Inferring New Information from a Knowledge Graph in Crisis

Management: A Case Study

Julie Bu Daher

, Tom Huygue

, Nathalie Hernandez

and Patricia Stolf

IRIT, Universit

e de Toulouse, CNRS, Toulouse INP, UT3, Toulouse, France

IRIT, Universit

e de Toulouse, CNRS, Toulouse INP, UT3, UT2, Toulouse, France

Keywords:

Semantic Web, Crisis Management, Knowledge Management, Information Inference.

Abstract:

Natural crises are dangerous events that can threaten lives and lead to severe damages. Crisis-related data

can be heterogeneous and be provided from multiple data sources. These data can be formally described

using ontologies and then integrated and structured forming knowledge graphs. Inferring new information

from knowledge graphs can strongly assist in the various phases of the crisis management process. Different

approaches exist in the literature for inferring new information from knowledge graphs. In this paper, we

present a case study of a ﬂood crisis where we discuss three approaches for inferring ﬂood-related information,

and we experimentally evaluate these approaches using real ﬂood-related data and synthetic data for further

analysis. We discuss the interest of using each of these approaches and detail its advantages as well as its

limitations.

1 INTRODUCTION

Natural crises, such as ﬂoods are adverse events re-

sulting from natural processes of the Earth. They

could lead to severe consequences such as loss of

lives, disruption of normal life of the population and

materialistic damage in properties, infrastructure and

economy. From here comes the urgent need of the cri-

sis management in order to limit these consequences.

Crisis-related data can be exploited for taking impor-

tant decisions that can assist in the crisis management

process. These data are heterogeneous and can be

provided from multiple sources. Managing these data

is important for two main reasons. First, integrating

and structuring the crisis-related data allows the ac-

tors involved in the management process to access

the needed data at the right time. Second, structur-

ing the data allows managing its heterogeneity and

thus inferring new information that enriches the ini-

tial data shared by the actors in real-time during the

crisis. This can be attained using semantic web tech-

nologies that allow the structuring of the data and the

inference of new information from it.

An ontology allows a structuring and a logical rep-

resentation of the knowledge through concepts and

relations among concepts of an ontology. They are

known for managing heterogeneity and having a con-

sistent shared understanding of the meaning of in-

formation (Elmhadhbi et al., 2019). Heterogeneous

crisis-related data can thus be integrated and struc-

tured using the concepts and relations of an ontol-

ogy to form a knowledge graph. This enhances the

interoperability of the data among the various actors

involved in the crisis management process, and it al-

lows inferring new information and making it explicit

which helps in the decision making process of the cri-

sis management.

Inferring new information from the crisis-related

data can help in past, present and future aspects. It

can help in improving a past experience through in-

ferring new information from past crisis-related data.

The situation of a current crisis can also be analyzed,

and we can infer new information that helps taking

current actions. In addition, we can infer new infor-

mation that helps in predicting a future crisis or its

phases and properties.

Crisis management has been described in the lit-

erature as a lifecycle, and it is categorized into four

main phases: mitigation, preparedness, response and

recovery (Franke, 2011). This work is in the frame of

a project that aims at integrating several disciplinary

expertises to limit the consequences of ﬂash ﬂoods. It

is a case study on real ﬂood-related data where the aim

is to propose solutions for limiting the consequences

of a ﬂood during the ﬂood response phase. When a

ﬂood occurs, the safety of the population is the most

Bu Daher, J., Huygue, T., Hernandez, N. and Stolf, P.

Inferring New Information from a Knowledge Graph in Crisis Management: A Case Study.

DOI: 10.5220/0011511200003335

In Proceedings of the 14th International Joint Conference on Knowledge Discovery, Knowledge Engineering and Knowledge Management (IC3K 2022) - Volume 2: KEOD, pages 43-54

ISBN: 978-989-758-614-9; ISSN: 2184-3228

important concern; therefore, an evacuation process

of the population in demand points should take place

where a demand point represents a place that can be

impacted by the ﬂood and thus needs to be evacu-

ated. This process is handled by the ﬁreﬁghters who

are responsible for taking rapid decisions and actions

concerning evacuation. In our case study, we aim at

proposing evacuation priorities to demand points in

order to assist in the evacuation process of ﬂood vic-

tims. In this frame, we have proposed an ontology

that formally describes the ﬂood-related data and, and

we have integrated the heterogeneous data in a knowl-

edge graph using the shared vocabulary of the ontol-

ogy (Bu Daher et al., 2022). Using this knowledge

graph, we aim at inferring new information repre-

senting evacuation priorities to all the demand points

in our study area, and we then aim at enriching the

knowledge graph with this information about prior-

ities and updating it constantly with real-time data.

The aim of this paper is to evaluate three approaches

for inferring evacuation priorities of demand points

through a case study on real data representing a past

occurring ﬂood.

The paper is organized as follows. Section 2 dis-

cusses the related work in this domain. Section 3

presents the problem statement containing our prob-

lematic, data description and the used ontology and

knowledge graph. Section 4 presents the three ap-

proaches that are proposed as solutions for our prob-

lem which are later evaluated in section 5. Finally,

section 6 discusses the conclusion and the future

work.

2 RELATED WORK

Ontology-based approaches have been proposed in

the literature in the domain of crisis management.

The main purpose behind proposing ontologies in this

domain is the information management and sharing

among different actors involved in the crisis manage-

ment process.

An ontology has been proposed by (Katuk et al.,

2009) for integrating ﬂood-related data to allow the

coordination of response activities among different

agencies involved in the management process and to

provide up-to-date information that facilitates the de-

cision making by the management committee chair-

man. Another ontology was proposed by (Yahya

and Ramli, 2020) to formally integrate ﬂood-related

data in order to be shared by all related agencies in

the management process. They propose an ontology

for each agency including one describing data about

evacuation centers, and they then aim at integrating all

the ontologies in a global one that shares information

among all agencies. An ﬂood ontology was proposed

by (Khantong et al., 2020) for managing and sharing

ﬂood information among different responders, orga-

nizers or processes that are handled by different sys-

tems in organizations in order to carry out disaster re-

lief operations. The ontology manages static and dy-

namic data. Static data represent the data that don’t

change during a ﬂood, while dynamic data represent

the data evolving throughout a ﬂood. Their static data

are described through concepts including area and re-

source, and their dynamic data represent coordination

and production acts concerning the crisis.

