This paper presents a visualization technique to assist legal experts in formalising their interpretation of legal texts in terms of regulatory requirements. (Semi-)automation of compliance processes requires a machine-readable version of legal requirements in a format that enables e ective compliance assessment. The use of a semi-structured controlled natural language as an intermediate step of the translation from a human-readable text to a machine-readable and understandable format ensures that the process of interpretation of those requirements is as simple as possible. However, it does not ensure that the formal representation resulting from the interpretation faithfully represents the intended semantics provided by the legal expert. Visualization techniques such as property graphs in Neo4j could ll this gap, allowing legal experts to understand and control the formal representation of the result of their act of interpretation.
Ensuring compliance with regulatory requirements represents a considerable
challenge for industries that deal with large amounts of data across di erent
jurisdictions. This is particularly true for safety-critical industries such as the
international nancial industry, driven by the proliferation and complexity of
the nancial regulatory environment in the aftermath of the global nancial
crisis. The wide acceptance in the industry that traditional Governance, Risk, and
Compliance (GRC) information systems are in de cit is leading to a growing
interest in semantic technologies as a solution [
One of the problems with representing the interpretation of legal rules in
a machine-readable format lies in the lack of understanding and control of the
formal models by legal experts. Some research approaches [
The research behind the present paper follows an approach that allows legal experts to interpret a legal statement by rewriting it in a semi-structured controlled natural language using a dedicated software called SmaRT. The core translation process relies on the markup and annotation of strings in the rewritten text in terms of given vocabulary elements. However, this is not su cient to put legal experts in control of all the elements that compose the logical formulation of the legal rule. The paper thus presents a possible solution that relies on property graphs to visualize the interpreted rule and all the relevant elements in a format that a lawyer can understand and manipulate. This should allow to ensure that the human-readable text and the machine-readable output of a legal rule carry the same semantics, thus achieving a semi-automatic translation of legal requirements for compliance purposes.
The rest of the paper is structured as follows. In the next section, we introduce the Regulatory Interpretation Methodology (RIM) and the controlled natural language. In Section 3, we look into relevant details of the RIM and the rule-editing tool SmaRT to understand the visualisation needs for the rule interpretation task. In Section 4, we present a possible visualisation solution that addresses these needs. 2
Understanding regulations is a complex task and legal experts face a number of
challenges in interpreting a regulatory text, including: following and eshing out
references and citations; identifying, delimiting, and disambiguating de nitions;
making sense of complex sentences; clarifying ambiguities resulting from legalese;
accounting for exceptions [
The translation of regulatory text into machine-readable information is
articulated around a methodology that de nes a process for transforming a regulatory
text into a formal representation using an intermediate human-readable
representation in semi-structured English. The process (see Fig. 1) is designed as a
collaborative one, involving the legal expert as a subject matter expert (SME)
and the modeller as a semantic technology expert (STE) through multiple
iterations [
The solution allows SMEs to represent the semantics of regulatory
requirements in a machine-readable format through a SME-friendly process. This is
ensured through the use of SBVR (Semantic of Business Vocabularies and Business
Rules), a Object Management Group speci cation based on formal logics and
well known to the industry [
To overcome this, our research group at the Governance, Risk, and
Compliance Technology Center (GRCTC) has devised a resource called Mercury,
composed of:
{ Mercury-SE, a semi-structured English based on SBVR used as an
intermediate language between the legal text and the machine-readable format,
and
{ Mercury-ML, a persistence model in RDF format, capturing the semantics
of Mercury-SE in a machine-readable format.
Mercury represents rule statements contained in regulations and describes the
concepts used in those rules in a vocabulary. The process of translation from the
legal text to the Mercury-SE language is manual and tool-assisted. The software,
SmaRT, is speci cally designed to make the translation process as intuitive as
possible, while at the same time reducing the user's time devoted to the most
repetitive work. The reader is invited to refer to [
Mercury relies on both (a) technologies from the Semantic Web (SW) stack
at the RDF layer and (b) non-SW technologies (SBVR and the RIM). It relies on
upper SW layers, particularly OWL, for advanced classi cation and reasoning
on rules and vocabulary. The GRCTC is currently developing a set of
ontologies called FIRO (Financial Industry Regulatory Ontology) to enable semantic
applications such as classi cation, querying, and reasoning, and has devised a
mapping of all SBVR elements relevant to Mercury into RDF/OWL to assist
the STE in translating Mercury rulebooks and vocabularies [
Because of limitation of OWL and Description Logics [
In order to perform these tasks, we need the legal rules { represented in Mercury-ML { to be complete in terms of semantics and computable. Mercury being a semi-structured controlled natural language, there is a need for solutions that allow SMEs to have greater control over the formal representation of their interpretation of a legal rule without having to learn any speci c formalism. We propose to achieve this by providing a graphical solution that allows the SMEs to visualise rules in a more schematic way. For example, Fig. 2 shows a graphical representation of the following rule: Rule 1. It is obligatory that a market operator that operates a trading venue makes public credits and debts.
