In April 2013, the International Consortium of Investigative Journalists (ICIJ) exposed the details of 130,000 offshore accounts. Although there are legitimate businesses which use such accounts, there exist a number of accounts which are possibly linked to international tax fraud and money laundering. The ICIJ investigation was based on 2.5 million records of offshore holdings linked to 170 countries. All these records have been made available for further examination and analysis. Based on these records, a set of facts, rules and queries have been formulated in Datalog+ RuleML 1.01/XML and these rules have been tested and validated against the Relax NG schema for Datalog+ in RuleML 1.01/XML. The usefulness of such rules for offshore holdings analytics is demonstrated here via incremental step-by-step approach, through the discovery of relationships between persons hiding large sums of assets and the offshore companies where the assets are purported to be hidden.
The International Consortium of Investigative Journalists (ICIJ) [
In this article, we demonstrate the effectiveness of expressive Datalog+ RuleML
1.01 [
The article is organized as follows: In section 2, the formalization of our rulebase is illustrated in detail by authoring facts, rules, and executing queries on them, often by checking the consistency of the rulebase. In section 3, the expressivity and the effectiveness of rules used here have been briefly discussed. Finally, we conclude by summarizing the tools and procedures used for rule authoring, schema design, and their validation in Appendix A. 2
Our Datalog+ RuleML 1.01/XML Document incrementally asserts into, retracts
from, and queries, a rulebase (within the <RuleML> element) for a total of 37
transactions: 14 Asserts, 3 retracts, and 20 distinct Queries. The rules formalized
in this document not only cover the smallest Datalog sublanguage, but also cover
the highlights of RuleML 1.01, the superlanguage of the decidable Datalog+/- [
The fact “The president of Azerbaijan is Ilham” is formalized with a binary president relation in RuleML. A query immediately follows, formalizing “Who is the president of Azerbaijan?”, and resulting in an x binding to Ilham. <?xml-model
href=="http://deliberation.ruleml.org/1.01/relaxng/naffologeq_relaxed.rnc" ?> <RuleML xmlns="http://ruleml.org/spec"> <Assert> <Atom> <Rel>president</Rel> <Ind>Ilham</Ind> <Ind>Azerbaijan</Ind> </Atom> </Assert> <Query> <Atom> <Rel>president</Rel> <Var>x</Var> <Ind>Azerbaijan</Ind> </Atom> </Query>
A variation of the above query can be formalized as “Which country does Ilham preside over?”, resulting in a y binding to Azerbaijan.
A generic formalization of the above query can be expressed as “List al l the individuals and countries where the individual is the president of a country ”. Given the only ground fact, x and y represent the individual and the country binding to Ilham and Azerbaijan, respectively.
The rule “No one is both the president and the prime minister of a country” is formalized as an integrity rule with falsity in rule head.
<Assert> <Forall> <Var>x</Var> <Implies> <if> <And> <Atom> <Rel>president</Rel> <Var>x</Var> <Var>y</Var> </Atom> <Atom> <Rel>primeMinister</Rel> <Var>x</Var> <Var>y</Var> </Atom> </And> </if> <then>
<Or/> </then> </Implies> </Forall> </Assert>
The formalization of the fact “Ilham is the prime minister of Azerbaijan”, when asserted with the previous rule, creates an inconsistency because of the already existing assertion “Ilham is the president of Azerbaijan”.
<Assert> <Atom> <Rel>primeMinister</Rel> <Ind>Ilham</Ind> <Ind>Azerbaijan</Ind> </Atom> </Assert>
The inconsistency can be detected by the query “Is there any inconsistency?”. It succeeds because, according to the rule with falsity, Ilham cannot hold both the post of a president and a prime minister of Azerbaijan.
To remove inconsistency from the rulebase, the fact “Ilham is the prime minister of Azerbaijan” is retracted.
The consistency of the rulebase is checked again by the same query. This time, the query fails and the rulebase regains consistency due to the retraction.
Universally quantified equations are useful for expressing more complex formulas. The following rule “Two persons have family ties if they are either married to each other, or there is a parent-child relationship between them, or they are distant relatives” is formalized using the binary relations hasFamilyTies, marriedTo, hasChild, and hasDistantRelative. The facts “Ilham is married to Mehriban”, “Ilham has a daughter Arzu”, and “Ilham has a daughter Leyla” are formalized using the relations.