Some of the proposed ontologies include concepts

related to victims’ evacuation such as victim, ﬂood

as well as evacuation areas, resources and centers

(Khantong et al., 2020; Yahya and Ramli, 2020);

however, these concepts are not exploited for infer-

ring new information that assist in the evacuation pro-

cess.

Although not widely existing, some approaches in

the literature propose inferring various kinds of infor-

mation in this domain. An ontology-based framework

for risk assessment is proposed by (Scheuer et al.,

2013) to manage and share the knowledge of stake-

holders and decision makers in risk management and

to infer information from the ontology concerning el-

ements at risk against certain event types based on the

user’s input. Different types of events such as ﬂoods

are deﬁned in the ontology where the user chooses an

event type and deﬁnes the event intensity in order to

assess elements at risk against this event. The pro-

posed framework identiﬁes the intensity parameters

that are suitable for this ﬂood event using the rela-

tion ”IntensityOf” deﬁned in the ontology between

”Event” and ”Intensity” classes. Then, the frame-

work infers the elements at risk that are susceptible to

this event through matching susceptibility functions

against the chosen event type using the “isSusceptib-

lityTo” relation that is used to link each susceptibility

function to the respective event types. A susceptibil-

ity function takes one or more intensity parameters

as input and allows obtaining a damage ratio. The

relevant elements at risk against the event type are

then inferred by matching susceptibility functions us-

ing the ”susceptibilityOf” relation deﬁned in the on-

tology between ”Susceptibility Function” and ”Ele-

ment AtRisk” classes. (Wang et al., 2018) propose a

hydrological sensor web ontology, based on three ex-

isting ontologies: SOSA

, Time

and GEOSPARQL

https://www.w3.org/TR/vocab-ssn/

https://www.w3.org/TR/owl-time/

https://opengeospatial.github.io/ogc-geosparql/geosp

arql11/index.html

KEOD 2022 - 14th International Conference on Knowledge Engineering and Ontology Development

to integrate heterogeneous data provided from differ-

ent sensors effectively during natural disasters. They

then use SWRL rules on their constructed knowledge

graph to infer ﬂood phases from the precipitation of

water level and observation data. (Kurte et al., 2017)

propose an ontology that captures dynamically evolv-

ing phenomena to understand the dynamic spatio-

temporal behaviour of a ﬂood disaster. ”SIIM” (Kurte

et al., 2016) and ”Time” ontologies were used to de-

scribe geospatial and time concepts, and SWRL rules

were then used to retrieve image regions based on

their temporal interval relations. (Sun et al., 2016)

propose an ontology that allows inferring ﬂood states

as well as their properties, such as precipitation and

water course in the frame of a context-aware system.

Jess rules (Hill, 2003) are used to infer new infor-

mation that enrich their context-aware system so that

its components can be adapted according to context

changes.

We notice from the approaches proposed in the

domain of crisis management that new information

can be inferred using the relations deﬁned among con-

cepts of the ontology (Scheuer et al., 2013) and using

rules such as SWRL rules (Wang et al., 2018; Kurte

et al., 2017). SPARQL queries are used to query

the knowledge graphs in order to extract information

(Wang et al., 2018).

Inferring new information using deﬁned concepts

and relations’ characteristics of the ontology is an ap-

proach that can be conducted using existing tools and

reasoners. SPARQL query language, that is usually

used for querying the knowledge graph, can be also

used for inferring new information from the knowl-

edge graph. Rules have been usually used in the do-

main of crisis management in order to infer new in-

formation. There exists a different kind of rules for

inferring new information from knowledge graphs. In

our work, we propose to evaluate three different ap-

proaches for inferring new information in the domain

of crisis management.

3 PROBLEM STATEMENT

In this section, we discuss our problem statement in-

cluding our problematic, the data representing our

case study, our proposed ontology and knowledge

graph, the evacuation priorities deﬁned for the study

area and the proposed solutions for our problematic.

3.1 Problematic

Flood crisis management is a critical process as it con-

cerns severe consequences where the most adverse

consequence is having victims in danger. Therefore,

the evacuation of victims is an essential process in the

ﬂood response phase. The proposed solutions for as-

sisting in this process should respect the delicacy of

the situation in being rapid and efﬁcient. In our work,

we propose evaluating three different approaches for

inferring information that can assist in the evacua-

tion process of ﬂood’s victims. These approaches are

based on approaches existing in the literature for in-

ferring new information from knowledge graphs. The

evaluated approaches propose inferring the same in-

formation differently. The ﬁrst approach proposes

inferring the information that enrich the knowledge

graph from the concepts and relations of the ontol-

ogy using OWL constructors, the second approach

uses SPARQL query language to infer new informa-

tion through directly enriching the knowledge graph

using insert and delete queries, and the third approach

proposes inferring the information using rules and en-

riching the knowledge graph with the inferred infor-

mation. Rules are frequently used in the literature for

inferring information from knowledge graphs. Dif-

ferent kinds of rules have been used for this purpose

including SWRL rules

. In this approach, we propose

using SHACL rules

, a more recent kind of rules that

overcomes certain limitations of other kinds and that

hasn’t been used in the domain of crisis management

yet.

Each approach relies on the knowledge graph con-

taining the ﬂood-related data to infer evacuation pri-

orities of demand points and to enrich the knowledge

graph with the inferred priorities. We experimentally

evaluate the performance of the three approaches on

a real use case of a ﬂood crisis in order to discuss the

importance of inferring new information that helps in

the crisis management process and to ﬁnd the most

efﬁcient approach applicable to this use case. In addi-

tion, we extend this use case with synthetic data when

needed in order to further analyze the speciﬁcities of

each approach.

3.2 Data Description

The data in our study area concern the Pyr

ees ﬂood

that occured in June 2013 in Bagn

eres-de-Luchon,

south-western France. It was a torrential ﬂood, de-

structive and dangerous for the population. The con-

sequences of this ﬂood include destructed houses,

cut roads, ﬂooded campsites and damaged farms.

