Advances on cognitive automation at LGI2P / Ecole des Mines d'Als - - PDF document

▶

Mar 21, 2024 25 likes •716 views

Advances on cognitive automation at LGI2P / Ecole des Mines d'Als Doctoral research snapshot 2016-2017 July 2017 Research report RR/17-01 Foreword This research report sums up the results of the 2017 PhD seminar of the LGI2P lab of IMT

SLIDE 1

July 2017 Research report RR/17-01

Advances on cognitive automation at LGI2P / Ecole des Mines d'Alès Doctoral research snapshot 2016-2017

SLIDE 2

SLIDE 3

Foreword

This research report sums up the results of the 2017 PhD seminar of the LGI2P lab of IMT Mines Ales. This annual day-long meeting gathers presentations of the latest research results

f LGI2P PhD students.

This year’s edition of the seminar took place on June 22nd. Thirteen PhD students presented their work for the past academic year. All presentations were followed by extensive time for very constructive questions from the audience. The aggregation of abstracts of these works constitute the present research report and gives a precise snapshot of the research on cognitive automation led in the lab this year. I would like to thank all lab members, among which all PhD students and their supervisors, for their professionalism and enthusiasm in helping me prepare this seminar. I would also like to thank all the researchers that came to listen to presentations and ask questions, thus contributing to the students’ PhD thesis defense training. A special thank to Yannick Vimont, head of the LGI2P lab since numerous years who will soon be joining Mines ParisTech. Thank you for providing me with the context to organize this yearly event and our monthly research seminars since 2009 ! I wish you all an inspiring reading and hope to see you all again for next year’s 2018 edition ! Christelle URTADO

SLIDE 4

SLIDE 5

Contents First year PhD students

Sabbah DIMASSI

Page 5

Recommandation multicritère par filtrage collaboratif sur des modèles de préférences et des similarités fonctionnelles : le cas de la recherche d'un logiciel d'entreprise sur un comparateur Meyi DULEME

Page 9

Monitoring de la charge cognitive dans le cadre de la gestion de crise Alexandre LE BORGNE

Page 13

Towards versioning three-level component-based architectures Cécile LHERITIER

Page 17

Contribution à l’amélioration des pratiques du retour d’expérience : Application au secteur de l’aide humanitaire Jean-Christophe MENSONIDES

Page 21

Evaluation d’un risque politique par l’analyse d’articles de presse au travers de techniques de traitement automatique du langage naturel : le cas des déchets radioactifs Behrang MORADI KOUTCHI Page 25 Formalisation et évaluation des propriétés non fonctionnelles pour l'ingénierie de système de système : application à la résilience Jocelyn PONCELET Page 31 Définition et gestion centralisée de l’assortiment idéal dans un réseau de magasins de la grande distribution

Second year PhD students

Valentina BERETTA Page 37

Evaluation of data veracity using value confidence and source trustworthiness

Gildas TAGNY NGOMPE Page 41

Extracting information from court decisions

Third year PhD students

Pierre-Antoine JEAN Page 46

On the definition of a knowledge inference model from relation extraction

SLIDE 6

Massissilia MEDJKOUNE Page 51

Towards a non-oriented approach for the evaluation of odor quality

Diadie SOW Page 55

A possibilistic framework for identifying the most promising oerformance for improving complex systems

Fourth year PhD students

Hasan ABDULRAHMAN Page 61

Oriented filters for feature extraction in digital images: application to contours, corners detection, evaluation

SLIDE 7

First year PhD students

Page 3

SLIDE 8

Page 4

SLIDE 9

preferences and similarity of products:

Application to SaaS search engine

Sabah Dimassi1,2, Abdelhak Imoussaten1, Jacky Montmain1, François Trousset1, Marc Ricordeau2

1ÉLGI2P, 69 Rue Georges Besse, Nîmes, France 2Cloud is Mine, appvizer, Cap Omega, Rond-point Benjamin Franklin, Montpellier, France

!"#"$%&'("!!')('*+!,"-+!%./

Abstract. r-
their capabilities to quickly offer to users the most relevant results according to their needs. To

improve the relevance of the search results, recommender systems are generally used within search engines to process needs and preferences. In our project we are focusing our research on BtoB1 recommender systems in order to improve the quality of the returned results by wever, time and data complexity represent con- intend to set up a professional refined recommender system that offers relevant results in reasonable time without bothering users with tedious questioning to get the required data in order to build their preference models. We suggest using MCDA2 s- inferred from collaborative filtering on usersfunctional similarities between products. This project is supported by Cloud is Mine, creator of the business SaaS3 search engine appvizer. Keywords: Recommender system (RS), MCDA2, Preference model, Similarity measure.

1 Introduction

Nowadays, recommender systems (RS) are being used in various fields and new research are done to improve [1] the relevance of their results. The interest in this area is high because of the abundance of practical applications that help users to deal with a big amount of data and provide personalized relevant recommendations in reasonable time without being too data greedy while interacting with users [2]. Amazon.com and other vendors have incorporated recommendation capabilities into their commerce servers in order to be as efficient as possible while giving personalized recommendations [3]. Despite of all these advances, the current generation of RSs still requires further improvements to reach a higher level of efficiency. When dealing with BtoB1 recommendation, we may access to more information especially on users and then applications can provide more personalized services by focusing on . RSs are usually classified into different categories of recommendation: Content-Based (recommend similar products), Collaborative (recommend same products to similar users), Hybrid (combination of the two). In this kind of application, we are dealing with preference on

1 Business to business 2 Multiple Criteria Decision Analysis 3 Software as a Service

Page 5

SLIDE 10

criteria). MCDA2 methods are then suita- ble to deal with this kind of problems [5]. MCDA2 methods are mainly divided into two categories called Outranking methods and Aggregation methods. The choice of the method will be directed by constraints in term of complexity and relevance preference expression. At this stage of our research, we believe that aggregation methods should be the more appropriate MCDA2 models for our purposes because these models are to be merged into optimization problems whose solving will be clearly simplified thanks to the analytical formalization of the multi attribute utility theory (MAUT) [6]. To precisely identify literature but generally require a lot of information from user [7, 8]. In our case, the application cannot be so intrusive. Our approach must be able to take into account data issued from well as possible while minimizing interactions with the user. The method we are proposing will tend to respect all these constraints to give the best recommendation based on hybrid methods enhanced by the use of MCDA2 methods: effi- non-intrusiveness (as less interaction with user as possible). Furthermore, our RS will have to improve its performance using user feedback by updating user and product similarity functions. Moreover as this project is part of an industrial collaboration with a BtoB1company (Cloud is Mine) the result will be applied in a real industrial case.

2 Proposed approach

Our approach can be modeled using the following elements (see Figure 1): User Mining: profile similarity metrics Content Mining: Request: list of criteria representi Response: list of products representing the results provided to users Profile: set of characteristics related to a specific user Configuration: preferences combining process Feedback: data provided by user about his satisfaction/dissatisfaction MCDA: method that computes the relevance of a product Expert: domain specialist that takes part in the preference initialization The MCDA2 assessment of products w.r.t. a set of criteria in N is based on aggregation

perator
1

( , ) : ... 0,1

X X

H x

, where x is the vector of characteristics of the as-

sessed product and the parameter vector of the preference model H . ( , ) H x provides the relevance of product (x) relatively to the preference model H of a user parameterized by . First a new user is identified by the RS (e.g., through LinkedIn or appvizer account). Secondly, this user formulates his request Q . Our RS then searches similar requests in the appvizer database. This search provides a set of users U (and their related preference parameters

) that have launched similar requests. An initial

is then computed by combining the

parameters associated to each user u in U (the more similar user u is with

the new user, the higher the contribution of

). The model (., ) H

can now provide

a list of relevant products according to preference model (., ) H

. The user may then ex-

press his contentment or dissatisfaction ( button) on each results in the list (Feedback). This feedback information is used to compute more accurate preference pa- Page 6

SLIDE 11

rameters

w.r.t. the user feedback by using products similarity (See 2.1 for details). The process can of course be reiterated. All the products similar to those that have been rejected (resp. approved) should be bad (resp. well) scored with the new evaluator function

(., ) H

Advanced control functions must also be defined to ensure consistency of the products and users similarities in the database (See 2.2 and See 2.3). While the first process is related to the on line improvement of user preferences model, the two other ones are done off line for control purposes when sufficient relevant feedbacks to warranty the consistency of the evolving database. Figure 1. Proposed approach for the intended multicriteria recommender system 2.1 On line user preference improvement (Feedback) At each feedback from useru , a set of constraints that capture the actions of

u is generated. The idea beyond the following constraints is that an approved product must

remain well scored with the new identified model

u i

whereas a rejected product must be badly scored with this new evaluation function.