<Assert> <Forall> <Var>x</Var> <Var>y</Var> <Implies> <if> <Or> <Atom> <Rel>marriedTo</Rel> <Var>x</Var> <Var>y</Var> </Atom> <Atom> <Rel>hasChild</Rel> <Var>x</Var> <Var>y</Var> </Atom> <Atom> <Rel>hasDistantRelative</Rel> <Var>x</Var> <Var>y</Var> </Atom> </Or> </if> <then> <Atom>
The query, formalizing “List the individuals having family ties to Ilham? ”, results in a y binding to three individuals named Mehriban, Leyla, and Arzu. <Query> <Atom> <Rel>hasFamilyTies</Rel> <Ind>Ilham</Ind> <Var>y</Var> </Atom> </Query>
On the other hand, the query “Which individuals are tied to Mehriban as family members?” can be formalized to retrieve Ilham using the marriedTo relation. Similarly, the binding of x to Ilham establishes family ties for both Leyla and Arzu because of the hasChild relation.
<Query> <Atom> <Rel>hasFamilyTies</Rel> <Var>x</Var> <Ind>Mehriban</Ind> </Atom> </Query> <Query> <Atom> <Rel>hasFamilyTies</Rel> <Var>x</Var> <Ind>Leyla</Ind> </Atom> </Query> <Query> <Atom> <Rel>hasFamilyTies</Rel> <Var>x</Var> <Ind>Arzu</Ind> </Atom> </Query>
The query “Find al l individuals having family ties to each other ”, results in multiple combinations of x and y where Ilham binds to an x and each of Mehriban, Leyla, and Arzu binds to a y, respectively.
The rule “Everything which operates either as an intermediary company or as an offshore company is designated as a company” is formalized as a universally quantified rule with disjunctions in the rule body and it asserts a subclass relationship. Three ground facts “ArborInvestmentsLtd operates as an intermediary company”, “NaziqAndPartners operates as an intermediary company ”, and “HarvardManagementLtd operates as an offshore company” are also formalized and added below.
The above rule implies that ArborInvestmentsLtd, NaziqAndPartners, and HarvardManagementLtd are all designated as companies. The formalization of the query “Which entities are designated as companies? ” results in ArborInvestmentsLtd, NaziqAndPartners, and HarvardManagementLtd as companies binding to the variable x.
From the above assertions, we can also find out if a particular company operates as an intermediary company or as an offshore company. For instance, the query “Does ArborInvestmentsLtd operate as an intermediary company?” succeeds as the company asserts as an intermediary company.
On the other hand, the query “Does ArborInvestmentsLtd operate as an offshore company?” fails as its operation does not assert as an offshore company but as an intermediary company.
Another integrity rule with falsity in rule head is formalized as “Nothing is both an onshore company and an offshore company”.
<Query> <Atom> <Rel>offshoreCompany</Rel> <Ind>ArborInvestmentsLtd</Ind> </Atom> </Query> <Assert> <Forall> <Var>x</Var> <Implies> <if> <And> <Atom> <Rel>onshoreCompany</Rel> <Var>x</Var> </Atom> <Atom> <Rel>offshoreCompany</Rel> <Var>x</Var> </Atom> </And> </if> <then>
<Or/> </then> </Implies> </Forall> </Assert>
The fact “HarvardManagementLtd is an onshore company” creates an inconsistency.
Inconsistencies can be detected by the query “Is there any inconsistency?”.
It succeeds because, according to the rule above, HarvardManagementLtd cannot be both an onshore company and an offshore company.
To remove inconsistency from the rulebase, the fact “HarvardManagementLtd is an onshore company” is retracted.
When consistency of the rulebase is checked by the same query, it fails as the rulebase became consistent with the retraction.
The rule “It is unlikely that an offshore company reveals their clients’ investment information” is formalized as an example of a negative integrity constraint. <Assert> <Forall> <Var>x</Var> <Implies> <if> <And> <Atom> <Rel>offshoreCompany</Rel> <Var>x</Var> </Atom> <Atom> <Rel>revealInvestmentInfo</Rel> <Var>x</Var> </Atom> </And> </if> <then>
<Or/> </then> </Implies> </Forall> </Assert>
The fact “HarvardManagementLtd reveals investment information” creates inconsistency when added to the rulebase.
The query “Is there any inconsistency?” succeeds because HarvardManagementLtd does not reveal investment information while operating as an offshore company.
The fact “HarvardManagementLtd reveals investment information” is retracted to remove inconsistency from the rulebase.
Now that the consistency is checked again by the same query, it fails.
The following assertion contains a rule and a ground fact. The rule “Anyone is a shareholder of a company if and only if there exists a certain level of stocks of that company such as Smal l-cap level or Medium-cap level or Large-cap level, which he owns” is formalized as an equivalence formula. The ground fact “Arzu owns Medium-cap level of stocks in ArborInvestmentsLtd” is formalized and added below.
