Business process compliance has become a crucial aspect for companies due to severe nes that can be imposed if constraints and rules emerging from regulatory documents are violated. Regulatory documents are often written in natural language and analyzing them is mainly done manually since only limited tool support is available. Therefore, we present RegMiner, a web service for discovering and visualizing constraints from regulatory documents. By employing NLP and data mining techniques, compliance constraints can be automatically extracted, grouped, and visualized leading to a separation of relevant and nonrelevant document parts and insights into, e.g., duties of stakeholders. A case study based on a current document from the European parliament regarding the nancial domain demonstrates RegMiner's maturity.
Analyzing regulatory documents as well as assessing the compliance of business
processes with regulations poses a tremendous challenge on companies and is
still mostly done manually though the amount of regulatory documents is
constantly increasing. Compliance violations, in turn, can cause severe problems,
e.g., more than 20 million euros in case of violating the General Data Protection
Regulation (GDPR)1. Hence, providing support for digitalized compliance
management is crucial. Yet, only few tools for extracting, analyzing and visualizing
compliance constraints have been proposed. Most tools rather focus on a one
or bidirectional mapping of natural language texts to imperative or declarative
process models [
RegMiner can be accessed via
http://regminer.wst.univie.ac.at/ A tutorial as well as example data sets are available at
http://gruppe.wst.univie.ac.at/projects/RegMiner/index.php?t=prototypes A screencast is available at
http://gruppe.wst.univie.ac.at/~gallm6/RegMiner/Video/
RegMiner is based on the command line prototypes of previous publications
[
The presentation tier consists Presentation Tier of a single page application rseigcneaivlweoursdesr,ignrpouutp(indgoccuhmoiecnet), language, based on HTML markup and pdaisspslauyserersinupltustretoceloivgeicdtfireorm logic tier JavaScript components. In particular, we use the Bootstrap itvnoidgoelkdinitbf2oyramtnhdaetdiuo3sn.ejrs.m3.uTsthebfeollporwo-- ppaptrrrnaioegdscsgseedeinsanrsttfNaaoitnLrtimPfoioenacrrmtoaimonandtpiotodonnaeptraneretctseieeivrnetadtivoian Logic Tier cegaxroltNcruuaLpclaPtctocecnocosnotsmrntarsiaptnritnoastinnstegnrtaph Document. The input can be provided in two ways. Either as Document DAO: ID, name, [paragraphIDs]Data Tier ZIP le which can contain a) PCaornafiggruarpahtiDonAOD:AIDO,:nIDa,mger,o[uspeinntgeonpcteiosn], language, signalwordsID, userdefinedtermsID one document, b) a set of docu- FGirlaepDhADOA: OID:,IfiDl,ednoacmuem,ecnotnIDte,ngtra(cpohntaesnJtScOaNn be list of keywords or user-defined terms) ments, or c) a set of paragraphs.
Option c) represents a parti- Fig. 1: RegMiner { Architecture
tioning of one or several
documents, e.g., based on the document structure into its sections (cf. [
As a second option, we decided to integrate support for the EUR-Lex platform4 which contains an extensive collection of legal documents such as the General Data Protection Legislation a ecting stakeholder of various domains. The user can provide an URL referencing the HTML markup of the desired EUR-Lex document. The document is downloaded and automatically split into sections based on HTML tags and attributes.
Language. English and German are supported as document languages.
Signal Words. At least one signal word to identify constraints, e.g., \shall",
\should" or \must" (cf. [
Grouping Options. Three options for grouping constraints are available Clustering Constraints are clustered based on term frequency using k-means++; in this case the user needs to specify the number of clusters.
Subject determined by Sentence Structure The subject of each constraint is identi ed based on NLP tags determined by a NLP parser. Per subject one group is created.5 This option is fully automatic, i.e., requires no further user input. User-de ned Terms Constraints are grouped based on terms de ned by the user which need to be uploaded as .txt le. Each term de nes one group, e.g., if \authority" and \user" are contained in the .txt le, all constraints containing \authority" are assigned to one group, all constraints containing \user" are assigned to another group. If neither of them is present, the constraint is shifted to group \unde ned", if both of them are present, the constraint will be contained in the \authority" as well as \user" group.
On submission, the information is passed to the logic tier. As soon as the results were retrieved, a graph consisting of several dots accumulated in clusters is displayed. Each dot represents one constraint and the color indicates the corresponding cluster. By hovering over the dots the constraint is shown. By double-clicking onto a dot the paragraph containing the constraint is displayed. 2.2
The logic tier is written in Python 3 and consists of two components. The rst
one processes incoming and outgoing requests from and to the presentation tier,
i.e., delivering the single page application to the client's browser, handling user
input and returning constraint graphs using the JSON le format and is based
on the web application framework Flask6. The second component, called NLP
component, conducts the actual constraint extraction and grouping procedure. It
is based on the NLP framework SpaCy [
For storing and retrieving data RegMiner uses the MongoDB7 database. Each document uploaded by the user, consists of a name and a list referencing the corresponding paragraphs, stored in separate les. Each paragraph consists of a name and content, i.e., the sentences within the paragraph. Signal words as well as the user-de ned terms are stored in separate les. Furthermore, a conguration le holding the grouping choice, the language and a reference to the signal words resp. user-de ned terms le is created. Information stored within the database becomes identi able via a unique hash-value. This has two bene ts. First, if a document having the same hash value already exists, it is not stored in the database multiple times. Secondly, as the results of the NLP component are also stored in the database, recomputing can be avoided. This enhances the usability as results can be displayed immediately. The latter is important when analyzing extensive documents, which can take up to several minutes. 3
The underlying concepts of the NLP
component have proven to yield
valuable insights into regulatory documents
from various domains, such as security,
the medical domain, nancial domain
and generally applicable regulatory
documents like the GDPR [
8 https://eur-lex.europa.eu/legal-content/EN/TXT/HTML/?uri=CELEX: 32020D0701&qid=1591613404669&from=DE assistance and Commission form the largest groups (black and green). By further examining the underlying constraints by hovering over the dots, the actions and duties for the commission are immediately apparent and can be easily summarized. A user interested under which conditions macro- nancial assistance applies or which aims it has, can directly gain this information based on the graph. If a user needs the context of the constraint, the corresponding paragraph can be displayed by double-clicking onto the constraint. Thereby, a user can directly retrieve the right position within the document. The ability to retrieve such a graph out-of-the-box for an arbitrary and recent document proves the applicability of RegMiner. Finally, RegMiner is tested and discussed with stakeholders from the nancial domain through transfer project RegMiner9.
As future work we plan to improve the pre-processing of documents, provide support for document formats such das PDF, integrate further visualization concepts, especially for complex and extensive documents and investigate domain speci c language models as well as the integration of ontologies.
This work has been funded by the Vienna Science and Technology Fund (WWTF) through project NXT19-003. 9 https://www.wwtf.at/programmes/new_exciting_transfer_projects/NXT19-003