![](bg1.png)
User-adaptable Natural Language Generation for Regression Testing
within the Finance Domain
Daniel Braun
a
, Anupama Sajwan and Florian Matthes
Technical University of Munich, Department of Informatics, Munich, Germany
Keywords:
Natural Language Generation, Regression Testing, Finance.
Abstract:
Reporting duties and regression testing within the financial industry produce huge amounts of data which has
to be sighted and analyzed by experts. This time-consuming and expensive process does not fit to modern,
agile software developing practices with fast update cycles. In this paper, we present a user-adaptable natural
language generation system that supports financial experts from the insurance industry in analysing the results
from regression tests for Solvency II risk calculations and evaluate it with a group of experts.
1 INTRODUCTION
Companies within the finance industry, like banks and
insurance companies, have to fulfil many regulatory
requirements. Some of the most prominent directives
within the European Union (EU) include “Basel III”
for banks and “Solvency II” for insurance companies.
In order to fulfil the requirements introduced by these
legislations, companies have to continuously report
risk relevant corporate results and investments to their
respective regulatory authority. These reports deter-
mine how much money companies have to put aside
as a security.
From a company’s perspective, it is desirable to
keep this amount as low as possible, because they
only can create profit from money which they can
actively invest. Therefore, big insurance companies
use tailored internal risk models instead of the stan-
dard risk model provided by Solvency II. The soft-
ware which runs these internal risk models has to be
updated regularly, in order to meet the regulatory re-
quirements and the company’s interests. Before a new
version of such a software is put into production use,
regression testing is used to ensure proper behaviour.
A single run of such regression tests produces thou-
sand of numbers which have to be compared to pre-
vious results and interpret by financial experts, which
then have to report back to developers. This is a cost
and time-intensive process which also hinders com-
panies to deploy updates more often.
In this paper, we present a natural language gener-
a
https://orcid.org/0000-0001-8120-3368
ation (NLG) system which creates textual reports for
the results from regression tests for Solvency II risk
capital calculations. By identifying and highlighting
salient patterns within the results, we want to support
the work of financial experts and speed up the process.
Moreover, the system is built in a way which aims to
empower expert users, which are non-programmers,
to adapt the system regarding the analysis which is
conducted but also regarding the textual representa-
tion of the results of the analysis. The system was
designed and evaluated with financial experts from a
major international insurance company.
2 RELATED WORK
Most transactions on the international financial mar-
kets are nowadays not only executed but also trig-
gered by machines. (Banulescu and Colletaz, 2013)
Therefore, the relevant data is available in a machine-
readable format and, due to the nature of the domain,
mostly numerical. Given these circumstances, it is no
surprise that the finance domain is of great interest to
the NLG community, from a scientific and a commer-
cial perspective.
One of the most prominent applications of NLG
within the finance domain today is robot journalism.
Together with weather, traffic, and sports, finance is
one of the most popular domains for robot journal-
ism. (D
¨
orr, 2016) Examples for such systems were
build by Kukich (1983), and more recently Liu et al.
(2004), Haarmann and Sikorski (2015), Murakami
Braun, D., Sajwan, A. and Matthes, F.
User-adaptable Natural Language Generation for Regression Testing within the Finance Domain.
DOI: 10.5220/0009563306130618
In Proceedings of the 22nd International Conference on Enterprise Information Systems (ICEIS 2020) - Volume 1, pages 613-618
ISBN: 978-989-758-423-7
Copyright
c
2020 by SCITEPRESS – Science and Technology Publications, Lda. All rights reserved
613