, ,

u u u k i k i k

H x H x

if the user has approved

, ,

u u u k i k i k

H x H x

if the user has rejected

* 1

argmax

u i

u u i k k

This set of constraints is completed by introducing similarity between products, i.e.,

these constraints are extended to products that are similar to approved or rejected products. 2.2 Off line control based on users similarity The idea beyond this control function is to provide the best initial

users. If

there is no control loop, the consistency of the user database may be degraded when feedbacks are inserted into the database. The following condition must hold to warranty the database consistency and integrity: Page 7

SLIDE 12

( , ')

U u u

sim u u

Ensuring this constraint holds for any pair of users entails to update the definition of the

similarity

sim over the set of users. This control process is carried out off line.

2.3 Improving Product Classification The idea beyond this second control function is to make more robust the user preference improvement procedure. Two similar products characterized by x and

' x must have simi-

lar evaluations according to any evaluation function

(., )

. The following condition

must hold to warranty the database consistency and integrity:

( , ') , ( , ) ( ', )

X u u

sim x x u H x H x

Ensuring this constraint holds for any pair of products entails to update the definition of

the similarity

sim

ver the set of products. This control process is carried out off line.

3 Conclusion and perspectives

Through this paper, we have presented the academic issues related to our industrial project and proposed new perspectives for professional multicriteria recommender systems (RS). We propose to improve the relevance of results by introducing We suggest using collaborative filtering on users and products profiles and feedbacks. We have first studied all types of MCDA2 methods to get an initial

verview about the possible multicriteria preference models. The MAUT models a priori

seem to be relevant formalizations in our RS problematic because of their simplicity of use in optimization problems. We also intend to introduce approximate reasoning in our RS by managing similarities among users and products. Further research is engaged to refine and validate our modeling assumptions. Research about MCDA2 methods is still ongoing while these modeling hypotheses are tested on appvizer database.

References

1. F.Lorenzi and F.Ricci.: Case-Based recommender systems: A unifying view. In B. Mobasher and
S. S. Anand, editors, ITWP, pages 89113. Springer, 2003
2. G.Adomavicius, N.Manouselis and Y.Kwon.: Multicriteria Recommender Systems. Springer,

2011.

3. G.Adomavicius and A.Tuzhilin.: Toward the next generation of recommender systems: A survey
f the state-of-art and possible extensions. IEEE Trans.Knowl.DataEng, 2005.
4. .: Trust in recommender systems. Proceeding of the 10th International

Conference on Intelligent User Interfaces, San Diego, California, USA, ACM Press, 2005.

5. B.Roy and D.Bouyssou.: Aide multicritère à la décision : méthodes et cas. Economica, 1993.
6. R. L Keeney and Howard Raiffa. Decision with multiple objectives, 1976.
7. Bana e Costa, C. A., De Corte, J. M., Vansnick, J. C. On the mathematical foundation of

MACBETH (pp. 409-437), Springer New York, 2005.

8. Saaty, T. L. What is the analytic hierarchy process? (pp. 109-121), Springer Berlin Heidelberg,

1988.

Page 8

SLIDE 13

Cognitive Load Monitoring in Crisis Management

Méyi Dulème1,2, Gérard Dray1, Stéphane Perrey2, Florian Tena-Chollet3

1! 2 Euromov, Université Montpellier, Montpellier, France 3

de Clavières, Alès, France Abstract: Human stake holder cognitive limitation is a big issue in the crisis management field. Being able to highlight cognitive load increase is a matter of concern. Here, we are investigating a way to monitor the cognitive load increase during an emergency situation, in the crisis cell simulator (ISR, EMA). It has been proved that cognitive load can modulate performances in a secondary task, as well as brain imagery signal, heart rate and motor control. In our work, healthy human subjects will perform a n- back task, with a secondary auditory detection task. Performances in the secondary task, electroencephalography (EEG) and functional near-infrared spectroscopy (fNIRS) signal, ECG and motion position and acceleration will be measured. Exploratory analysis, using support vector machine (SVM) will potentially allow us to find robust physiological signatures of cognitive load increase. Keywords: Crisis management, Cognitive load, Sensorimotor activity, Neuroimagery, Machine Learning, Classification clustering

1 Introduction

holder.

Cognition is the mental action or process of acquiring knowledge and understanding through thought, experience, and the senses. It comprises processes like memory, attention or reasoning, that take place in the brain. Despite its great complexity (around 1010 interconnected neurons), the brain is a limited system. One limit is the storage space. Stake- limits are then an actual hazard in crisis management [1]. Since it would increase the efficiency of crisis management, to be able to monitor the cognitive load could be a major issue. Cognitive load (or Mental Workload) is a function of the cognitive demands associated with task performance and

f the mental processing ability of the person performing the task [2]. The

concept of cognitive load is then inseparable from the concept of limited cognitive resources. This latter has been quite well defined (for review see [3]). In a strictly cognitive point of view, according to this review, there are three illustrations of limited processes. The change detection task highlights visual short term memory (VSTM) limited resources. In the change detection task, subjects have to detect if a stimulus has changed in a set of visually presented stimuli (see fig. 1a). There is a threshold of the number of stimuli for which subjects are no more able to detect change. The inattentional blindness illustrates attention limitation. Subjects have to detect two stimuli rapidly presented (Rapid Serial Visual Presentation, Page 9

SLIDE 14

RSVP). The closer in time these stimuli are, the less subjects are capable of detecting the last presented one (see fig. 1b).. The Psychological Refractory Period (PRP) shows a . This task consists of answering differently to two stimuli will be. The shorter the inter-stimuli interval is, the longer the response time to the second stimulus (see fig. 1c).. This review only focus on the central nervous system but we must bear in mind that physical load exists too. a) b) c)

Figure 1: Three illustrations of limited cognitive processes. (a) Rapid Serial Visual presentation with two targets (T1 and T2) to detect. SOA = Stimulus onset asynchrony (b) Change detection task (c) Psychological Refractory period. Source : [3]

Our work is placed in the context of the crisis cell simulator (Institut des

to improve stake-training in realistic

conditions. We can then question if there are physiological signatures of an increasing cognitive load, compatible with body movement. Also, what analysis methods will permit us to detect them? Our hypothesis is that we will find robust physiological signatures of an increasing cognitive load. Page 10

SLIDE 15

2 Cognitive load measurement techniques

Several metrics of the increase of cognitive load are relevant: the evolution

f task performance, subjective reports, brain imagery (EEG and fNIRS),
cculomotricity and heart rate (ECG) measurement (for review see [4]).

To assess cognitive load, an interesting tool would be the performance of a (genuine) secondary task, like a peripheral detection task (PDT). Participants are instructed to press a button attached to their index finger each time they see a randomly presented LED signal. Performance in the PDT actually decreases with the increase of the cognitive load imposed by the primary task. Decrease in performance in the PDT means an increase of reaction time and an increase of missed signals as well. Subjects are also able to make cognitive load self-reports. The more widely used is the NASA Task-Load Index (NASA TLX) [5], but there are various existing subjective report questionnaires. Concerning brain imagery, the increase of cognitive demand of the task typically leads to a modulation of EEG signal: a suppression of alpha (8-12 Hz) activity and an elevation of beta activity (12-40 Hz) [6]. The fNIRS (functional Near Infrared Spectroscopy) is also robustly modulated: there is an increase of oxygenated blood flow in the left prefrontal area [7]. In addition, an increasing mental workload can increase heart rate and decrease heart rate variability at the same time [4]. A modulation of

cculomotricity (saccades) and pupilar diameter can also be found.