The food-related data are heterogeneous and are pro-

vided from various data sources. These sources in-

https://www.w3.org/Submission/SWRL/

https://www.w3.org/TR/shacl-af/

Inferring New Information from a Knowledge Graph in Crisis Management: A Case Study

clude institutional databases such as BD TOPO

and

GeoSirene

providing data about hazards, vulnerabil-

ity, damage and resilience. Some sources provide

data about geographical locations of roads, buildings,

companies and establishments in France. Other data

sources provide various data including data sensors

providing data about water levels and ﬂows, a hy-

drological model computing ﬂood generation, a hy-

draulic model for ﬂood propagation as well as other

sources providing other kinds of data such as socio-

economic and population data as well as danger and

vulnerability indices of the ﬂood calculated by the do-

main experts. The vulnerability index measures the

vulnerability of a demand point, and it is calculated

using topographic and social data like population den-

sity, building quality and socio-economic conditions.

The danger index measures the level of danger of a de-

mand point, and it is calculated using the water speed

and level obtained from a hydraulic model. These

data can be categorized as static and dynamic data.

Our static data include the number of ﬂoors and geo-

graphic locations, while our dynamic data include the

water level and the number of population in a demand

point.

3.3 Ontology and Knowledge Graph

Construction

In a previous work, we have proposed an ontol-

ogy that formally describes our heterogeneous

ﬂood-related data with all the needed concepts

and relations, and we constructed our knowl-

edge graph through integrating the heterogeneous

ﬂood-related data of our study area (Bu Daher

et al., 2022). Our ontology consists of 41 classes,

6 object properties and 23 data properties, and

it is available online via the following URL:

https://www.irit.fr/recherches/MELODI/ontologies/i-

Nondations.owl. It is used in a further step to infer

new information assisting in the ﬂood response phase

of the management process. We present as follows

the concepts and relations that are strictly related

to inferring new information concerning evacuation

priorities.

We deﬁne in our ontology a class named ”de-

mand point” representing either an infrastructure or

an infrastructure aggregation. This class is char-

acterized by four subclasses representing evacuation

priorities that are used in a further step for infer-

ring new information concerning the priorities of de-

mand points. The class ”Material infrastructure” de-

https://www.data.gouv.fr/en/datasets/bd-topo-r/

https://data.laregion.fr/explore/dataset/base-sirene-v3-

ss/

scribes all possible types of infrastructure in the study

area including homes, working places and facilities

such as healthcare facilities. We also deﬁne a class

named ”Infrastructure aggregation” that allows man-

aging different types of infrastructure in an aggre-

gated manner by regrouping them in districts, build-

ings and ﬂoors. For example, we can describe that

the district has buildings, the building has ﬂoors, and

the ﬂoor has apartments using the relation ”has part”.

This class is useful when data about different kinds

of infrastructure are unavailable. For example, when

data about a building are not available, we consider

the data about its district.

The deﬁned class ”population”describes the pop-

ulation in an infrastructure including fragile and non-

fragile population deﬁned using the relations ”is in”

and ”contains”.

The object and data properties are divided into

static and dynamic properties to represent static and

dynamic ﬂood-related data. The static object prop-

erties represent the relations between concepts de-

scribing static data such as ”has part” describing an

infrastructure or an infrastructure aggregation, while

the dynamic object properties represent the relations

between concepts describing dynamic data such as

”contains” that describes the population in an infras-

tructure or an infrastructure aggregation.

Concerning the data properties, the static data

properties include building’s vulnerability index and

number of ﬂoors, and the dynamic data properties

include danger index, submersion height, ﬂood du-

ration, number of population and whether a demand

point is inhabited or not.

Using the previously described ontology, we have

constructed our knowledge graph integrating static

and dynamic ﬂood-related data. The static data were

integrated only once at the beginning of the ﬂood,

transformed into RDF triples and added to the ontol-

ogy. In contrary, the dynamic data were transformed

into RDF triples and updated in real-time throughout

the ﬂood. The transformation of static and dynamic

data into RDF triples has been performed using ”rd-

ﬂib” library

in python that maps data according to

the corresponding concepts and relations of the ontol-

ogy (Bu Daher et al., 2022).

3.4 Evacuation Priorities

There are four evacuation priorities deﬁned by the do-

main experts in ﬂood management for the evacuation

process as follows:

”Evacuate immediately”, ”Evacuate in 6 hours”,

”Evacuate in 12 hours” and ”No evacuation”. Each

https://rdﬂib.readthedocs.io/

KEOD 2022 - 14th International Conference on Knowledge Engineering and Ontology Development

evacuation priority is represented as a set of con-

ditions deﬁned for certain properties of the demand

points where the conditions of each priority are also

deﬁned by the domain experts. The properties used

for deﬁning the evacuation priorities are: danger in-

dex, duration of ﬂood, number of ﬂoors, submersion

height, vulnerability index, housing type and whether

the building is inhabited or not. The conditions are

deﬁned such that the evacuation priorities are exclu-

sive and consider all the possible values of properties

describing the demand points in the study area.

3.5 Inferring New Information

concerning Priorities

We recall that our aim is to manage and share hetero-

geneous ﬂood-related data among different actors in-

volved in the ﬂood crisis management in order to help

them access the needed data at the right time and take

decisions with the help of the inferred information.

Based on the knowledge graph that integrates all

the ﬂood-related data, we propose inferring new in-

formation concerning the evacuation priorities of the

demand points of the study area to help the ﬁreﬁghters

take rapid decisions and actions in the evacuation pro-

cess of ﬂood victims. As previously detailed, we have

four evacuation priorities that are represented as sub-

classes of the class ”Demand point” in the ontology.

Each demand point will then be typed with one of

these four evacuation priorities according to its prop-

erties. In other words, we classify the instances of

demand points in our study area into four categories

where each category represents one of the four evacu-

ation priorities. The inferred information is then used

to enrich the knowledge graph so that it is shared by

different actors involved in the management process.

4 THREE APPROACHES FOR

INFERRING EVACUATION

PRIORITIES

In this section, we discuss the three approaches for in-

ferring the evacuation priorities of the demand points

in our study area.

