=Paper=
{{Paper
|id=Vol-1963/paper485
|storemode=property
|title=Using Lightweight Semantic Models to Assist Risk Management in a Large Enterprise
|pdfUrl=https://ceur-ws.org/Vol-1963/paper485.pdf
|volume=Vol-1963
|authors=Shirin Sohrabi,Anton Riabov,Octavian Udrea,Fang Yuan
|dblpUrl=https://dblp.org/rec/conf/semweb/SohrabiRUY17
}}
==Using Lightweight Semantic Models to Assist Risk Management in a Large Enterprise==
https://ceur-ws.org/Vol-1963/paper485.pdf
Using Lightweight Semantic Models to Assist
Risk Management in a Large Enterprise
Shirin Sohrabi, Anton Riabov, Octavian Udrea, and Fang Yuan
IBM T.J. Watson Research Center
1 Introduction
In this paper we summarize our experience and the initial results from imple-
menting and operating IBM Scenario Planning Advisor (SPA), a decision sup-
port system that uses lightweight semantic models to assist finance organizations
in identifying and managing emerging risk, a category of risk associated with the
changes in the global or local economies, politics, technology, society, and others.
SPA is designed to support the business process called “scenario planning”
[3] that consists of preparing several future scenarios, followed by identifying
the implications for the business, and finally choosing the mitigation actions to
be taken. For example, prior to the Brexit referendum in 2016, an international
company operating in the UK could consider alternative future scenarios for
changes in trade and employment treaties assuming the majority voted to leave
the EU, identifying the implications for the company’s finances and its ability to
hire, enabling the company to act immediately to minimize the negative impacts.
The main functions of SPA are: 1) discovering active risk drivers by aggre-
gating relevant news from the Web and social media, and generating lists of
candidate observations corresponding to the detected risk drivers; 2) generat-
ing multiple alternative future scenarios highlighting their business implications
and leading indicators, based on user-selected observations, and using domain
knowledge about driver relations, cascading effects, and implications.
A key design decision in SPA is that the system does not compute probabili-
ties for the generated scenarios. Instead, we recommend that the domain experts
assign probabilities to the final 3-5 scenarios when necessary. In our experience
that approach delivers value to the users much faster and requires significantly
less work from the experts, compared to creating prediction models for over 200
often non-stationary risk drivers (e.g., oil prices or election results), and their
interactions, as would be required to derive scenario probabilities automatically.
We have identified the following major challenges in developing SPA: 1) cap-
turing observations from news and social medial; 2) capturing the domain knowl-
edge about the risk drivers quickly and efficiently, while preventing conflicts;
3) reasoning with incomplete and biased input to include sufficiently complete
and minimally biased sets of risks and opportunities in generated scenarios.
In the rest of the paper, we describe the components of the SPA system,
explaining their role in addressing the challenges, and highlighting the use of
semantic technologies throughout the SPA system. We then present some details
of the deployment that indicate our initial success in overcoming the challenges.
2 Domain Knowledge Workflow
Increasing debt levels Trapped cash Acme to
Challenging environment for Contoso
identify opportunities to invest & expand
Currency
High inflation depreciation against Acme workforce capital available at better rates
Contoso
US dollar
Increasing trade deficit Lower domestic demand Economic decline
Domain Experts Mind Maps
Observations
Concepts
AI ddd
Local Experts Planning Scenarios
Customizations
News High
Aggregator Inflation
Weakening
Economy
Scenario Planning
Social Media, Teams
RSS/Atom WIKIDATA Observations
2 Semantic Components in SPA
The above figure shows the interactions between SPA components and its users. 2
Capturing Observations. The News Aggregator component aggregates
news from RSS and Atom feeds and social media posts, e.g., Twitter, in mul-
tiple languages, by monitoring user-configured keywords for each candidate ob-
servation, for each country. To further refine and filter the information, News
Aggregator uses the structured semantic knowledge available in Wikidata. Coun-
try relevance is determined based on the mentions of the local people and or-
ganizations found using Wikidata Query Service [7]. News Aggregator also uses
Wikidata for source discovery. The end-users (local experts) then choose relevant
observations for scenario planning.
Capturing Domain Knowledge. The Mind Maps and The Customiza-
tions components store knowledge about risk drivers and business implications
elicited from the domain experts and the local country experts correspondingly.
While the reasoning engine in SPA supports a rich representation of risk drivers
as actions in Planning Domain Description Language (PDDL) [2], the knowl-
edge representation used by domain experts is drastically simplified, to prevent
conflicts and reduce overheads in knowledge elicitation and maintenance. The do-
main experts use Mind Maps created in FreeMind [1] to capture directed graphs
of risk drivers and business implications, with edges having hidden semantics of
pairwise cause and effect. The Customizations are elicited using generated ques-
tionnaires that request country-specific likelihood and impact for selected cause
and effect pairs. Due to Customizations, the same observations may generate
different scenarios in different countries.
Reasoning With Incomplete Knowledge. The AI Planning component
applies plan recognition and top-k planning techniques to reason with incomplete
knowledge and generate scenarios [4–6]. The scenarios are clusters of high-quality
plans that include a trajectory of cause-effect transitions from the Mind Maps,
explaining the largest possible subset of observations (rather than achieving a
predefined goal), and such that each plan ends with a business implication.
The scenarios, presented as generated text summaries and graphically, are then
reviewed and refined by scenario planning teams.
3 SPA Deployment
The deployed system has over 30 active teams of users, 700 scenarios generated,
and is processing over 50,000 social media messages per hour. There are 230 risk
drivers and business implications and 382 edges between them in the lightweight
Mind Map representation of the domain knowledge used for scenario generation.
3
4 Acknowledgements
We thank Fang Yuan and Finn McCoole at IBM for providing the domain exper-
tise. We thank Nagui Halim and Edward Shay for their guidance and support. We
also thank our LAS collaborators. This material is based upon work supported in
whole or in part with funding from the Laboratory for Analytic Sciences (LAS).
Any opinions, findings, conclusions, or recommendations expressed in this mate-
rial are those of the authors and do not necessarily reflect the views of the LAS
and/or any agency or entity of the United States Government.
References
1. FreeMind: http://freemind.sourceforge.net
2. McDermott, D.V.: PDDL — The Planning Domain Definition Language. Tech.
Rep. TR-98-003/DCS TR-1165, Yale Center for Computational Vision and Con-
trol (1998)
3. Peterson, G.D., Cumming, G.S., Carpenter, S.R.: Scenario planning: a tool for con-
servation in an uncertain world. Conservation biology 17(2), 358–366 (2003)
4. Sohrabi, S., Riabov, A., Udrea, O.: Plan recognition as planning revisited. In: Pro-
ceedings of the International Joint Conference on Artificial Intelligence (IJCAI)
(2016)
5. Sohrabi, S., Riabov, A., Udrea, O.: State projection via AI planning. In: Proceedings
of National Conference on Artificial Intelligence (AAAI) (2017)
6. Sohrabi, S., Riabov, A., Udrea, O., Hassanzadeh, O.: Finding diverse high-quality
plans for hypothesis generation. In: Proceedings of the 22nd European Conference
on Artificial Intelligence (ECAI) (2016)
7. Wikidata: https://query.wikidata.org/