There is also evidence that movement (postural sway, for example) is impacted by cognitive load [8].

3 Experimental protocol

About 20 healthy subjects will be placed in front of a screen. They will have to complete a dual n-back task with a keyboard. In addition to a classical visuo-spatial n-back task - that is, a matching task at various lags - , they will have to detect a specific auditory stimulation. There will be 3 levels of difficulty, to manipulate the level of cognitive load (1-back, 2-back and 3- back task). During this task, brain signal will be recorded with EEG and

fNIRS. Heart rate and movement will be measured too, respectively with

ECG and motion capture and accelerometers. Performance and subjective reports will be assessed as well.

4 Analysis methods

Performances and subjective assessment will be analyzed with classical descriptive statistics. A correlation calculation will be made to check if there is dissociation: that is, incongruence of cognitive load metrics. To find potential physiological signatures, we will use mainly exploratory methods such as machine learning. An EEG frequency analysis and a binary classification (Support Vector Machine, SVM) of physiological measures will be made: 1-back VS rest, 2-back VS rest and 3-back VS rest. In addition, it has been shown that fNIRS + EEG system, which we choose for Page 11

SLIDE 16

this work, allows better classification accuracy [9]. Interaction between physiological metrics and performance will be assessed as well.

5 Continuation of work

If we find interesting results, next experiments will be about looking for physiological signatures during crisis management (crisis cell simulator), with the same measurement and analysis methods. We will also investigate if the physiological signatures highlighted are impacted by social

interaction. Indeed, the crisis cell is based on group work.

We will also consider to set a biofeedback protocol and to model the physiologic markers evolution with the increase of cognitive load.

References

1 Dehais, F., Causse, M., Vachon, F., Régis, N., Menant, E., & Tremblay,

S. (2014). Failure to detect critical auditory alerts in the cockpit: evidence

for inattentional deafness. Human Factors, 56(4), 631-644. 2 Parasuraman, R., Sheridan, T. B., & Wickens, C. D. (2008). Situation awareness, mental workload, and trust in automation: Viable, empirically supported cognitive engineering constructs. Journal of Cognitive Engineering and Decision Making, 2(2), 140-160. 3 Marois, R., & Ivanoff, J. (2005). Capacity limits of information processing in the brain. Trends in cognitive sciences, 9(6), 296-305. 4Young, M. S., Brookhuis, K. A., Wickens, C. D., & Hancock, P. A. (2015). State of science: mental workload in ergonomics. Ergonomics, 58(1), 1-17. 5 Hart, S. G., & Staveland, L. E. (1988). Development of NASA-TLX (Task Load Index): Results of empirical and theoretical research. Advances in psychology, 52, 139-183. 6 Borghini, G., Vecchiato, G., Toppi, J., Astolfi, L., Maglione, A., Isabella, R., . . . Babiloni, F. (2012). Assessment of mental fatigue during car driving by using high resolution EEG activity and neurophysiologic indices. Proceedings of Engineering in Medicine and Biology Society (EMBC), 2012 Annual International Conference of the IEEE (pp. 64426445). New York, NY: IEEE. 7 Ayaz, H., Shewokis, P. A., Bunce, S., Izzetoglu, K., Willems, B., & Onaral, B. (2012). Optical brain monitoring for operator training and mental workload assessment. Neuroimage, 59(1), 36-47. 8 Bucci, M. P., Seassau, M., Larger, S., Bui-Quoc, E., & Gerard, C. L. (2014). Effect of visual attention on postural control in children with attention-deficit/hyperactivity disorder. Research in developmental disabilities, 35(6), 1292-1300. 9 Aghajani, H., & Omurtag, A. (2016, August). Assessment of mental workload by EEG+ FNIRS. In Engineering in Medicine and Biology Society (EMBC), 2016 IEEE 38th Annual International Conference of the (pp. 3773-3776). IEEE. Page 12

SLIDE 17

Towards versioning three-level component-based architectures

Alexandre Le Borgne1, David Delahaye2, Marianne Huchard2, Christelle Urtado1, and Sylvain Vauttier1

1 IMT Mines Al`

es {alexandre.le-borgne, christelle.urtado, sylvain.vauttier}@mines-ales.fr

2 LIRMM / CNRS & Montpellier University, France {David.Delahaye,

Marianne.Huchard}@lirmm.fr

Keywords: Component-Based Software Engineering, Architecture evolution, Archi- tecture versioning, Component substitutability, Version propagation.

1 Introduction

Nowadays, softwares are getting more and more complex. This complexification brings a deep need of managing software at every moment of its life-cycle. Indeed, during its life-cycle, a software may be subject to many changes, which makes versioning one

f the main issues of software evolution management [5]. It is important for software

architects and developers to keep track of those changes since they may occur from early specification steps (e.g., specification) to post-production activities (e.g., software maintenance). However performing changes all along a software life-cycle may imply side-effects which need to be well managed in order to keep track of valid software con-

figurations. Moreover, collaborative developments make versioning systems even more
necessary. In this context of very complex software systems, developers need to be able

to reuse specific versions of libraries /packages /components [13] while users need to maintain up-to-date versions of their applications. A lot of version control approaches have been proposed to track changes within source code, objects, models etc. [3]. How- ever, only a few of these approaches deal with architectural versioning issues. Some of the main issues of architectural versioning are related to the traceability of every architectural decisions that occur during the life-cycle and also to the prediction of versioned architectural artifacts compatibility which is intended to improve reusability and guide architects and developers within software development process. Section 2 introduces a three-level architecture description language (ADL) named Dedal [14], which aims at representing software life-cycle. Next, Section 3 briefly introduces our work around versioning architectures. Finally, Section 4 concludes this paper by giving some perspectives for future works.

2 Covering software life-cycle

In order to cover the entire life-cycle of component-based software, a three-level ADL named Dedal has been proposed. Indeed, Dedal aims at giving component-based representations of the main steps of software engineering. The three levels of Dedal are as Page 13

SLIDE 18

follows: (a) The specification which corresponds to the actual specification stage of the development process. It is composed of component roles that are abstract types which describe the functionalities of the software. (b) The configuration is the representation of the implementation stage of the development process. This architectural level is composed of component classes that realize the component roles from the specification (n to m relation). (c) The assembly corresponds to the deployment stage and is composed of all the component instances that instantiate the component classes. This ADL ensures architectural integrity within and between its abstraction levels by providing two essential properties:

1. intra-level consistency which ensures that all components from an architecture

level are connected to each other, and

2. inter-level coherence which is the description of the relations that exist between the

three levels of Dedal: (i) all the roles which are described in the specification must be realized in the configuration by component classes and (ii) all the component classes from the configuration level must be instantiated in the assembly level. However, the architectural integrity of Dedal may be broken because of a change. Thus, it has to be recovered through propagation mechanisms within between architecture levels. To do so, Dedal has been formalized [9] following principles of Liskov et

al. [8] with the B language [1], and because of an evolution manager, Dedal makes it

possible to automatically calculate an evolution plan that, if it exists, restores architectural integrity after any its abstraction level has been subject to change. This multiple-level evolution management makes necessary to version component- based architectures at several abstraction levels.