The graph in Fig. 2 makes explicit the information about the logical formulation of the rule that is not immediately apparent in the textual form. For example, it shows that market operator is the subject of three verb concepts: market operator operates trading venue, market operator makes public credit, and market operator makes public debt. It also shows which verb concepts express the deontic condition of the rule { see the dashed box in red { distinguishing it from the other verb concept, which expresses the rule's applicability condition. Finally, it shows how the former and the latter verb concepts relate to each other { they are connected via the same subject, market operator.
In the next section, we describe the tasks of the Regulatory Interpretation Methodology in more detail to specify further our visualization needs. 3
To create a rule from a legal document, SMEs follow the methodology de ned
in the RIM. First, they identify a unit of analysis within a given legal document.
Second, they rewrite the text of the unit of analysis, eshing out references.
Third, they identify and de ne the noun concepts (i.e., entities) and verb
concepts (i.e., actions with their participants and other attributes) and mark the
keywords (e.g., the logical operators) composing the rule. Finally, they specify
the logical formulation of the rule by linking verb concepts to each other. This
logical formulation follows speci c patterns for constitutive rules [
All these steps are carried out using SmaRT, our web-based rule editing tool. Fig. 3 shows the editor window at the rst step of the methodology. The original regulatory text is displayed in the upper box and the text rewritten by the SME is displayed in the lower box. 3.1
The task of creating verb concepts consists of identifying all the noun concepts { usually domain-speci c terms { and linking them to the appropriate verbs or verb phrases using one of the roles that specify the semantic relation between a noun concept and the action denoted by the verb, such as hasSubject, hasObject, hasIndirectObject, and hasLocation. A verb concept is created when all the relevant noun concepts are linked to the verb concept's verb with one (and only one) of these relations. The verb of a verb concept should be captured in its active form and should account for any of its variant forms, e.g., its passive form. Therefore, the rst visualisation requirement to assist the SMEs in the task of creating verb concepts is to display (i) the building blocks of a verb concept { the verb and the noun concepts { and (ii) the links representing the roles relating noun concepts to the verb in a verb concept. 3.2
The second task of connecting verb concepts to each other to form the rules brings about further requirements for useful visualisation aids. This task consists of two subtasks: { Identifying the logical operators (and /or ) forming a conjunction or disjunction of two or more noun concepts that play the same role with respect to a verb concept. This results in the creation of two or more verb concepts denoting two or more actions with their participants and other attributes (see Fig. 4). { Identifying noun concepts that play a role in more than one verb concept (see Fig. 5).
Therefore, the second visualisation requirement to assist the SMEs in the task of linking verb concepts to each other is to display these larger building blocks and their links in a user-friendly manner that captures and gives a clear understanding of the logical structure of a rule (see Fig. 6). 3.3
To help the SMEs in these tasks, the current version of SmaRT integrates display functions that address only part of the visualisation needs. The tool allows SMEs to visualise the di erent elements of a verb concept as shown in Fig. 7. These are displayed using color-coded tags that encapsulate the respective text spans using distinct colors for noun concepts (in green) and verbs { verb parts of speech and verbal phrases { (in blue).
Selecting a verb highlights the entire verb concept, that is, the verb and the related noun concepts. This is illustrated in Fig. 8, where the verb operates in the verb concept market operator operates trading venue is marked in blue and the related noun concepts, market operator and trading venue, in red. However, the current solution is limited in that the roles played by each noun concept within a verb concept are not graphically represented in the main window. They are speci ed and displayed in a separate pane (see the text box on the right-hand side of the editor in Fig. 8).
Similarly, the current implementation does not support the visualisation of more complex relations between verb concepts resulting from conjunctions or disjunctions of noun concepts { \make public credits AND debts", nor of how two or more verb concepts are linked to each other { market operator is the subject of makes public and of operates and thus links the corresponding verb concepts.
