=Paper=
{{Paper
|id=Vol-1417/paper8
|storemode=property
|title=Automated Decision Support for Financial Regulatory/Policy Compliance, using Textual Rulelog
|pdfUrl=https://ceur-ws.org/Vol-1417/paper8.pdf
|volume=Vol-1417
|dblpUrl=https://dblp.org/rec/conf/ruleml/GrosofBFKGCS15
}}
==Automated Decision Support for Financial Regulatory/Policy Compliance, using Textual Rulelog==
https://ceur-ws.org/Vol-1417/paper8.pdf
Automated Decision Support for Financial
Regulatory/Policy Compliance, using Textual Rulelog
Benjamin Grosof1 , Janine Bloomfield1 , Paul Fodor1 , Michael Kifer1 , Isaac Grosof1 ,
Miguel Calejo2 , and Terrance Swift1
1
firstName.lastName@coherentknowledge.com; Coherent Knowledge Systems, USA
2
mc@interprolog.com; InterProlog Consulting, Portugal
Abstract. We present a novel technological approach, based on Textual Rulelog,
to automated decision support for financial regulatory/policy compliance, via
a case study on banking Regulation W from the US Federal Reserve. Legal
regulations and related bank operational policies in English documents are en-
coded relatively inexpensively by authors into Rulelog, a highly expressive log-
ical knowledge representation. Key compliance queries are automatically an-
swered accurately and fully explained in English, understandable to non-IT com-
pliance staff and auditors. The prospective business impact of our approach over
the next decade or two is significantly increased productivity and systemic stabil-
ity, industry-wide, worth many billions of dollars.
1 Business Case
A complex set of regulations and associated policies govern a wide range of operations
and activities that financial institutions, such as banks and investment firms, engage
in every day. Compliance, and proof of compliance, are essential: for both external
regulators and internal management. The complexity and amount of regulations and
related policies are ever-growing. The stakes are high: the global financial crisis of 2008
cost the overall economy trillions of dollars3 while individual institutions spent over 100
billion dollars on fines, penalties, and legal settlements4 . To mitigate the probability
and severity of another financial crisis, the US and many other advanced industrial
countries enacted a broad range of new regulations and policies. While these measures
have improved somewhat the systemic risk of the global financial system, much more
remains to be done to improve it further. Additional regulations, along with associated
policies, are continuously being developed and issued, and likely will continue to be for
many years in the foreseeable future.
3
‘Financial Regulatory Reform: Financial Crisis Losses and Potential Impacts of the Dodd-
Frank Act’, GAO-13-180, 2013. http://www.gao.gov/products/GAO-13-180
Accessed 28 May 2015.
4
Banks pay out $100bn in US fines by R. McGregor and A. Stanley,
Financial Times, March 25, 2014. http://www.ft.com/cms/s/0/
802ae15c-9b50-11e3-946b-00144feab7de.html#axzz3a5MXr6vx Accessed
28 May 2015.
2 Grosof et. al
Automated support is needed for compliance decisions and associated analysis, but
currently deployed (i.e., previous) methods are expensive, unwieldy, and often quite in-
accurate. These current methods often take the form of stand-alone or loosely connected
components/services which review transactions or other activities, answer queries, and
issue alerts. Typically, these methods require humans in the loop at run time, often be-
cause the scope of automation only captures part of the substance of the relevant regula-
tions and policies. Overall, compliance is best viewed as mostly an aspect of operations,
whose automation should be woven tightly into operational systems as a whole rather
than as a separate step that sits architecturally only loosely coupled to other operational
activities.
The market landscape in automated financial/regulatory compliance includes sev-
eral kinds of vendors. One is providers of outsourced operations services; they often
use IT automation. A second kind of provider furnishes IT development services. A
third kind of provider furnishes software products (or software as a service) that facili-
tate building compliance systems. Others provide software solutions or data that are part
of implementing particular regulations or policies. Even when these solutions are more,
rather than less, complete for a particular regulation area, the solutions usually must
accommodate extension to, and integration of, institution-specific policies, especially
for larger institutional customers. In addition to utilizing vendors, larger institutions
tend to implement quite a bit in-house as well. As with software applications generally,
there is a momentum towards subscription and frequent updating, rather than traditional
(infrequent-major-release) licensing, in both the software and the related data.