<Assert> <Forall> <Var>x</Var> <Var>y</Var> <Equivalent> <Atom> <Rel>shareHolderOf</Rel> <Var>x</Var> <Var>y</Var> </Atom> <Exists> <Var>z</Var> <Atom> <Rel>ownsStockin-atLevel</Rel> <Var>x</Var> <Var>y</Var> <Var>z</Var> </Atom> </Exists> </Equivalent> </Forall> <Atom> <Rel>ownsStockin-atLevel</Rel> <Ind>Arzu</Ind>
The benefit of this equivalence formula is that the ownsStockin-atLevel relation can be used when the level of stock ownership is important, otherwise shareHolder relation is sufficient as it does not care about the level of stocks to become a shareholder.
The ownsStockin-atLevel relation is defined as a functional relation, hence together with its existential formalization in the equivalence formula above, it must have uniqueness. This is achieved by formalizing the rule “Every stock owner owns at most one level of stocks of a company” with equality in the head.
The query “Is there any individual who owns Smal l-cap level of stocks in ArborInvestmentsLtd?” fails because there are no individuals binding to an x who own Small-cap level of stocks. A similar query with a different quantity Large-cap in place of Small-cap also fails.
On the other hand, the query “Is there any individual who owns Medium-cap level of stocks in ArborInvestmentsLtd?” succeeds by providing the answer Arzu as the owner binding to an x.
<Query> <Atom> <Rel>ownsStockin-atLevel</Rel> <Var>x</Var> <Var>ArborInvestmentsLtd</Var> <Var>Medium-cap</Var> </Atom> </Query> <Query> <Atom> <Rel>shareHolderOf</Rel> <Var>x</Var> <Var>ArborInvestmentsLtd</Var> </Atom> </Query>
The query “List al l shareholders of ArborInvestmentsLtd” provides the individual Arzu as the only shareholder binding to an x.
The rule “There is a link between a person and a company if the person is a director or a shareholder of the company or he owns a level of stock in the company” is formalized as a universally quantified rule with disjunctions in rule body.
The query “Which individuals and companies are linked to each other?” results in the name Arzu binding to the variable x and the company ArborInvestmentsLtd to a y.
The rule “If a company C1 manages assets of another company C2 and the latter i.e. C2 manages assets of yet another company C3, then company C1 manages the assets of the company C3” is formalized as a universally quantified rule implying a transitive relation. With this rule, relations between two distant companies can be established even if there is a chain of companies sitting in the middle managing assets. The facts “ArborInvestmentsLtd manages assets of NaziqAndPartners” and “NaziqAndPartners manages assets of HarvardManagementLtd” are formalized using the relation manageAssets and are added at the end of the assertion.
From the transitive rule and two facts mentioned above, it can be deduced that the company ArborInvestmentsLtd is, in fact, managing assets of the distant company HarvardManagementLtd. The query “Find al l the companies whose assets are being managed and the companies managing them” can be used to infer such relations among companies.
Finally, by formalizing the rule “The president of a country (possibly) has offshore investments in a company if his family members have links to companies managing assets in that offshore company”, it can be discovered if a president has invested in an offshore company to hide his assets using someone in his family.
Based on the facts and rules provided, the query “Find the presidents and companies in which the presidents possibly have offshore investments” is formalized to find the presidents who have invested in offshore companies to hide their assets from the concerned national and international authority.
<Query> <Atom> <Rel>possiblyHasOffshoreInvestments</Rel> <Var>p</Var> <Var>oc</Var> </Atom> </Query> </RuleML>
The answer to the above query is president Ilham of Azerbaijan whose daughter Arzu has links with the company ArborInvestmentsLtd, which indirectly manages the assets in the offshore company called HarvardManagementLtd. 3
A rulebase formulated using Datalog+ RuleML 1.01/XML rules is presented in this article. The incremental formalization of rules demonstrates how interesting relationships can be discovered among different entities from financial records, which are difficult to deduce in plain sight. The discovery of relationships is used for offshore holdings analytics.
Our rulebase is comprised of expressive RuleML 1.01 rules which, in addition to the existing RuleML sublanguages, cover two newly introduced anchor sublanguages called datalogplus min and disdatalogplus min. Thus our rules not only include positive facts and universally quantified implications, but also allows existential variables, “Equal” predicates, and falsity in the heads of implications. Moreover, conjunctions within existetials and non-empty disjunctions, arbitrarily nested with existentials and conjunctions are permitted in the heads of implications. Some of the rules in our rulebase have disjunctions both in the body and head of implications.
With the addition of new features from the two sublanguages, expressive
rules are now easier to formulate in RuleML 1.01. The ability to use existential
variables in the heads of implications lets us make some progress in solving old
problems in semantic technologies. For instance, an earlier attempt mentioned in
[