4.1 Inferring Evacuation Priorities

using FaCT++ Reasoner in Prot

The ﬁrst approach consists of enriching the knowl-

edge graph using the semantics of the ontology. Each

class representing an evacuation priority is deﬁned

with axioms that express the conditions that a demand

point should satisfy, and a reasoner is used to auto-

matically classify the demand points’ instances ac-

cording to the four priority classes.

Several reasoners already exist in the literature

for inferring new information from knowledge graphs

where some of them can be plugged in Prot

in-

cluding Pellet (Sirin et al., 2007), HermiT (Glimm

et al., 2014) and FaCT++ (Tsarkov and Horrocks,

2006) reasoners. This allows inferring new informa-

tion directly through the ontology editor and visualiz-

ing the demand points and their corresponding evac-

uation priorities. We choose the FaCT++ reasoner

in Prot

e to infer new information from our knowl-

edge graph which represent evacuation priorities to

demand points of the study area. FaCT++ is more ef-

ﬁcient on our knowledge graph than other reasoners

plugged in Prot

e. For example, The information

about the evacuation priorities is inferred from the

knowledge graph using FaCT++ in 1.24 hours, while

it takes 10 hours to be inferred using Hermit reasoner.

Figure 1 displays the four deﬁned classes of prior-

ities in our ontology visualized in Prot

e, and ﬁgure

2 displays the axioms deﬁned for the priority class

”Evacuate in 12h”. Similarly, axioms are deﬁned for

the corresponding classes of the three other evacua-

tion priorities.

Figure 1: Classes of Evacuation priorities in Prot

Figure 2: Evacuation priority: ”Evacuate in 12 hours”.

When loading the knowledge graph in Prot

FaCT++ reasoning can be directly conducted to in-

fer the evacuation priorities of demand points. The

instances of demand points are thus classiﬁed among

the four evacuation priorities according to their prop-

erties where each demand point can be classiﬁed and

viewed under the priority class that it satisﬁes. Figure

3 displays an example of some demand points classi-

ﬁed under the priority class ”Evacuate in 12h”

https://protege.stanford.edu/

Inferring New Information from a Knowledge Graph in Crisis Management: A Case Study

Figure 3: ”Inferences for ”Evacuate in 12h” priority in

Prot

We then obtain the knowledge graph enriched

with the inferred priorities, and this enriched knowl-

edge graph can thus be shared with the different actors

of the ﬂood management process, particularly the ﬁre-

ﬁghters concerned in the victims’ evacuation in order

to help them take decisions concerning the evacuation

process.

4.2 Inferring Evacuation Priorities

using SPARQL Queries

The second approach that we evaluate is enriching the

knowledge graph with inferred information about the

evacuation priorities using SPARQL queries

. In this

approach, the knowledge graph integrating the ﬂood-

related data is ﬁrst stored in a triplestore. We have

chosen ”Virtuoso” triplestore for storing knowledge

graphs as it is proved to be efﬁcient in storing a big

number of triples in a relatively short time

. For in-

stance, the results of a benchmark show that Virtuoso

loads 1 billion RDF triples in 27 minutes while it takes

hours to load the same triples in other triplestores such

as BigData, BigOwlim and TDB

After storing the knowledge graph in Virtuoso,

SPARQL insert and delete queries are implemented

to classify the demand points among the four de-

ﬁned evacuation priorities according to their proper-

https://www.w3.org/TR/rdf-sparql-query/

https://virtuoso.openlinksw.com/

http://wbsg.informatik.uni-mannheim.de/bizer/berlin

sparqlbenchmark/results/V7/#exploreVirtuoso

ties. The deﬁnitions of each priority class in the ﬁrst

approach are now expressed as conditions of each

evacuation priority in its ”WHERE” condition of the

SPARQL query.

We deﬁne an ”insert” statement in the SPARQL

query of each evacuation priority in order to type each

demand point that satisﬁes the conditions of the query

with this evacuation priority after making sure that

we have deleted all possible priorities with which this

demand point could have been typed before. A new

triple is thus obtained for each demand point express-

ing its typed priority. Figure 4 displays the SPARQL

query implemented for the priority ”Evacuate in 12h”.

Figure 4: ”SPARQL query of the priority ”Evacuate in 12”.

The knowledge graph is directly enriched with the

new triples deﬁning the evacuation priorities of de-

mand points on the triplestore, and it can be shared

with different actors to assist in the evacuation pro-

cess.

4.3 Inferring Evacuation Priorities

using SHACL Rules

The third evaluated approach is inferring the evacua-

tion priorities of demand points using SHACL rules.

A Shapes Constraint Language (SHACL) rule is a re-

cent kind of rules that can be used for inferring new

information from knowledge graphs while having ad-

vantages over other kinds of rules used in the litera-

ture, and it has not been used in the domain of crisis

management yet. A SHACL rule is identiﬁed through

a unique Internationalized Resource Identiﬁer (IRI)

not like other kinds of rules. In addition, it can be

activated or deactivated upon its usage purpose where

a deactivated rule is ignored by the rules engine and

is not executed. An execution order can also be deter-

mined for SHACL rules when more than one rule is

implemented.

Rules are executed using TopBraid SHACL API

which is an open source implementation of the W3C

SHACL based on Apache Jena

https://github.com/TopQuadrant/shacl

https://jena.apache.org/

KEOD 2022 - 14th International Conference on Knowledge Engineering and Ontology Development

There exist different types of SHACL rules in-

cluding SPARQL rules, which are based on SPARQL

query language that allows writing rules in SPARQL

notation. In this approach, we use SPARQL rules to

infer the evacuation priorities of demand points. For

each priority, we deﬁne a rule as follows. We ﬁrst

deﬁne the node shapes representing the classes that

describe the priorities and the property shapes rep-

resenting the properties used to deﬁne the priorities;

then, we deﬁne the rules. Referring to the axioms de-

ﬁned in the ﬁrst approach for each priority class, they

are now deﬁned as rules determining the conditions

of each evacuation priority. The SPARQL rule deﬁn-

ing the priority ”Evacuate in 12h” in the shape ﬁle is

displayed as follows.

sh:rule [

rdf:type sh:SPARQLRule ;

sh:preﬁxes ns1: ;

sh:construct ”””

PREFIX ns1: <https://www.irit.fr/recherches/

MELODI/ontologies/i-Nondations.owl# >

CONSTRUCT

{?this ns1:priority ?priority.}

WHERE

{ ?this ns1:danger index ?danger index.