3 Versioning three-leveled architectures

Initially, versioning activity aims at keeping an history of changes for representing and retrieving past states of a file and most of the times relies on text-based mechanisms like in version control systems such as Git [12] and CVS [10]. However models and architectures cannot be versioned as text. Surprisingly, little work copes with versioning architectures. Some of the few approaches we can cite are ADLs like SOFA [2], Mae [11] or xADL [4]. However, those ADLs are not covering the whole life-cycle

f software developments as they only provide two abstraction levels and SOFA ADL

does not provides a fine grained enough typing for predicting version propagation. Thus Dedal is the ideal candidate for studying architetural versioning since it covers the entire life-cycle of a software and provides an evolution manager that can be used for studying version propagation in the context of multiple-leveled architecture descriptions. Then, a study on version propagation in Dedal has been lead [6,7] for predicting the compatibility of a new version of a component at any of the three architecture levels in terms of impact the new version may have on the different levels. Figure 1 is an example of version propagation within a very simple base case of a three-level architecture. This example describes a three-level architecture Arch which specification Spec is composed of two component roles R1 and R2. Those component roles are realized in the configuration Config respectively by component classes C1 Page 14

SLIDE 19

Fig. 1: Example of version propagation after a perturbation

and C2 which are themselves instantiated in the assembly Asm respectively by component instances I1 and I2. All the described components are connected through a func- tionality f() with different levels of refinement within the architecture description. To make this figure understandable, we note: (a) T1 T2: T1 is a subtype of T2 or equal to T2 (b) T1 ⌫ T2: T1 is a supertype of T2 or equal to T2 (c) T1 k T2: T1 is not comparable to T2 (d) T2 # T1: T2 replaces T1. In this figure, the reader can see how a new version C

1 of the component C1 of the configuration may impact its own architecture level but

also the whole architecture description. Indeed, C

1 breaks the interface compatibility

with C2 which implies the need of a new version C

2 of C2. In addition, this change

also broke inter-level coherence rules that exist between Spec and Config but also between Config and Asm. Then the change needs to be propagated to the specification and to the assembly in order to recover global architecture integrity. Thus version is propagated to the component roles of Spec and to the component instances of Asm. Then one can feel that the version must be propagated not only to the components of the adjacent architecture levels but also to the architecture levels themselves and even more to the whole architecture. This version propagation allows software architects and developers not only to keep an history of individual components but also to keep histories of architectural level and global architecture. Then it makes it possible of keeping track of valid software configurations that would realize a given specification and which would potentially be instantiated by n assemblies. Software users, architects or developers could therefore adapt architectures according to a desired version of a specification, configuration or assembly.

4 Conclusion and future work

As a result of this first study on version propagation, we could identify multiple criteria for guarantying the integrity of the three architecture levels of Dedal. Then we could Page 15

SLIDE 20

identify non-propagation scenarios but as a consequence we could identify specific condition for intra-level and or inter-level propagation. Thus we are now able to predict the impact that a versioned artifact may have on a three-leveled architecture. Among the future works, it will be important to focus on the representation of versioning concepts. This will imply the definition of a versioning metamodel, and also versioning paterns before we can formalize versioning concepts and automate versioning activity in Dedal architectures.

References

1. Abrial, J.R.: The B Book - Assigning Programs to Meanings. Cambridge University Press

(Aug 1996)

2. Bures, T., Hnˇ

etynka, P., Pl´ aˇ sil, F.: Sofa 2.0: Balancing advanced features in a hierarchical component model. In: Software Engineering Research, Management and Applications. 4th International Conference on. pp. 40–48. IEEE (2006)

3. Conradi, R., Westfechtel, B.: Version models for software configuration management. ACM

Computing Surveys 30(2), 232–282 (1998)

4. Dashofy, E.M., Hoek, A.v.d., Taylor, R.N.: A comprehensive approach for the development
f modular software architecture description languages. ACM Transactions on Software En-

gineering and Methodology 14(2), 199–245 (2005)

5. Estublier, J., Leblang, D., Hoek, A.v.d., Conradi, R., Clemm, G., Tichy, W., Wiborg-Weber,

D.: Impact of software engineering research on the practice of software configuration man-

agement. ACM Transactions on Software Engineering and Methodology 14(4), 383–430

(2005)

6. Le Borgne, A., Delahaye, D., Huchard, M., Urtado, C., Vauttier, S.: Preliminary study on

predicting version propagation in three-level component-based architectures. In: Proceed- ings of the 10th Seminar Series on Advanced Techniques & Tools for Software Evolution (SATToSE) (2017), (to appear)

7. Le Borgne, A., Delahaye, D., Huchard, M., Urtado, C., Vauttier, S.: Substitutability-Based

Version Propagation to Manage the Evolution of Three-level Component-Based Architec-

tures. In: Proceedings of the 29th International Conference on Software Engineering &

Knowledge Engineering (SEKE) (2017), (to appear)

8. Liskov, B.H., Wing, J.M.: A behavioral notion of subtyping. ACM Transactions on Program-

ming Languages and Systems (TOPLAS) 16(6), 1811–1841 (1994)

9. Mokni, A., Urtado, C., Vauttier, S., Huchard, M., Zhang, H.Y.: A formal approach for manag-

ing component-based architecture evolution. Science of Computer Programming 127, 24–49 (2016)

10. Morse, T.: CVS. Linux Journal 1996(21es), 3 (1996)
11. Roshandel, R., Hoek, A.V.D., Mikic-Rakic, M., Medvidovic, N.: Mae—a system model and

environment for managing architectural evolution. ACM Transactions on Software Engineer- ing and Methodology 13(2), 240–276 (2004)

12. Torvalds, L., Hamano, J.: Git: Fast version control system. URL http://git-scm. com (2010),

last visited: 03.05.2017

13. Urtado, C., Oussalah, C.: Complex entity versioning at two granularity levels. Information

systems 23(3-4), 197–216 (1998)

14. Zhang, H.Y., Urtado, C., Vauttier, S.: Architecture-centric component-based development

needs a three-level ADL. In: Babar, M.A., Gorton, I. (eds.) Proceedings of the 4th Euro- pean Conference on Software Architecture. LNCS, vol. 6285, pp. 295–310. Springer, Copen- hagen, Denmark (Aug 2010)

Page 16

SLIDE 21

Contribution to Lessons learned process with application to logistical response in humanitarian emergencies

1,2, Gilles Dusserre1, Sébastien Harispe1, Abdelhak Imoussaten1 and Benoit Roig2

1 IMT Mines Alès, LGI2P, 69 Rue Georges Besse, Nîmes, France

!"#"$%#&%'#($)$#(*+,$%%#-&./--#((#*+-#01-)$#2&'1($-3#*+

10.#%1'4&$56/--1)#2785$2#-91%#-&:(+

2 Université de Nîmes, CHROME, Rue du Dr Georges Salan, Nîmes, France

0#26$)&(6$,8/2$5#-&:(

Abstract. Because of its critical impacts on performance and competitivity,
rganizations is today considered to be an invaluable asset. In this

context, the development of methods and frameworks aiming at improving knowledge preservation and exploitation is of major interest. Within Lessons Learned framework which proposes relevant methods to tackle these challenges , this paper briefly presents our work on an approach mixing Knowledge Representation, Multiple-Criteria Decision Analysis and Inductive Reasoning for inferring general learnings by analyzing past experiences. This is done by studying the similarities of past experiences shared features, patterns and their potential influence on the overall success of cases through the identification of a set of criteria having a major contribution on this success. The considered work will be developed and validated through the scope of a humanitarian organization, Médecins Sans Frontières, with a focus on the logistical response in emergency situations. Keywords: Lessons Learned, Knowledge Representation, Multiple-Criteria Decision Analysis, Inductive Reasoning

1 Introduction

Knowledge is what makes organization go [1]; indeed, because of its critical impacts on their performance and competitivity, most organizations today fully understand the invaluable asset their knowledge represents [1, 2] an asset which, in capital, often contributes defining their values, their

DNA. In this context, the development of methods aiming at improving knowledge

preservation and exploitation is of major interest for organizations. It requires devel-

ping robust approaches dealing with the heterogeneous and complex nature of organ-

knowledge: a complex and weakly defined object most often composed of (i) explicit formalized and codified elements of knowledge, and (ii) tacit elements that are hard to define since usually based on intuition, experience and context [3]. Knowledge capitalization and exploitation have long been critical open challenges studied by several disciplines and communities. In particular, Knowledge Management, defined as n- gible assets [4], presents strategic features of interest for improving knowledge capture, sharing and leveraging. Among various approaches, the Lessons Learned framework proposes relevant methods to tackle this issue through the analysis of past

experiences. Within this framework, we study an approach mixing Knowledge Repre-

sentation, Multiple-Criteria Decision Analysis and Inductive Reasoning for inferring general learnings by analyzing past experiences. In the following sections, we briefly Page 17

SLIDE 22

introduce our positioning; we discuss in particular the main stakes and limitations of existing practices, to next introduce our problematic and the approach we propose to tackle aforementioned limitations.