An example of a recent regulation issued after Dodd-Frank is the US Federal Re-
serve Act’s Regulation W (“RegW”).5 It concerns activities/transactions between a bank
and its counterparties (mainly, companies) that are defined as “affiliates” which share
ownership, control, or advisory relationships. It is designed to limit concentrations of
risks to an individual institution, and the banking system as a whole. Compliance re-
quires avoiding transactions, or reporting transactions, under certain conditions.
2 Technological Challenges
There are a number of factors that make it difficult for previous methods to solve the
problem well. The regulations (and associated policies) are frequently very compli-
cated in both their logical/semantic substance and their English syntax (“legalese” is
notorious). The regulations are full of meta information and rife with important excep-
tion cases requiring defeasibility. The body of regulations is voluminous, continually
increasing, and ever changing. For these often high-stakes compliance decisions, ac-
curacy and reasoning efficiency (near real time) are necessary, as are provenance and
audit trails to demonstrate (often as potential legal evidence) why decisions were taken.
The complicated English rules (and definitions) must be encoded into automated form
suitable for reasoning. And a variety of enterprise data, not just transactions, needs to
5
Part 223 Transactions between member banks and their affiliates (Regulation W), Elec-
tronic Code of Federal Regulations (e-CFR), GPO http://www.ecfr.gov/cgi-bin/
text-idx?tpl=/ecfrbrowse/Title12/12cfr223_main_02.tpl Accessed 28
May 2015.
Automated Decision Support for Financial Regulatory/Policy Compliance 3
be integrated with the implemented regulations in order to perform that reasoning. Sub-
ject matter experts without extensive training in logic or programming (SME’s) need to
be involved closely in developing, testing, and debugging the encoded implementation.
Yet it’s hard for them to understand, much less contribute to, the implemented form.
3 Rule-based Solution
We developed and applied an approach based on Textual Rulelog [3, 1, 2], implemented
in our (Coherent Knowledge Systems’) Ergo SuiteTM platform (“Ergo”)6 . Ergo includes
both a Reasoner and an integrated development environment (IDE), Studio. Ergo pro-
vides: very high logical expressiveness, including for higher-order, defeasibility, quan-
tifiers, head disjunction, and meta knowledge; and efficient dynamic automated rea-
soning capabilities, including fully detailed user-navigable explanations (for each an-
swer) that are quite understandable by SME’s. Ergo integrates tightly some English
natural language capabilities both for authoring (i.e., developing) rules – going from
English phrases to logical expressions (“text interpretation”) – and for generating an-
swers with explanations – which leverages mapping from logical expressions to English
phrases (“text generation”). Ergo also includes connectors that import knowledge from
OWL/RDF and other forms (e.g., relational databases and spreadsheets), then tightly
integrate the imported knowledge into overall reasoning.
Our methodology for rule development (“authoring”) had several steps. Starting
from English source sentences in published regulations and other documents, we ar-
ticulated a set of English encoding sentences that were clearer and syntactically more
self-contained, along with some additional background knowledge encoding sentences.
Then we encoded each encoding sentence into a logical rule in Ergo syntax, using a
textual terminology style in which each English phrase is mapped closely (nearly iso-
morphically) to and from a corresponding Ergo logical term (typically higher-order);
English words were used as Hilog functors, then composed into phrasal terms. We en-
coded additional rules that specified textual templates for text generation (used heavily
in explanation) and text interpretation (as an extension of the original work). We also
developed ontology mapping rules that supported effective reasoning with imported
OWL knowledge. We iteratively tested (queried) and debugged the rules, intensively
using Ergo’s explanations capability. Explanations helped to find both missing knowl-
edge and incorrect knowledge.
Ergo is flexibly deployable. The implemented Ergo based component, including
its knowledge bases, can be integrated into an overall compliance product, service, or
enterprise application, in several different ways. It can be queried, started or stopped,
loaded, and configured via its Java API. Under development are additional methods
including a RESTful web service wrapper.
Ergo scales well computationally in several important regards. First, its fundamen-
tal logic (i.e., semantic knowledge representation), Rulelog, is an extension of database
logic, i.e., of well-founded declarative logic programs, and is equipped with restraint,
a form of bounded rationality that allows one to ensure that inferencing is worst-case
polynomial-time. Second, its fundamental reasoning algorithms have many performance
6
http://coherentknowledge.com
4 Grosof et. al
optimizations, including compilation, transformation, indexing, and subgoal reorder-
ing; they employ a form of LP tabling [4, 5] that caches inferences and thereby reuses
previous computations on subgoals when computing an answer to a new goal. Ergo
inferencing/computation is primarily in main memory; however, it hooks up to other
components, such as databases and triple stores, that are disk-based. On an ordinary
current laptop or desktop, it scales up to millions, but not billions, of complex infer-
ences and associated dependencies and fact assertions, while occupying roughly 5-10
gigabytes of RAM. One can take reasoning to yet larger scales by using processors
equipped with more RAM, and/or by distributing reasoning computations across mul-
tiple Ergo instances on multiple processors.