?this ns1:duration of ﬂooding

?duration of ﬂooding.

?this ns1:number of ﬂoors ?number of ﬂoors.

?this ns1:submersion height ?submersion height.

?this ns1:vulnerability index ?vulnerability index.

?this ns1:is habitated ?is habitated.

FILTER

(?danger index >0

&& ?danger index <50

&& ?duration of ﬂooding ≥ 12

&& ?number of ﬂoors ≥ 1

&& ?submersion height >0.0

&& ?submersion height ≤ 1.0

&& ?vulnerability index <50.0

&& ?is habitated = true ) .

BIND (”12h before evacuation” AS ?priority).

}

””” ;

sh:condition ns1:12h before evacuation ;

The rules are executed on the knowledge graph to

infer new triples. Each inferred triple consists of a de-

mand point typed with an evacuation priority accord-

ing to its properties. The knowledge graph is then

enriched by adding the inferred triples to it, and it can

be shared by different actors in the crisis management

process.

5 EXPERIMENTAL EVALUATION

In this section, we discuss the experimental evalu-

ations conducted in order to evaluate the three ap-

proaches for inferring the evacuation priorities of the

demand points. The evaluations are divided into three

main categories. The ﬁrst category concerns analyz-

ing the impact of the variation of the number of in-

stances in a knowledge graph on the complexity of

the process of inferring new information. The sec-

ond category concerns analyzing the complexity of

the evacuation priorities. The third category concerns

analyzing the impact of the variation of the number of

evacuation priorities on the complexity of the process

of inferring new information.

All the conducted experiments run in 4h and 1min

on 8 CPUs Core i7-1185G7, and draw 0.28 kWh.

Based in France, this has a carbon footprint of 11.01 g

CO2e, which is equivalent to 0.01 tree-months (calcu-

lated using green-algorithms.org v2.1 (Lannelongue

et al., 2021)).

5.1 Variation of Number of Instances

In our knowledge graph, evacuation priorities are in-

ferred for demand points; therefore, the number of in-

stances of demand points determines the number of

evacuation priorities inferred. It also represents the

number of times the conditions of each evacuation

priority need to be tested in order to infer a priority

for a demand point. We thus aim at analyzing the

impact of the variation of the number of instances in

the knowledge graph on the complexity of three ap-

proaches of inferring the priorities in terms of execu-

tion time.

5.1.1 Knowledge Graph

The knowledge graph containing all the ﬂood-related

data of our study area is composed of 472,594 triples.

There are 15,078 demand points in our study area;

therefore, 15,078 new triples representing evacuation

priorities of demand points are inferred.

5.1.2 Experimental Results

A demand point is described by different proper-

ties and thus by different instances representing these

properties. To analyze the impact of the variation of

the number of instances, we study the execution time

with decreasing percentages of demand points in the

knowledge graph from 75% to 25% of the total num-

ber of demand points.

Table 1 displays the percentage of demand points

in the knowledge graph (named KG in the tables) with

Inferring New Information from a Knowledge Graph in Crisis Management: A Case Study

the number of demand points for each percentage

which thus represents the number of inferred evacu-

ation priorities of demand points.

Table 1: Percentage of demand points in KG.

% of demand points

in KG

Number of

evacuation priorities

100 % 15,078

75 % 11,308

50 % 7,539

25 % 3,769

Table 2 presents the execution times (in seconds)

of the three approaches for inferring the priorities

with decreasing percentages of demand points in the

knowledge graph. We recall that the three approaches

are as follows: FaCT++ in Prot

e, SPARQL insert

and delete queries as well as SHACL rules.

Table 2: Execution times (s) of inferring priorities using 3

approaches with decreasing percentages of demand points.

100%

of KG

75%

of KG

50%

of KG

25%

of KG

FaCT++ 4464.2 1809.5 1131.5 164.1

SPARQL

queries

80.62 41.37 23 8.09

SHACL

rules

12.86 8.78 6.55 4.56

From the results in table 2, we notice that the exe-

cution times decrease with decreasing percentages of

demand points in the knowledge graph using the three

approaches. Inferring the priorities using SHACL

rules takes 12.86 seconds for generating 15,078 pri-

orities (100% of demand points in knowledge graph)

and only 4.56 seconds for generating 3,769 priorities

(25% of demand points in knowledge graph). It is

the most efﬁcient approach in terms of execution time

compared to the two other approaches.

In the SPARQL approach, the knowledge graph

should ﬁrst be loaded on the triplestore which takes

from 3 to 4.5 seconds depending on its size. After

that, 2 SPARQL queries (delete and insert queries)

are executed for each evacuation priority in order to

deﬁne the priorities for all demand points and obtain

the enriched knowledge graph. On the other hand,

the chosen SHACL implementation is independent of

a triplestore where the knowledge graph, the node-

shapes of classes and properties as well as the rules

are deﬁned in a single Turtle RDF ﬁle. The new in-

ferred triples are then added to the knowledge graph

to obtain the enriched one.

FaCT++ in Prot

e takes around 36 times longer

than SHACL rules to infer the priorities for 25% of

the demand points in the knowledge graph. We note

that we use FaCT++ reasoner in Prot

e which could

not be conducted independently of all other Prot

functionalities; this would thus increase the complex-

ity of the process which explains the long execution

time of this approach compared to the other two ap-

proaches.

5.2 Evaluation of Complexity of

Evacuation Priorities

We deﬁne the complexity of an evacuation priority as

the number of demand points that are typed with this

priority and the time taken to classify them using the

three different approaches. An evacuation priority is

deﬁned as a set of conditions on certain properties that

describe the demand points where these conditions

are combined through ”and” and ”or” logic operators

such that all the possibilities of demand points’ prop-

erties are considered for the four evacuation priorities

in an exclusive manner. We deﬁne the worst case sce-

nario of an evacuation priority as the case where all

its conditions should be tested in order to determine if

a demand point is typed with this priority.