2 Background and motivation

Lessons learned process (LL) is considered as a Knowledge Management approach to collect, store, analyze, disseminate and reuse tacit experiential knowledge [5]. Learn- ing from experience is a field of growing interest proposing to address major stakes. Indeed, LL mainly aims to avoid loss of knowledge due to turnover, and to prevent the recurrence of undesirable outcomes - or conversely promoting the desirable outcomes - through lessons acquired from past projects [3]. LL is therefore future oriented since it intends to improve the performaprocess-

es. However, the implementation of a LL methodology based upon databases of past

experiences, knowledge capitalization through predefined models, problem solving methods, etc. involves both human and technological challenges [5]. Regarding the technical aspects, the conceptual modeling of knowledge remains a time-consuming and cumbersome process to setup [3]. As an example, the selection

f an appropriate data representation abstraction is not trivial; a balance must be

reached to provide a representation understandable by organizaexperts with no specific skills in data modeling , while improving the data exploitation capabili-

ties. The heterogeneous nature and complexity of knowledge extracted from experi-

ence, as well as its fuzziness, uncertainty, or subjectivity challenge both theoretical and practical limitations of knowledge representation. Lastly, generalization consist- ing in building lessons from capitalized experiences is today not automated, and pri- marily relies on the analyze and expertise of a cross-disciplinary committee [2]. Considering service oriented organizations, particularly in the humanitarian sector, experience capitalization is also a major concern, subject of various guidelines [6]. There, oral testimony is intuitive and settled in their habits; hence, LL set up, as a formal process, is especially challenging. As a consequence, we are here often facing non-systematic data collection or elicitation after projects, as well as manual data

processing. This limits cross-referencing a lot of sources of information and further

limits complex analysis. Discussions with Médecins Sans Frontières (MSF), an international humanitarian organization, have stressed the complexity of project management in this domain. Decisions are often (i) based on intuition relying on expert experience and knowledge; (ii) context-dependent, considering instable environments

f intervention, unpredictable events and singular contexts. Decisions also take into

account inter-related actions: repercussion of choices on the sequence of events. Fi- nally, multiple and sometimes conflicting objectives are often considered. Neverthe- less, despite decision-making complexity, and according to MSF, the logistical strategy of some projects e.g. as different as vaccination campaigns and food distributions could be, to a certain extent, generalized and reproduced analyzing similarities between projects. s- tical response, it remains however extremely challenging to support such entailments estimating/learning projects similarity from the low number of available observed cases. Page 18

SLIDE 23

3 Problem statement and proposed approach

With regard to aforementioned interests and limitations, the aim of our work is to contribute to the LL process, with a central question: how to capitalize and add value to past experiences to turn them into exploitable knowledge and extract lessons. To address this issue, we propose a specific case-based reasoning approach that specifi- cally focuses on linking the performance assessments of similar past experiences to the features impacting them. This study, at a crossroad of various disciplines, is developed through an application case: the logistical response implemented to support ination or food distribution campaigns. Our proposal is next briefly introduced. Hereafter, we refer to a past project experienced by an organization, using the term , i.e. the implementation of an alternative in a given context, towards project

achievement. The alternative denotes a specific set of decisions also commonly

defined as a strategy. The starting point of our work rests upon the availability of information related to cases; usually explicitly registered in several heterogeneous representations, mostly in natural language. However, some elements may remain tacit the preference of the decision-maker, and must be collected, through interviews for example. Due to their lack of structure and formalism, raw data are neither automatically an- alyzable nor exploitable as it. Consequently, the first step of the proposed approach aims at formalizing cases using Knowledge Representation paradigms to build a re- pository of machine-processable cases enabling advanced data analysis. The objective is to store cases using a generic, formal and shared structure, e.g. a tuple (context, alternative, measures1). For this purpose, we study the use of the well-established Resource Description Framework with additional semantics provided by RDFS or OWL profiles (relying on Semantic Web technologies and Description Logics). We then propose to compare cases comparing their contexts and implemented alterna-

tives. Similarities between cases will be estimated through semantic similari-

ty/proximity measures [7]. These similarities will then further be exploited in the last stage of the approach. In the context of LL we need to learn lessons from past experiences, and particularly about choices made in specific contexts. Thus, a posteriori, it aims at determining if choices were good or bad compared to project objectives, and why; we note that the definition of what are the best alternatives naturally depends on different factors enabling to criticize the relevancy of alternatives w.r.t a decision/project context. Indeed, the selection and evaluation of an alternative can only be discussed in a context-dependent manner. Technically, it involves (i) evaluating the overall success of implemented alternatives and (ii) identifying criteria contributing to case success/

failure. We propose to rest upon MCDA methods designed to address these issues [8],

[9]. We underline that made decisions may have consequences according to several points of view. Also, it must be taken into consideration that the criteria are dependent, and sometimes antagonist e.g. reaching a maximum of beneficiaries while minimizing the costs. Similarly, taking into account decision-maker preferences is necessary since decisions made regarding an alternative may differ from one decision-

1 observed measures for the case, being the alternative assessment on evaluation criteria.

Page 19

SLIDE 24

maker to another. These elements are part of the, so-called in MCDA, preference model; different methods exist to define it (e.g. Learning process, disaggregation). However, since we also intend to identify the subset of criteria with the major contribution on the success of cases, it constrains the method choice. Our LL approach aims at inferring general learnings from formalized cases. To address this purpose, we propose, in this last step, to highlight the link between the previous outcomes: characterization of the similarity of cases and associated success estimation, as well as the determination of a subset of criteria with a major contribution to this overall success. We intend to identify, within a case, which elements from its context or its alternative have a significant influence on the values observed for each contribution criterion. Hence, we propose to verify if cases appraised as success- ful share some features. The observations and similarity measures, previously assessed, are now exploited to identify potential patterns: shared features of cases regarding either the context or the alternative. Then, evidencing the link between the shared features identified and the evaluated success of a case will enable to state lessons learned. Thus, to a further extent, the lessons learned will provide guidance to decision-makers in the selection of a strategy to adopt for future projects, that is, which strategy will be the most efficient and the most appropriate for a given context.

4 Conclusion and perspectives

Our study, within the framework of LL methods, aims at inferring general learnings by analyzing several past experiences of an organization. Future work will address case representation issues in order to improve exploitation capabilities of cases. It will focus on the representation of four real cases from MSF past projects. Then, an exploitation phase will involve similarity and pattern discovery, as well as, the identification of contribution criteria through the evaluation of cases success.

References

[1]

T. H. Davenport and L. Prusak,

they know. Harvard Business School Press, 1998. [2] ation in experience feedback processes: An ontology- Comput. Ind., vol. 59, no. 7, pp. 694710, 2008. [3]

Int. J. Hum. Comput. Stud., no. 51, pp. 567598, 1999.

[4] Expert

Syst. Appl., vol. 20, no. 1, pp. 16, 2001.

[5]

R. Weber, D. W. Aha, and I. Becerra-

Expert Syst. Appl., vol. 20, no. 1, pp. 1734, 2001. [6] Traverses, 2004. [7]

S. Harispe, S. Ranwez, S. Janaqi, and J. Montmain, Semantic similarity from natural

language and ontology analysis. Synthesis Lectures on Human Language Technologies, 2015. [8]

B. Roy, , Economica. Paris, 1985.

[9]

R. L. Keeney and H. Raiffa, Decision analysis with multiple conflicting objectives,

Wiley & So. New York, 1976.

Page 20

SLIDE 25

Towards the evaluation of political risks related to radioactive waste management using natural language processing methods

Jean-Christophe Mensonides1, S´ ebastien Harispe1, Jacky Montmain1, and V´ eronique Thireau2

1 ´

Ecole des mines d’Al` es, LGI2P Research Center, Parc scientifique G. Besse 30035 Nˆ ımes Cedex 1, France firstname.lastname@mines-ales.fr

2 Universit´

e de Nˆ ımes, CHROME, Rue du Dr Georges Salan, Nˆ ımes, France veronique.thireau@unimes.fr

Abstract. We propose to study the development of automatic analysis

techniques based on Natural Language Processing for helping studying political risks related to radioactive waste management in France. We focus more particularly on detecting relevant information from large text corpora analysis in order to characterize power games in that domain. To this aim we study the problems of detecting (i) links between news articles and confronting groups of actors (i.e. to assess which group the conveyed information supports - pros or cons roughly speaking), as well as (ii) topics of interest mentioned in those articles. Several supervised approaches have been analyzed for those tasks. This paper presents results obtained so far in that project as well as working perspectives.