Ergo has special performance optimizations for RDF data in bulk to be queried in
SPARQL from Ergo, translated into Ergo, and loaded into Ergo (for use in reasoning),
at high speed. This import and tightly integrated reasoning scales to millions of RDF
triples. Under development are additional methods for distributed processing methods
in order to scale this up further.
4 Results
Our approach was originally developed in coordination with, and in support of, an
overall application-piloting proof-of-concept (PoC) effort, called “FIBO Rules”, con-
ducted by the Enterprise Data Management Council (EDMC), a leading financial-sector
international industry-government consortium. The PoC was conceived at the FIBO
Summit, held in June 2013 at the SemTechBiz SJ conference. RegW was identified
by banking participants as a challenge regulation that had significant complexity and
industry urgency. One of us (B. Grosof) acted as technical lead for the PoC. Besides
Coherent, other participants in the PoC included Wells Fargo Bank, SRI International,
and GRCTC (Ireland). The PoC very successfully reached its goal of demonstrating
how a representative subset of RegW could be effectively automated using Textual
Rulelog while also leveraging the EDMC/OMG Financial Industry Business Ontol-
ogy (FIBO). The business benefits of our rule-based solution were dramatic, includ-
ing: higher accuracy; solution development that has lower cost, greater agility, higher
reusability, and more SME participation; and greater scope of automation including
SME-understandable explanations/provenance. The PoC was presented to industry au-
diences in 2014 at the OMG Technical Meeting plenary session on March 26, an EDMC
webinar on June 26, and multiple sessions of the SemTechBiz SJ industry conference
on August 21.7,8,9,10
7
Semantics - Crossing the Chasm Workshop, OMG Technical Meeting Special Event,
March 26, 2014, Reston, VA, http://www.omg.org/news/meetings/tc/va-14/
special-events/Sem-Fin_Day.htm Accessed 28 May 2015.
8
EDM Council Industry Webinar 26 June 2014, https://www.youtube.com/watch?
v=8c9WmMW16t8&feature=youtu.be#t=11m00sAccessed 28 May 2015.
9
‘Leading Financial Industry Consortium Proof-Of-Concept Features Coherents Ergo
Suite’, Coherent Knowledge Systems website, http://coherentknowledge.com/
leading-financial-industry-consortium-proof-of-concept-features-
coherents-episto/ Accessed 28 May 2015.
10
Semantic Technology and Business Conference, August 19-21, 2014, San Jose, CA, Ses-
sions titled Detailed Explanations in English of Rich Reasoning for E-Learning and
Automated Decision Support for Financial Regulatory/Policy Compliance 5
The technical benefits of our approach, that enable the business benefits, revolve es-
pecially around expressiveness, authoring, and explanations. Textual Rulelog provides
very high expressiveness of rules and answers, by combining: defeasible higher-order
logic formulas; justification graphs that are selectively expandable/collapsible; and text
generation/interpretation. This reduces the cost/time, and logical skill, needed for au-
thoring and increases the understandability of explanations.
Next, we give some illustrative examples of rules and explanations for this RegW
case study. An example query in Ergo is:
What proposed transactions are prohibited by RegW?
Show (?Bank,?Company,?Amount).
Here, the prefix “?” before “Bank” indicates a logical variable; likewise, before
“Company” and before “Amount”. The set of answers for a query can be displayed in
Ergo Studio as a table of answer tuples, where each tuple has one element per vari-
able in the query. An example answer tuple for this particular query is: ’Pacific
Bank’, ’Maui Sunset’, 23.0. This answer tuple corresponds to the variable
tuple “?Bank, ?Company, ?Amount”. Here the “23.0” indicates millions of dol-
lars (for brevity’s sake).