The number of conditions deﬁning each of the

four evacuation priorities as well as the properties

used in these conditions vary from one priority to an-

other. For example, the property ”duration of ﬂood” is

not used in the conditions deﬁning the priority ”Evac-

uate immediately”; however, it is used in the condi-

tions deﬁning the three other priorities. Table 3 dis-

plays the number of conditions deﬁning each evacua-

tion priority with considering the worst case scenario

for all the priorities.

Table 3: Number of conditions constituting evacuation pri-

orities.

Evacuation priority Number of conditions

Evacuate in 12 hours 8

Evacuate in 6 hours 16

Evacuate immediately 18

No evacuation 27

Due to the different number of conditions and

properties among the four evacuation priorities, we

can tell that they have different complexity. We eval-

uate the complexity of each of the four evacuation pri-

orities using the three different approaches.

5.2.1 Evaluation on Knowledge Graph

Containing Real Data

The knowledge graph that contains the real ﬂood-

related data corresponding to our study area consists

KEOD 2022 - 14th International Conference on Knowledge Engineering and Ontology Development

of 15,078 demand points with the various instances

of properties describing them. In this experiment, we

evaluate the complexity of the four evacuation pri-

orities on this knowledge graph using the three ap-

proaches for inferring the priorities of demand points.

Table 4 shows the execution times (in seconds) of

each evacuation priority in the three approaches, and

table 5 presents the number of demand points that are

typed with each priority.

Table 4: Execution times (s) of evacuation priorities.

Evacuation

priority

FaCT++

SPARQL

queries

SHACL

rules

Evacuate in 12h

2080.72 12.88 6.61

Evacuate in 6h

2080.83 7.2 7.28

Evacuate

immediately

2054.88 13.19 7.27

No evacuation 2424.88 64.52 9.96

Table 5: Number of demand points per evacuation priority.

Evacuation priority Number of demand points

Evacuate immediately 412

Evacuate in 12h 398

Evacuate in 6h 31

No evacuation 14,237

From these results, we can conclude that the

SHACL approach is the most efﬁcient one when be-

ing applied on the real ﬂood-related data of our study

area, and the SPARQL approach is more efﬁcient

than the approach that uses FaCT++ in Prot

e which

proves to be inefﬁcient on this knowledge graph.

However, our interest is to analyze the complexity of

the evacuation priorities further and in a more generic

manner. Using this knowledge graph, it is difﬁcult

to draw more precise conclusions due to two main

reasons. First, the number of conditions deﬁning

each evacuation priority is not identical among dif-

ferent priorities, and they are not necessarily deﬁned

by the same properties. Second, the number of de-

mand points that are typed with each evacuation pri-

ority is not identical as we can see in table 5. For

example, although the number of conditions deﬁning

the priority ”Evacuate in 12 hours” is less than that

of the priority ”Evacuate in 6 hours”, inferring the

information related to the former priority takes more

time (12.88 seconds) than the latter (7.2 seconds) us-

ing SPARQL queries. The reason is that there are 398

demand points typed with the former priority while

only 31 demand points typed with the latter.

5.2.2 Evaluation using a Synthetic Knowledge

Graph

In order to evaluate the complexity of the evacua-

tion priorities more precisely, we propose to evalu-

ate the impact of the number of conditions deﬁning

each evacuation priority on its complexity. To do that,

there are two important factors to be taken into con-

sideration. First, the number of demand points sat-

isfying each evacuation priority should be ﬁxed as it

impacts the complexity. Second, the worst case sce-

nario should be considered for all evacuation prior-

ities in order to ensure a precise evaluation of their

complexity. Considering the worst case scenario rep-

resents having demand points whose properties en-

sure the need of testing each and every condition in

all the evacuation priorities. Therefore, we propose a

new synthetically generated knowledge graph for the

purpose of evaluating the complexity of the priorities.

This synthetic knowledge graph is generated such

that there are four categories of demand points; each

category contains 4,000 demand points whose prop-

erties satisfy the conditions of an evacuation prior-

ity. The total number of demand points in the knowl-

edge graph is then 16,000. This knowledge graph

is generated as follows. Random data are automati-

cally generated representing random ID’s of demand

points with their properties respecting the conditions

of the evacuation priorities. These data are then trans-

formed into RDF triples to form the knowledge graph

relying on the concepts and relations of our ontology.

The RDF triples constituting the knowledge graph are

generated using JENA Java library

As the experimental results of the approach of in-

ferring evacuation priorities using FaCT++ in Prot

has proved to be inefﬁcient compared to the two other

approaches in terms of time, we exclude this approach

from the coming experiments, and we will thus only

evaluate the two other approaches, that are SPARQL

insert and delete queries and SHACL rules.

Table 6 presents the experimental results in terms

of execution time (in seconds) of each evacuation pri-

ority using the two approaches.

We can notice from these results that the execu-

tion time increases as the number of conditions deﬁn-

ing each priority increases in both approaches (re-

fer to table 3 for the number of conditions deﬁn-

ing each priority). This conﬁrms that the number of

conditions deﬁning an evacuation priority impacts its

complexity. In addition, inferring the priorities us-

ing the SHACL approach takes less time than the

SPARQL approach for all the evacuation priorities

which proves that it is the most efﬁcient approach for

https://jena.apache.org/

Inferring New Information from a Knowledge Graph in Crisis Management: A Case Study

Table 6: Execution times (s) of evacuation priorities of the

synthetic knowledge graph.

Evacuation

priority

SPARQL

queries

SHACL

rules

Evacuate in 12h 104.54 7.97

Evacuate in 6h 128.49 8.59

Evacuate

immediately

158.99 9.01

No evacuation 187.89 9.13

inferring the priorities.

5.3 Variation of Number of Evacuation

Priorities

The number of evacuation priorities represents the

number of times that the properties of each demand

point are tested against the conditions of all the evac-

uation priorities in order to type this demand point

with its appropriate evacuation priority. Therefore,

the number of evacuation priorities tested in an inte-

grated manner on the knowledge graph has an impact

on the complexity of the process of inferring the evac-

uation priorities.

There are four evacuation priorities deﬁned by the

domain experts concerning the evacuation of victims.