1 Introduction

France’s energy policy is heavily based on nuclear energy, making France one of the countries in which the public debate related to radioactive waste management is the most prevalent, and of upmost interest. More generally, worldwide, public health and environmental protection are major challenges for our societies - radioactive waste management being a critical topic in that domain. Traditionally, power games analysis in public decision making (in France), such as power games encountered in the radioactive waste management domain are studied manually. Such studies are of particular interest to better understand contemporary trans- formations related to public decision making. As an example, in France, OCN3 aims at monitoring, collecting, and analyzing relevant information and analyses related to conflicts in that domain. OCN and similar groups analyses are made manually, for instance by analyzing large corpora of news articles in order to detect and better understand power games among concerned groups of actors. We propose to study the use of Artificial Intelligence based approaches, and to develop automatic analysis techniques based on Natural Language Processing

3 Observation et analyse des Conflits dans l’industrie nucl´

eaire civile: http://ocn.unimes.fr

Page 21

SLIDE 26

(NLP), in order to enable the analysis of large text corpora. We focus more particularly on detecting relevant information from large text corpora analysis in order to characterize power games. To this aim we study the problems of detecting (i) links between news article and confronting groups of actors (i.e. to assess which group the conveyed information supports - pros or cons roughly speaking), as well as (ii) topics of interest mentioned in those articles. We consider a simplified modeling where two groups of actors of interest in the power game study confront each other [1]: – regulators try to maintain the existing system as it is. Simply put, these actors seek to promote the development of nuclear industry in France. The French National Radioactive Waste Management Agency (ANDRA) is an example of regulator player. – balancers try to alter the existing system’s trajectory. Simply put, they act with the aim of leading France towards the abandonment of nuclear industry. The network ”R´ eseau Sortir du nucl´ eaire” is an example of a balancer player. We aim at providing macroscopic elements in order to analyze power games evolution between these two groups of actors. Our main focus has been to highlight:

1. the news articles publication frequencies according to which group their con-

tent benefits. That is, we consider a news article to be regulator if its content makes the game of regulators (i.e, after reading it, a ”regular citizen” would be favorably influenced toward the nuclear industry), and vice-versa.

2. the frequencies of appearance of some predefined topics in news articles

related to nuclear waste management. Those two tasks are modeled as classification problems and are further introduced in the following sections.

2 Detecting regulator/balancer news articles polarity

We are interested in the task of classifying news articles into two classes, (regulator and balancer), in a context of supervised learning. Supervised text classification has been widely studied in NLP, and several methods achieved good performance on related tasks (e.g. on opinion mining, sentiment analysis): Naive Bayes (NB) [2], Support Vector Machines (SVM) or NBSVM [3] to cite a few. Most approaches consider the hypothesis that target labels are independent of words order and represent documents as bags of n-grams. They are high-bias low-variance models and tend to generalize well, even with few training data. Recently, recurrent neural networks (RNN) have also been used with success on text classification [4, 8]. Compared to aforementioned approaches, these models consider documents as sequences of words (order matters) and encode them into fixed length vectors. They are generally high-variance models and tend to

verfit quickly on small training sets. Experiments evaluating the performance
f several approaches on our problem have been performed.

Page 22

SLIDE 27

Evaluation has been made considering 630 labeled french news articles from 1988 to 2017, splitted in training and test sets. These articles range from 30 to 8400 words, with an average length of 930 words. The BN and NBSVM approaches have been tested on our dataset, along with a 1-layer Gated Recurrent Units (GRU) [5] network and a 1-layer bidirectional GRU network. Note that we used GRU networks with gradient clipping to reduce the vanishing/exploding gradient issue commonly observed in the training of recurrent neural networks. The accuracy results obtained with the different approaches are the following: NB 0.776; NBSVM 0.804; 1-layer GRU 0.593; 1-layer bidirectional GRU 0.644. Best results have therefore been obtained using NBSVM, followed by NB. This seems not surprising, since the (reduced) size of the training dataset implicitly tends to favor high-bias low-variance models. Indeed, RNN-based approaches used are quickly overfitting, even with the use of dropout [6] (a regularization technique) and batch normalization [7].

3 Detecting topics of interest

Beyond the regulator/balancer classification, we are also interested in the task

f detecting topics of interest in news articles; topics are here predefined and
rganized into a taxonomy. Considering that an article can be associated to

several topics, this problem can be modeled as a supervised learning problem called hierarchical multi-label classification [9]. Formally, we consider a taxonomy of labels as a directed acyclic graph representing a partially ordered set G = (Y, ⇥), with Y the set of possible labels. We aim at defining a model f : X ⌅ P(Y) enabling to distinguish a subset Yx Y of labels tagging a text x ⇧ X. Naturally, we are particularly interested in models providing hierarchically consistent results w.r.t to G, i.e. if (a, b) ⇧ Y2 and a ⇤ b, any text associated to label a should necessarily be associated to label b. Considering a specific class of models depending on a set of parameters θ, the hierarchical multi-label classification consists of finding a set of parameters (i) providing a hierarchically consistent multi-label b y = f(x; θ) for any x ⇧ X, which (ii) minimizes the following risk function (considering a discrete setting): R(f) = 1 n

i=1

∆(y(i), f(x(i); θ)) with n = |X|. ∆ : P(Y), P(Y) ⌅ R is a loss function measuring the discrepancy between a prediction b y and its true multi-label y. Intuitively, loss functions should respect the following properties [10]: (i) partially correct solutions should result in lower losses than completely incorrect solutions, (ii) correctly classifying

ne level down should result in a lower loss than staying at the parent node, and

incorrectly classifying one level down should result in a higher loss than staying at the parent node, and finally (iii) errors made at higher level of the hierarchy should be punished more heavily. We are currently studying the various approaches focusing on performance and complexity analysis. Page 23

SLIDE 28

4 Positioning

We propose to use AI-based techniques for studying the evolution of power games in public decision-making related to radioactive waste management in France, through automatic news article analysis. We have divided this task into two learning problems related to text analysis: (i) regulator/balancer (binary) classification and (ii) hierarchical multi-label classification for topic tagging. Considering the empirical and theoretical analyses performed so far, we propose to study how to make RNN-based classification approaches more competitive for solving these problems - indeed, our results have shown their limitations in our context despite good performances for solving other complex problems. Those results are mainly explained by the lack of labeled data; we are in particular studying the consideration of additional unlabeled data as well as prior knowledge in the form of terminologies and ontologies to improve those mod-

els. We are for instance currently studying how to pre-train RNN parameters

by training unsupervised RNNs, e.g. using sequence auto-encoders that learn to encode and decode their inputs [8].

References

1. Perroux F., Unit´

es math´ ematiques nouvelles, R´ enovation de la th´ eorie de l’´ equilibre g´ en´ eral, Dunod, 1975, 274 pages.

2. Mehran Sahami, Susan Dumais, David Heckerman, and Eric Horvitz. A bayesian

approach to filtering junk e-mail. In AAAI-98 Workshop on Learning for Text Categorization, 1998.

3. Wang, Sida and Manning, Chris D. Baselines and bigrams: Simple, good senti-

ment and text classification. In Proceedings of the 50th Annual Meeting of the Association for Computational Linguistics, 2012.

4. Irsoy, O., & Cardie, C. (2014). Opinion Mining with Deep Recurrent Neural Net-
works. In Proceedings of the 2014 Conference on Empirical Methods in Natural

Language Processing (EMNLP), pp. 720728, Doha, Qatar. Association for Com- putational Linguistics

5. Cho, K., van Merrienboer, B., Gulcehre, C., Bougares, F., Schwenk, H., and Bengio,
Y. (2014a). Learning phrase representations using RNN encoder-decoder for sta-

tistical machine translation. In Proceedings of the Empiricial Methods in Natural Language Processing (EMNLP 2014).

6. Srivastava, N., Hinton, G., Krizhevsky, A., Sutskever, I. & Salakhutdinov, R.

Dropout: a simple way to prevent neural networks from overfitting. J. Machine Learning Res. 15, 19291958 (2014).