The user can then ask “Why” by right-clicking on that answer, and Ergo in response
will automatically generate and present an explanation. The explanation is a justification
graph, presented as a sideways tree, in which each line corresponds to one step of rule
inferencing and is analogous to a line in a natural-deduction style proof (remember
high school geometry?). Each explanation line has a handle at its left. By clicking on
the handle of a line, the user can selectively expand a portion of the explanation. This
can be done repeatedly to drill down into full detail of the explanation (leaf nodes of
the justification graph correspond to asserted facts). After the user has inspected that
portion of the explanation, the user can then collapse the portion (i.e., its subtree) by
again clicking on the same handle. Figure 1 is an example of a detailed such explanation,
for the example answer tuple above.
Ergo generates explanations by first creating a justification graph (JG) whose nodes
are logical facts, then transforming each node in the JG. Text generation occurs during
this transformation of a JG node. Text generation uses mappings of logical formulas to
English sentences; the mappings are specified via Ergo rules (textual templates, men-
tioned earlier).
Ergo’s high expressiveness makes it easier to author rules by mapping from English
to logic, as well as vice versa. Examples of Ergo facts are:
subsidiary(of)(’Pacific Bank’,’Americas Bank’).
advised(by)(’Maui Sunset’,’Hawaii Bank’).
bank(’Hawaii Bank’).
company(’Maui Sunset’).
capital(stock(and(surplus)))(’Pacific Bank’,2500.0).
Compliance, Semantics in Finance: Addressing Looming Train Wreck in Risk Manage-
ment, Regulatory Compliance and Reporting, Semantics: the Technical Lens for Finan-
cial Transparency and Risk Management and Grill the experts on FIBO and its underly-
ing technology, http://semtechbizsj2014.semanticweb.com/agenda.cfm?
confid=82&scheduleDay=08/21/14 Accessed 28 May 2015.
6 Grosof et. al
Fig. 1. An example of explanation
proposed(loan) (from(’Pacific Bank’))(to(’Maui Sunset’))
(of(amount(23.0))) (having(id(1101))).
previous(loan)(from(’Pacific Bank’))(to(’Hawaii Bank’))
(of(amount(145.0))) (having(id(1001))).
proposed(asset(purchase))(by(’Pacific Bank’))
(of(asset(common(stock)(of(’Flixado’)))))(from(’Maui Sunset’))
(of(amount(90.0)))(having(id(1202))).
Examples of non-fact rule assertions are:
/* A company is controlled by another company when the first
company is a subsidiary of a subsidiary of the second company.*/
@!{rule103b} /* declares rule id */
@@{defeasible} /* indicates the rule can have exceptions */
controlled(by)(?x1,?x2)
:- /* the "if" symbol */
subsidiary(of)(?x1,?x3) \and
subsidiary(of)(?x3,?x2).
/*A case of an affiliate is: Any company that is advised on a
contractual basis by the bank or an affiliate of the bank. */
@!{rule102b} @@{defeasible}
affiliate(of)(?x1,?x2) :-
( advised(by)(?x1,?x2)
\or
(affiliate(of)(?x3,?x2) \and advised(by)(?x1,?x3))).
Here “/*...*/” encloses a comment; “:-” means “if”; “@!{...}” is meta info
that indicates a rule id. “@@{defeasible}” means the rule can have exceptions.
“@{...}” encloses a rule tag, a label used for specifying prioritization information
in handling of conflicts between rules, for reasoning about exceptions. “\overrides” is
a predicate used to specify such prioritization-type precedence info. Example rules that
specify an exception case are:
Automated Decision Support for Financial Regulatory/Policy Compliance 7
@!{rule104e}
@{ready market exemption case for covered transaction’}
/* tag for prioritizing */
\neg covered(transaction)(by(?x1))(with(?x2))
(of(amount(?x3)))(having(id(?Id))) :-
affiliate(of)(?x2,?x1) \and
asset(purchase)(by(?x1))(of(asset(?x6)))(from(?x2))
(of(amount(?x3)))(having(id(?Id))) \and
asset(?x6)(has(ready(market))).
/* prioritization info, as one tag being higher than another */
\overrides(ready market exemption case for covered transaction’,
’general case of covered transaction’).
/* If a company is listed on the New York Stock Exchange (NYSE),
then the common stock of that company has a ready market. */
@!{rule201} @@{defeasible}
asset(common(stock)(of(?Company)))(has(ready(market))) :-
exchange(listed(company))(?Company)(on(’NYSE’)).
IRI prefixes can be specified in Ergo syntax via a directive that declares them. An
example is:
:- iriprefix fibof = /* declares an abbreviation */
"http://www.omg.org/spec/FIBO/FIBO-Foundation/20120501/ontology/".