These priorities are deﬁned such that they consider

all the possibilities of demand points, and they are

exclusive. In the following experiments, we aim at

evaluating the impact of the variation of the number

of evacuation priorities on the complexity of the pro-

cess of inferring the priorities of demand points. For

the sake of this evaluation, we propose deﬁning new

evacuation priorities in addition to the four priorities

deﬁned by the domain experts. As the demand points

of our knowledge graph representing the real data of

our study area are typed with the four existing evac-

uation priorities, we propose to enrich the knowledge

graph so that it contains new demand points whose

instances of properties satisfy the new evacuation pri-

orities.

5.3.1 New Evacuation Priorities

The evacuation priority having the lowest complex-

ity among the four priorities deﬁned by the domain

experts is ”Evacuate in 12h”. We thus propose to

add three new evacuation priorities that have the same

complexity as this priority for the sake of simplicity of

evaluations. As previously mentioned, the four evacu-

ation priorities are exclusive and consider all the pos-

sibilities. Therefore, in order to add new priorities,

we choose to modify the existing evacuation prior-

ity ”Evacuate in 12h” by adding only one condition

Table 7: The intervals deﬁning the property ”duration of

ﬂooding” in the new evacuation priorities.

Evacuation priority

”duration of ﬂooding”

interval

Evacuate in 12h

(modiﬁed priority)

>=12h , <15h

Evacuate in 15h >=15h , <18h

Evacuate in 18h >=18h , <24h

Evacuate in 24h >=24h , <1000h

such that we can deﬁne new priorities from this one

through modifying this condition and maintaining the

exclusiveness of the different priorities. Let’s recall

the conditions deﬁning this evacuation priority as pre-

sented in ﬁgure 2. We propose to add a condition con-

cerning the property ”duration of ﬂooding” to allow

the division of the values’ range of this property in

order to deﬁne several evacuation priorities.

We then deﬁne three new evacuation priorities

from the priority ”Evacuate in 12h”. The difference

among these priorities is the value range of the prop-

erty ”duration of ﬂooding” as displayed in table 7.

5.3.2 Knowledge Graph Enrichment

We propose to enrich our knowledge graph with ad-

ditional RDF triples representing demand points with

their instances satisfying the three new evacuation pri-

orities. For this purpose, we have generated synthetic

RDF triples of three categories. Each category con-

tains demand points whose instances satisfy the con-

ditions of one of the new evacuation priorities. We de-

ﬁne 398 demand points for each category that repre-

sents the number of demand points initially satisfying

the evacuation priority ”Evacuate in 12h”. The en-

riched knowledge graph is generated similarly to the

generation of the synthetic knowledge graph for the

previous experimental evaluation, and it thus contains

16,272 demand points with their instances.

5.3.3 Experimental Evaluation

We ﬁrst evaluate the complexity of the modiﬁed evac-

uation priority ”Evacuate in 12h” on the enriched

knowledge graph using the two approaches: SPARQL

insert and delete queries and SHACL rules. We then

evaluate the complexity of the four existing evacua-

tion priorities on the enriched knowledge graph us-

ing the two approaches. After evaluating the exist-

ing priorities, we evaluate the impact of the variation

of the number of evacuation priorities on the process

of inferring the priorities of demand points from the

enriched knowledge graph using the two approaches.

Table 8 displays the results of these evaluations in

terms of execution time (in seconds) where ”4 pri-

KEOD 2022 - 14th International Conference on Knowledge Engineering and Ontology Development

orities” represent the four existing evacuation prior-

ities deﬁned by the domain experts. We notice from

Table 8: Time for inferring priorities of demand points (s)

with 7 evacuation priorities progressively.

New evacuation priority

SPARQL

queries

SHACL

rules

Evacuate in 12h 12.88 6.61

4 priorities 84.32 13.1

4 priorities + Evacuate in

15h

94.93 15.34

4 priorities + Evacuate in

15h + Evacuate in 18h

104.59 17.61

4 priorities + Evacuate in

15h + Evacuate in 18h

+ Evacuate in 24h

115.08 19.18

the results that the time increases after the addition

of each evacuation priority using the two approaches.

The time increase after adding each new priority is

by around 10 seconds using the SPARQL approach

which is close to the time of inferring the priority

”Evacuate in 12 hours” for the corresponding demand

points. While the time increase after adding each pri-

ority is around 2 seconds using the SHACL approach

which is less than the time taken to infer the prior-

ity ”Evacuate in 12 hours” for the corresponding de-

mand points. It takes less time to infer the priori-

ties of 16,272 demand points using the SHACL ap-

proach (19.18 seconds) than using the SPARQL ap-

proach (115.08 seconds).

5.4 Discussion

The evacuation process of ﬂood victims is a critical

process that should be rapid and efﬁcient as it threat-

ens the lives of the population. The SHACL approach

is proved to be the most efﬁcient one among the three

approaches for inferring the evacuation priorities in a

real use case as well as using synthetic data that are

generated for various experimental purposes. It is ca-

pable of generating priorities for a big number of de-

mand points in a short time compared to the two other

approaches.

The constraint of inferring the evacuation prior-

ities of demand points rapidly in order to assist in

taking rapid actions concerning victims’ evacuation

is not satisﬁed in the approach of FaCT++ in Prot

due to the long time that it takes to infer the priori-

ties. It is thus considered inefﬁcient for inferring the

priorities.

The delicacy of the evacuation process requires

proposing solutions that are not only efﬁcient and

rapid but also simple to assist in this process. The per-

sons in charge of the evacuation process are usually

non-experts in novel technological techniques; there-

fore, the proposed solutions to assist in taking actions

and decisions should facilitate the process. SPARQL

queries deﬁning the priorities can be integrated in a

tool where a natural language query written by the

user concerning an evacuation priority can be trans-

formed into a SPARQL query that allows inferring

the information about this priority using existing ap-

proaches that propose query transformations (Shaik

et al., 2016; Ochieng, 2020). SHACL rules that are

used in our third approach are based on SPARQL

query language; therefore, it can take the same advan-

tage of transforming natural queries to rules and thus

can be integrated in a tool used by the users to obtain

information. In addition, integrating SHACL rules in

a tool would allow users to choose whether to activate

or deactivate rules as well as to set an execution order

to different rules based upon their needs.