7. Tim Cooijmans, Nicolas Ballas, Cesar Laurent, and Aaron Courville. Recurrent

batch normalization. arXiv preprint arXiv:1603.09025, 2016.

8. Dai, Andrew M and Le, Quoc V. Semi-supervised sequence learning. arXiv preprint

arXiv:1511.01432, 2015.

9. Silla, C.N. and Freitas, A.A. A survey of hierarchical classification across different

application domains. Data Mining and Knowledge Discovery, 22(1-2):3172, 2010.

10. Kiritchenko, S., Matwin, S., Nock, R., & Famili, A. F. (2006). Learning and eval-

uation in the presence of class hierarchies: Application to text categorization. In Proceedings of the 19th Canadian conference on AI (AI2006) (pp. 395406).

Page 24

SLIDE 29

Formalization and evaluation of non-functional requirements for system of systems engineering: Application to Resilience

Behrang MORADI1, Nicolas DACLIN 1 and Vincent CHAPURLAT1

1 Laboratoire de Génie Informatique et d’Ingénierie de Production – LGI2P

Ecoles des Mines d’Alès, Parc Scientifique Georges Besse, France surname.name@mines-ales.fr

Abstract :

This paper introduces the very first developments of a method for the specification, formalization and evaluation

f resilience in the context of System of Systems Engineering (SoSE). The developed method is based on three

working hypotheses. On the one hand, we study and analyze, several resilience’s metrics and indicators as well as the relationship between resilience and other requirements in its environment. On the other hand, the method relies on the principles of the multi-viewpoint modelling of complex systems applied on the field of SoSE. The results of the first working hypothesis have to be consistently related to the multi-viewpoint modeling context. The expected benefit of such method is to allow the modeling complex systems as SoS and characterize the resilience in order to enable its evaluation. Keywords: Formalization, Evaluation, Simulation, System of systems engineering, Resilience engineering, system of system architecture, Non-functional property

1. Introduction

“A system of systems (SoS) is an integration of a finite number of constituent systems which are independent and operable, and which are networked together for a period of time to achieve a certain higher goal.” [1] A SoS is composed of separate constituent systems and have a the following five characteristics [2] Operational independence, Managerial independence, Emergent behavior, Evolutionary development, Geographical

distribution. Many surrounding systems can be seen as SoSeg, air transport networks, Franco-German high

speed train networks, electric power distribution networks). Figure 1. shows an example of a SoS[3].

Fig.1. Air transport system, [3]

To fullfil its mission adequately, a SoS must satisfy functional and non-functional requirement. Among the non- fonctional requirements a set is called “-illities” and represent “the desired properties of systems, such as flexibility or maintainability (usually but not always ending in “ility”), that often manifest themselves after a system has been put to its initial use. These properties are not the primary functional requirements of a system’s performance, but typically concern wider system impacts with respect to time and stakeholders than are embodied in those primary functional requirements.” (e.g. interoperability, robustness, flexibility, resilience…), [4] [5] [6]. Currently there are between 74 and 83 '-ilities' identified in the literature The figure .2. presents a set

f identified 'ilities' and their relationship as defined by Ross [4]. Our research work focuses on the study of one
f these non-functional properties such asthe resilience. Resilience in the context of SdS engineering is an

important property because it must be mastered and maximized by systems in order to effectively cope with disruptive events and maintain acceptable levels of services and performance. It exists several definitions of

resilience. Generally speaking, the resilience is defined as as “the ability of the system to resist, absorb, recover
r adapt to disturbances and diminish the consequences as well as to recover quickly and effectively”. [7][8][9].

Page 25

SLIDE 30

Lastly, Resilience is practiced in several of application area as critical infrastructure, security etc. [9][10][11] [12][13].

Fig.2. Correlation network of ilities based [4]

The current problems of this research work are to define, specify and evaluate the resilience. Within this work

ur problem how and why and at what level to evaluate resilience by considering (analyzing) the formalization

its relationship with other non-functional properties, '-ilities'.

2. State of the Art

2.1. Resilience As claimed in [14] “conventional risks management approaches are based on hindsight and emphasize error tabulation and calculation of failure probabilities, Resilience Engineering looks for ways to enhance the ability at all levels of organizations to create processes that are robust yet flexible, to monitor and revise risk models, and to use resources proactively in the face of disruptions or ongoing production and economic pressures “[14]. There are currently several definitions of resilience

Resilience is “the ability to anticipate, prepare for, respond to, adapt to disruptions and to mitigate the

consequences as well as to recover in timely and efficient manner including preservation restoration of services” [15];

Resilience is “the ability of the system to withstand a change or a disruptive event by reducing the

initial negative impacts (absorptive capability), by adapting itself to them (adaptive capability) and by recovering from them (restorative capability)”[15];

Resilience is “the ability of an entity to recover from an external disruptive event”[21];

They all suggest that resilient systems are able to manage disastrous situation due to several capacities (discussed below) requested throughout the classical phases of the disaster management lifecycle: Preparation, Prevention, Response and Recovery as illustrated in Fig..

Fig.3. The classical disaster management lifecycle, [16].

Page 26

SLIDE 31

Zone 1 – Prevention: this phase aims at identifying and minimizing the risks posed by the CI, its equipment and fittings, and the natural hazards of the environment [16]. In other words, the goal is to prepare a system for, and increase its robustness [17] against, the most probable and anticipated threats by implementing different techniques (e.g., installing various safety barriers, security protocols, disaster management resources, fail-safes

r fail-secures). .

Zone 1 – Preparation: this phase is defined as a “getting ready to cope” [16]. It consists in:

developing a written preparedness, response and recovery plan;
keeping the plan up-to-date, and testing it;
keeping together supplies and equipment required in a disaster and maintain them;
establishing and training a disaster response team;
distributing the plan and documentation to appropriate locations on- and off-site; instituting procedures

to notify appropriate people of the disaster and assemble them rapidly; Zone 2 – Response: this zone takes place when a disaster happens [16]. It consists in following the previously established emergency procedures. Two sub-zones are distinguished:

Absorption;
Adaptation;

Zone 2.1 - Absorption: the absorptive capability refers to an endogenous ability of a system to reduce the negative impacts caused by disruptive events and minimize consequences with less effort [17] [18]. The capacity of a system to absorb negative effects can be improved during the Prevention phase by increasing system’s robustness. Alternatively, the absorption can be enhanced by improving system redundancy, which provides a way for the system to operate [17]. This zone takes place when a disruption event starts and lasts until the duration of the initial and propagated effects ends. Zone 2.2 - Adaptation: the adaptive capability refers to an endogenous ability of the system to adapt to disruptive events through self-organization in order to minimize consequences and it can be enhanced by using emergency systems [17]. According to [19], the ability of a system to adapt is related to three overarching concepts: 1) vulnerability to unpredictable shocks; 2) the resources available to a system to help it change—or “wealth”; 3) the “internal controllability” of relationships in a system—rigidity vs. flexibility. This zone takes place when a disruption event starts and lasts until the system recovers. Zone 3 - Recovery: this zone is defined as a “getting back to normal” phase by [16]. The restorative capability refers to an ability of the system to rapidly be repaired and returned to a as much as possible normal and a reliable functioning mode that meets the requirements for an acceptable and desirable level of quality of service and expected control [17]. According to [20], there are four resilience strategies to speed up recovery: 1) removing operating impediments (debris removal and related complications, and streamlining paperwork for insurance claims and government assistance); 2) management effectiveness (skills that promote restoration, repair and reconstruction); 3) speeding restoration (alternative means of access to repair sites and incentive contracts); 4) Input substitution, import substitution, inventories, (materials and labor needed for repair activities rather than normal production operations). 2.2. Resilience Metric Numerous works related to the evaluation of the resilience are existing in the literature. Among these works we can mention [21] which propose a metric method of resilience that can be associated in three types such as:

the focus attribute parameters, which usually consist of indices that are based on subjective

assessments;

the indicators built on databases, which quantify the system attributes that contribute to resilience;
the performance-based methods, which measure the consequences of system disturbances and the

impact that system attributes have on mitigating these consequences. On his side, Royce's [18], states the measurement of resilience is a function of the 3 capacities (absorption, Page 27

SLIDE 32

adaptation and recovery) as well as recovery time, through the measurement of the performance and its

evolution. The above resilience factor is then obtained:

!! !!, !!, !!, !! = !! !! !! !! !! where !! = !!!!

