant is now covered by the new benet (as opposed to being
to Routine Pap becomes visible. excluded by Diagnostic Benet).
In addition, the CSR may view various rules applic- 5 Discussion
able to the node by clicking on the appropriate button.
In Fig. 1, the CSR has selected the Combined button, 5.1 Evaluating the System
indicating that all rules applying to the service should be BenInq represents a signicant advance in existing tech-
displayed. The lower right portion of the screen shows nology for benets inquiry. The departure from a code-
the dialog box that pops us when the Combined button is based system to a knowledge-based system is a big
selected. Note that both the text and logical forms of the change for the insurance industry. This is, as far as
rules are given. Similarly, the CSR can view the associa- we know, one of the few large-scale industry applica-
tions at a particular node by clicking on the appropriate tions that explicitly models nonmonotonic reasoning and
button. The lower left portion of Figure 1 displays the uses formal nonmonotonic reasoning techniques. This
dialog box which pops up when the CSR clicks on the is clearly an advance for the insurance industry, which
associations button.5 This box lets the CSR know that previously could deal with exceptions only in an ad-hoc
if lab tests are performed in an inpatient hospital (resp. manner. It also validates the work of the nonmonotonic
hospice) setting, the procedure is covered by the Inpa- reasoning community, which has often been accused of
tient Hospital (resp. Hospice) Benet and is subject to being insular and producing no useful results for the real
those rules. If the CSR wishes to see these rules, she world.
clicks on the desired setting in the dialog box and the BenInq allows CSRs to immediately determine
appropriate regulations appear. whether or not a service is covered and which rules ap-
ply to that service. It further provides PMs with an
5
In reality, only one of the rules/associations dialog boxes easy method of updating and modifying insurance prod-
can be shown at one time. Both dialog boxes are shown for ucts. The current version of BenInq was completed after
explanatory purposes. our consulting engagement had ended. The medical in-
surance company received an earlier version of BenInq, be generalized to other domains as well. The construct
which incorporated some but not all of the reasoning of a FAN and its associated algorithms may prove useful
methods described in this paper. This version supported in other domains which satisfy the following criteria:
standard inheritance with exceptions and attaching ws  there exists a large amount of taxonomic information
to a node, but did not perform w-inheritance. Thus  there exists a signicant amount of non-taxonomic in-
PMs who modied the rules at a node had to modify formation, conceptually linked to the taxonomic infor-
the rules at all subtypes of the node as well in order to mation
ensure integrity. Despite this drawback, the earlier ver-  the non-taxonomic information can be mapped into
sion of BenInq received excellent reviews from the CSRs ws
and PMs who used it. They needed virtually no train- The construction of an internal representation which
ing, surpassing our most optimistic expectations. They closely parallels the external representation is a useful
were delighted with the ease of nding out information technique for domains where there is a non-technical
and modifying products. class of users who must directly manipulate the knowl-
The desiderata outlined in section 2 were largely sat- edge structure. This technique is best suited to domains
ised, although the last desideratum, supporting easy in which there is a natural translation between the user's
updates and modication, was only partly satised. In conceptual model and a knowledge structure which lends
particular, while it is easy for a PM to add, delete, or itself to ecient reasoning methods.
modify links and nodes in the network, it is not always Examples of such domains include other parts of the
easy to modify or add ws. The PM may easily modify insurance industry, such as life, property, and casualty
the text version of business rules, but cannot modify the insurance; legal reasoning, especially case law (cases are
logical form of business rules. often organized taxonomically, and dierent legal rulings
This is a direct consequence of the fact that there is are associated with cases); and medical reasoning.
no direct mapping between the internal (logical form) We are currently using these techniques in the devel-
and external (text) representations of business rules, as opment of a system to dynamically congure property
discussed in section 3. This is in direct contrast to all and casualty insurance products. In this system, the
other parts of the semantic network, in which there is constructs of composition and subtyping are both im-
a natural, direct mapping between the internal and ex- portant. The methods developed to reason with sub-
ternal representations. Fixing the problem requires the types are also applicable for reasoning with composition,
construction of a mapping between the text and logical with appropriate modications. This indicates that the
form representations; this problem is in general dicult techniques described in this paper are useful not only in
and beyond the scope of this work. However, one can domains in which taxonomic information is central but
identify certain types of rules for which there is a rea- in domains in which composition is a useful construct.
sonable mapping. We have sketched out mappings for This has the potential to expand the applicability of this
cost-share and access rules, as well as for administra- work. We hope to report on this research in the future.
tive rules involving temporal and spatial (setting) con- Acknowledgements: We are grateful to Lynn Brick-
straints. The aim is the construction of templates which ing, Kathleen Jamison, Bill Wynne, Tracy Edmonds,
will allow the PM to directly enter rules in logical form, Terry Brown, and Kay Krimmer for their assistance in
and will generate a reasonable text equivalent. the knowledge engineering phase. An earlier version of
What were the alternatives? The success of BenInq BenInq was implemented with the assistance of Mike
was due to two facts: rst, the realization that the bulk Ashwell and Jason Springer. Thanks also to Ernie Davis
of the information was taxonomic, but that a signicant and Benjamin Grosof for helpful suggestions.
chunk was not, and second, the construction of a knowl-
edge structure, the FAN, which allowed the representa- References
tion of both taxonomic and non-taxonomic information, [Horty et al., 1990] J. Horty, R. Thomason, and D. Touret-
and linkage between the two. Existing benets inquiry zky. A Skeptical Theory of Inheritance in Nonmonotonic
systems had ignored the taxonomic nature of the medical Semantic Networks. Artif. Intell. 42, 311-349.
insurance domain. This resulted in systems that were [Morgenstern, 1996] L. Morgenstern. Inheriting Well-formed
extremely spotty in their coverage (text-based systems) Formulae in a Formula-Augmented Semantic Network.
or were unwieldy (codes). Even a more sophisticated Proc. KR-96, 268-279.
rule-based (or production) system that ignored the tax- [Morgenstern, 1997] L. Morgenstern. Inheritance Comes of
onomic nature of the domain would run into the problem Age. Invited Talk, IJCAI-97.
of redundant information and the need for multiple up- [Reiter, 1980] R. Reiter. A Logic for Default Reasoning. Ar-
dates. On the other hand, any attempt to force all of tif. Intell. 13, 81-132.
the information into a taxonomy would have resulted in [Selman and Kautz, 1993] B. Selman and H. Kautz.
highly articial nodes with no clear semantics. Domain-Independent Extensions to GSAT. Proc IJCAI.
5.2 Generalizations and Extensions [Stein, 1992] L. Stein. Resolving Ambiguity in Nonmonoto-
The techniques described in this paper, while developed nic Inheritance Hierarchies, Artif. Intell., 259-310.
specically for benets inquiry in medical insurance, can