In Ergo syntax, “foo#” indicates that foo is being used as a shortname IRI prefix.
Examples of imported knowledge from FIBO (facts translated from OWL/RDF into
Ergo) are:
rdfs#subClassOf(fibob#BankingAffiliate,fibob#BodyCorporate).
rdfs#range(fibob#whollyOwnedAndControlledBy,
fibob#FormalOrganization).
owl#disjointWith(
edmc#Broad_Based_Index_Credit_Default_Swap_Contract,
edmc#Narrow_Based_Index_Credit_Default_Swap_Contract).
In order to reason effectively with knowledge imported from OWL, Ergo includes
some general axioms that specify the semantics of OWL. Examples of such axioms are:
?r(?y) :- rdfs#range(?p,?r), ?p(?x,?y).
?p(?x,?y) :- owl#subPropertyOf(?q,?p), ?q(?x,?y).
Ergo is also very capable at representing ontology mappings. To tightly integrate the
knowledge imported from FIBO with the other knowledge in Ergo that represented the
RegW regulations and information that originated in English, we developed rules that
mapped between textual terminology and FIBO OWL vocabulary. Examples of such
ontology mapping rules are:
company(?co) :- fibob#BodyCorporate(?co).
fibob#whollyOwnedAndControlledBy(?sub,?parent) :-
subsidiary(of)(?sub,?parent).
8 Grosof et. al
5 Importance and Impact
Our rule-based approach offers the realistic promise to increase the productivity (i.e.,
reduce cost and risk) of financial regulatory/policy compliance by at least several per-
cent. Such a productivity improvement will, over the next decade or two, be worth
many billions of dollars to the global economy, and to the individual institutions that
are players in banking and investment. In addition to the productivity advantages of our
approach, increasing systemic stability has many non-economic benefits. Moreover, the
radically increased transparency afforded by our approach has the potential to signif-
icantly improve the governance of the process of writing and enforcing regulations,
thereby mitigating the critical cluster of problems arising from exploitative gaming of
the regulation system by financial-institution players. These problems are long-studied
and include over-complexity of regulation a.k.a. “overregulation”, “regulatory capture”,
and “regulatory arbitrage”11 .
Acknowledgements
Thanks to all the FIBO Rules effort’s participants for their insights and collaboration
on RegW and financial regulatory compliance via semantic technology, particularly:
Dennis Wisnosky, Michael Bennett, and Michael Atkin of EDMC; David Newman,
Wesley Moore, Cheryl Maske, and Patrick Greenfield of Wells Fargo (David also wore
an EDMC hat for the effort); Elie Abi-Lahoud of GRCTC; and Daniel Elenius, Grit
Denker, Susanne Riehemann, Reg Ford, and John Shockley of SRI International.
References
1. Grosof, B., Dean, M., Kifer, M.: Semantic Web Rules: Fundamentals, Applications, and
Standards (July 2013), tutorial presented at AAAI-2013; available at http://silk.
semwebcentral.org/
2. Grosof, B., Kifer, M., Fodor, P.: The Power of Semantic Rules in Rulelog: Fundamentals
and Recent Progress (August 2015), tutorial presented at RuleML-2015; available at http:
//www.coherentknowledge.com/publications/
3. Grosof, B.: Rapid Text-Based Authoring of Defeasible Higher-Order Logic Formulas, via
Textual Logic and Rulelog. In: Morgenstern, L., Stefaneas, P., Lvy, F., Wyner, A., Paschke,
A. (eds.) Theory, Practice, and Applications of Rules on the Web, Lecture Notes in Computer
Science, vol. 8035, pp. 2–11. Springer Berlin Heidelberg (2013), http://dx.doi.org/
10.1007/978-3-642-39617-5_2
4. Swift, T., Warren, D.: XSB: Extending the Power of Prolog using Tabling. Theory and Practice
of Logic Programming (2011)
5. Swift, T.: Incremental Tabling in Support of Knowledge Representation and reasoning. The-
ory and Practice of Logic Programming 14(4-5), 553–567 (2014), http://dx.doi.org/
10.1017/S1471068414000209
11
‘Wall Street is Using the Power of Dodd-Frank Against Itself’ by A. Davidson, NY Times
(Magazine), May 27, 2015 http://www.nytimes.com/2015/05/31/magazine/
wall-street-is-using-the-power-of-dodd-frank-against-itself.
html Accessed 28 May 2015