6 CONCLUSION

Natural crises management is a critical process; there-

fore, we consider that it is important to propose so-

lutions that help in the information management and

sharing among different involved actors as well as to

infer new information that assist in the crisis manage-

ment process. In this paper, we have proposed a case

study that relies on a previously proposed ontology

and a knowledge graph integrating real ﬂood-related

data (Bu Daher et al., 2022) in order to evaluate three

approaches for inferring new information represent-

ing evacuation priorities for demand points during a

ﬂood crisis. The three approaches were evaluated us-

ing real data as well as synthetic data for further anal-

ysis. The experimental results show that inferring the

priorities using SHACL rules is the most efﬁcient ap-

proach as it allows inferring the priorities in a short

time and assist in taking rapid decisions and actions.

As a future work, we ﬁrst aim at translating this

work into an industrial usage through an interface

that integrates SHACL Rules and allows to transform

users’ natural language queries to rules. This allows

the users to obtain information about evacuation pri-

orities of demand points in a study area, to activate or

deactivate different rules as well as to set their execu-

tion order according to their needs.

We also aim at relying on the ontology and the

knowledge graph in order to infer new information

concerning the management of the resources (evacu-

ation vehicles) that are used for the evacuation pro-

cess and the routes organization. In addition to in-

ferring new information that assists in the crisis re-

Inferring New Information from a Knowledge Graph in Crisis Management: A Case Study

sponse phase, we aim at inferring new information to

improve a past crisis experience. In this frame, we

aim at proposing a learning approach that learns from

the data of a past crisis and adjusts the values of the

properties that deﬁne the conditions of different evac-

uation priorities with the aim of improving the expe-

rience of this crisis.

ACKNOWLEDGMENTS

This work has been funded by the ANR in the context

of the project ”i-Nondations” (e-Flooding), ANR-17-

CE39-0011

REFERENCES

Bu Daher, J., Huygue, T., Stolf, P., and Hernandez, N.

(2022). An ontology and a reasoning approach for

evacuation in ﬂood disaster response. In 17th Interna-

tional Conference on Knowledge Management 2022,

page To appear. University of North Texas (UNT) dig-

ital library.

Elmhadhbi, L., Karray, M. H., and Archim

ede, B. (2019).

A modular ontology for semantically enhanced in-

teroperability in operational disaster response. In

16th International Conference on Information Sys-

tems for Crisis Response and Management-ISCRAM

2019, pages 1021–1029.

Franke, J. (2011). Coordination of Distributed Activities in

Dynamic Situations. The Case of Inter-organizational

Crisis Management. PhD thesis, Universit

e Henri

Poincar

e-Nancy I.

Glimm, B., Horrocks, I., Motik, B., Stoilos, G., and Wang,

Z. (2014). Hermit: an owl 2 reasoner. Journal of

Automated Reasoning, 53(3):245–269.

Hill, E. F. (2003). Jess in action: Java rule-based systems.

Manning Publications Co.

Katuk, N., Ku-Mahamud, K. R., Norwawi, N., and Deris, S.

(2009). Web-based support system for ﬂood response

operation in malaysia. Disaster Prevention and Man-

agement: An International Journal.

Khantong, S., Sharif, M. N. A., and Mahmood, A. K.

(2020). An ontology for sharing and managing infor-

mation in disaster response: An illustrative case study

of ﬂood evacuation. International Review of Applied

Sciences and Engineering.

Kurte, K. R., Durbha, S. S., King, R. L., Younan, N. H.,

and Potnis, A. V. (2017). A spatio-temporal ontologi-

cal model for ﬂood disaster monitoring. In 2017 IEEE

International Geoscience and Remote Sensing Sympo-

sium (IGARSS), pages 5213–5216. IEEE.

Kurte, K. R., Durbha, S. S., King, R. L., Younan, N. H.,

and Vatsavai, R. (2016). Semantics-enabled frame-

work for spatial image information mining of linked

https://anr.fr/Projet-ANR-17-CE39-0011

earth observation data. IEEE Journal of Selected Top-

ics in Applied Earth Observations and Remote Sens-

ing, 10(1):29–44.

Lannelongue, L., Grealey, J., and Inouye, M. (2021). Green

algorithms: Quantifying the carbon footprint of com-

putation. Advanced Science, 8(12):2100707.

Ochieng, P. (2020). Parot: Translating natural language

to sparql. Expert Systems with Applications: X,

5:100024.

Scheuer, S., Haase, D., and Meyer, V. (2013). Towards

a ﬂood risk assessment ontology–knowledge integra-

tion into a multi-criteria risk assessment approach.

Computers, Environment and Urban Systems, 37:82–

94.

Shaik, S., Kanakam, P., Hussain, S. M., and Suryanarayana,

D. (2016). Transforming natural language query to

sparql for semantic information retrieval. Interna-

tional Journal of Engineering Trends and Technology,

7:347–350.

Sirin, E., Parsia, B., Grau, B. C., Kalyanpur, A., and Katz,

Y. (2007). Pellet: A practical owl-dl reasoner. Journal

of Web Semantics, 5(2):51–53.

Sun, J., De Sousa, G., Roussey, C., Chanet, J.-P., Pinet,

F., and Hou, K.-M. (2016). Intelligent ﬂood adap-

tive context-aware system: How wireless sensors

adapt their conﬁguration based on environmental phe-

nomenon events. Sensors & Transducers, 206(11):68.

Tsarkov, D. and Horrocks, I. (2006). Fact++ description

logic reasoner: System description. In International

joint conference on automated reasoning , pages 292–

297. Springer.

Wang, C., Chen, N., Wang, W., and Chen, Z. (2018). A hy-

drological sensor web ontology based on the ssn on-

tology: A case study for a ﬂood. ISPRS International

Journal of Geo-Information, 7(1):2.

Yahya, H. and Ramli, R. (2020). Ontology for evacua-

tion center in ﬂood management domain. In 2020 8th

International Conference on Information Technology

and Multimedia (ICIMU), pages 288–291. IEEE.

KEOD 2022 - 14th International Conference on Knowledge Engineering and Ontology Development