In sum, the visualisation needs for a seamless execution of these tasks can be broadly categorized into two types: grouping textual spans at di erent levels of granularity and linking these spans to each other in a user-friendly manner that clari es and makes the logical structure of a rule explicit. Ideally, the rule-editing tool would allow the SMEs not only to visualise the rules, but also to interact with the visual environment to build the rules. Finally, the visualisation solution should also address end-user needs by providing them with additional query functions that would allow them to retrieve and display relevant rules for enhanced regulatory compliance veri cation. To address the current shortcomings of the tool and improve its user-friendliness, we have explored a visualisation solution that would meet these needs and assist the SMEs in the rule-editing task. 4
Graphs are helpful to visualise relationships. In this section, we present a method to represent vocabularies and legal rules in property graphs and store them in the graph database management system Neo4j1 that allows for native graph storage and processing. The property graph contains connected entities (the nodes) which can hold any number of attributes (key-value pairs). Nodes can be tagged with labels representing their di erent roles in a domain. In addition to contextualizing node and relationship properties, labels may also serve to attach meta data, index, or constraint information to certain nodes.
Neo4j uses Cypher2, a declarative, SQL-inspired language for visually describing patterns in graphs using an ascii-art syntax. Here, Cypher is used to represent and query graph data. 4.1
Graphical Representation of Noun Concepts. Noun concepts constitute building blocks of legal rules; they can be nancial concepts. Each noun concept includes attributes, such as label, concept type, context type, and their corresponding values. They also include object properties such as hasGeneralConcept, which indicates a subclass relationship between two noun concepts. We represent each noun concept as a graph node, with key-value pairs specifying its attributes and their corresponding values.
Fig. 9 shows an example of a noun concept, investment rm, with some of its attributes and corresponding values (shown in the bottom bar), and a relationship hasGeneralConcept between investment rm and the noun concept company, indicating that the concept investment rm is a subclass of the concept company.
Graphical Representation of Verb Concepts. Usually, verb concepts denote actions contained in rules and describe basic relationships between noun concepts. Each verb concept includes a subject and a verb or verbal phrase called verbSymbol. It may also include an object, and indirect object or other types of relations such as hasLocation. We represent each verb concept as a graph node that has edges (relationships) such as: hasSubject, hasVerbSymbol, hasObject, and hasIndirectObject. Fig. 10 shows an example of the verb concept market operator operates a trading venue. 1 https://neo4j.com/ 2 https://neo4j.com/developer/cypher-query-language/ 4.2
Legal rules are more complex than noun and verb concepts. Each rule includes one or more applicability conditions that determine whether a given action is relevant to a given rule or not, and one or more deontic conditions that determine whether a relevant action complies with or breaches a rule. In our current implementation, we represent a legal rule as a graph node that is connected to the graph nodes of applicability condition and deontic condition with the edges hasApplicabilityCondition and hasDeonticCondition. The nodes of applicability condition and deontic condition have edges associatedWith connecting them to action (verb concept) nodes.
Fig. 11 shows an example of graphical representation of the rule used throughout the paper (Rule 1). This rule has key-value pairs containing some basic information such as the original regulatory text shown in the bottom bar. The rule1 node connects two deontic condition nodes and one applicability condition node. Each of them is connected to a verb concept node by the edge associatedWith. The rules and vocabularies represented in property graphs can be queried and retrieved by users. Queries over graph databases are often graph patterns. Here, we use the Cypher language of the Neo4j graph database to formulate graph patterns and pose the Cypher queries in the graph database to retrieve graph data. Note that queries can also be visualised since they are graphs with variables. We also formulate some common queries as query templates which can help the users who have no knowledge of Cypher to formulate their queries. 5
This paper presents visualisation needs and enhancements for a rule-editing tool used by legal experts to achieve a machine-readable interpretation of legal requirements with a semi-structured controlled natural language. The paper presented knowledge graphs as a way to visualize an interpreted rule and all its relevant elements in a format that the lawyer can understand and manipulate. This is meant to ensure that the machine-readable output of the regulation interpretation process is semantically enriched as intended by the legal expert, thus ensuring the reliability of the interpretation stored in the machine-readable format. Next steps of the research include further investigation of graphing solutions. We anticipate that this might bring us to reconsider some basic formalization principles derived from SBVR, which in turn might lead us to reconsider some modelling choices in Mercury.