∗) exp −! !! − !! ∗

!"# !! ≥ !!

∗ !! !!

∗ !"ℎ!"#$%!

Where : Sp Recovery factor Fo Stable system performance Fr Performance after steady state a parameter controlling decay in resilience attributable to time to new equilibrium tδ slack time tr time to final recovery tr * time to complete initial recovery actions tδ Time of disturbance Fd Performance immediately after the disturbance As last example of resilience measurement, Cox in 2011 [20] proposes an operational measure of direct static economic resilience (DSER), which is the measure for which the reduction in estimated direct production differs from the likely maximum potential reduction given an external shock. He provides admittedly crude mathematical definitions of resilience. Adapting an established definition, direct static economic resilience (DSER) refers to the level of the individual firm or industry and corresponds to what economists refer to as ‘‘partial equilibrium’’ analysis, or the operation of a business or household entity itself. An operational measure

f DSER is the extent to which the estimated direct output reduction deviates from the likely maximum potential

reduction given an external shock, such as the curtailment of some or all of a critical input: DSER = %∆DYm ± %∆DY %∆!"# Where, ∆DYm is the maximum percent change in direct output and %∆DY is the estimated percent change in direct

utput.
3. Work-hypothesis

3.1. Hypothesis1. We consider the representation of the assessment of resilience by defining and analyzing the set of '-ilities'. It is a matter of analyzing and formalizing the relationship between resilience and other “-ilities”. Figure 4. In this hypothesis, we focus on the analysis of the resilience and its environment and highlight the various components to consider in order to evaluate resilience. The four componnets considered in our work are (also shown on Figure 4):

1. The dependence (1). It means to identify if the resilience depends upon other “-ilities”. Some

dependence are currently identified and defined in [4]. However, as claimed in [4] some could be existing but not yet identified. In order to evaluate the resilience based on the study of the set of ilities, it is required to identify all dependence between resilience and other “-ilities”.

2. The orientation (2). It means to define if the dependence is unidirectional (ex. quality → resilience) or

bidirectional (ex. quality ↔ resilience). This orientation must be considered “from” resilience “to” another “-ilities” as well as “from” another “-ilities” to resilience.

3. The intensity (3). The intensity defines in which measure a variation of a given “-ilities” can impact a

given “-ilities” in relation. The more important the intensity the more this relation must be considered in the evaluation. In other word an “intensity coefficient” related to the relation between “-ilities” must be defined and integrated to the evaluation of the resilience.

4. The chain (4). The chain reprensents a relation “starting from” the resilience and returning to the

resilience via another ilities (resilience → safety → sustainability). In this work we consider and limit a chain to a path with 3 “-ilities”. Page 28

SLIDE 33

Fig.4. Correlation network of ilities based [4]

3.2. Hypothesis.2. In order to evaluat the resilience, our method is based on the dynamic of the resilience (anticipation, impairment and recovery). To do this we mapped the '-ilities' and their dependencies in each zone of the dynamic of resilience (ex. flexibility can be expected during the prevention), Figure.6. The objective is to get a resilience indicator for each zone in order to identify precisely a problem of resilience. In the end, the aggregation of each indicator will provide an evaluation of the resilience of the SoS.The mapping allows us to choose indicators of resilience associated with these -'ilities'. This method requires the use of concepts and metrics to evaluate the resilience of SdS.

Fig.6. Framework of evaluation of resilience

3.3. Hypothesis3. The definition and the formalization of the rules, methods and indicators for assessing resilience take into account other '-ilities' but also the different (point of view) of SdS; Functional, operational, organizational and

physical. Now we adapt and integrate each indicator in the point of view studied. It is considered to define and

formalize the indicator as a mathematical tool, above; ! ! = !(S1, S2, S3, P1, P2, P3, -ilities, situation, environment, emergence, value) (1) P S = ( !"

! ! ! !!!

, !(!")

! !

!"#$%(!")

! !

), (2) [n: number of sous-system, S: system, P: property]

Fig.6. Framework of evaluation of resilience

Page 29

SLIDE 34

4. Conclusion

The work presented aims to evaluate the resilience relying on first, the dynamic of the resilience and then, on the study of the resilience as an “-ilities” belonging to a wider set of ilities. Three main hypothesis are highlighted including items to consider and to develop in the next steps of thi research. Futur work is related to the validation of the hypothesis and the development of indicators to evaluate resilience. References [1] Sauser, B., Boardman, J., & Gorod, A. (2009). System of systems management. Chap, 8, 191-217.1 [2] Maier, M. (1998). Principles of System of System Architecture. [3] Blanchard, 2011 [4] De Weck, O. L., Rhodes, D. H., & Ross, A. M.: Investigating relationships and semantic setsamongst system lifecycle properties (ilities). 3rd International Engineering Systems Symposium, TU Delft, the Netherlands (2012). [5] ESD Terms and Definitions (Version 12), ESD Symposium Committee, October 19, 2001, Massachusetts Institute of Technology Engineering Systems Division - Working Paper Series, SD-WP-2002-01 [6] Adam M. Ross and Donna H. Rhodes, Towards a Prescriptive Semantic Basis for Change-type Ilities, 2015 Conference on Systems Engineering Research, Procedia Computer Science 44 (2015) 443 – 453 [7] Haimes YY.On the definition ofresilience in systems. Risk Analysis 29(4) pp.498–501 (2009). [8] Resilience Engineering in Practice: A Guidebook, E.Hollnagel, J.Paries, D.Woods, J.Wreathall Eds., ASHGATE publishing company, ISBN: 978-0-4094-1035-5, 2011 [9] Royce F., Behailu B.: A metric and frameworks for resilience analysis of engineered and infrastructure

systems. Reliability Engineering and System Safety 121 pp.90–103 (2014).

[10] Christophe Dony, Jørgen Lindskov Knudsen, Alexander B. Romanovsky, and Anand Tripathi, editors. Advanced Topics in Exception Handling Techniques, volume 4119 of Lecture Notes in Computer Science. Springer, September 2006 [11] Mark A. Pflanz, Alexander H. Levis, On Evaluating Resilience in C3 Systems, INSIGHT April 2015, Volume 18, Issue 1, INCOSE [12] Alfred R. Berkeley, Mike Wallace, 2010. A Framework for Establishing Critical Infrastructure Resilience Goals Final Report and Recommendations by the Council-National Infrastructure Advisory Council. [13] Zoe Andrews, John S. Fitzgerald, Richard Payne, Alexander Romanovsky: Fault modelling for systems of

systems. ISADS 2013: 1-8

[14] Etkin, David. 2014. Disaster Theory: An Interdisciplinary Approach to Concepts and Causes. Butterworth- Heinemann. [15] Cutter, Susan L, Joseph A Ahearn, Bernard Amadei, Patrick Crawford, Elizabeth A Eide, Gerald E Galloway, Michael F Goodchild, et al. 2013. “Disaster Resilience: A National Imperative.” Environment: Science and Policy for Sustainable Development 55 (2): 25–29. doi:10.1080/00139157.2013.768076. [16] UNESCO. 2017. “Disaster Planning: Prevention, Preparedness, Response, Recovery.” [17] Nan, Cen, and Giovanni Sansavini. 2017. “A Quantitative Method for Assessing Resilience of Interdependent Infrastructures.” Reliability Engineering and System Safety 157 (January). Elsevier Ltd: 35–53. [18] Royce, Francis, and Bekera Behailu. 2014. “A Metric and Frameworks for Resilience Analysis of Engineered and Infrastructure Systems.” Reliability Engineering and System Safety. [19] Holling, Crawford S. 2001. “Understanding the Complexity of Economic, Ecological, and Social Systems.” Ecosystems 4 (5). Springer: 390–405. [20] Cox, Andrew, Fynnwin Prager, and Adam Rose. 2011. “Transportation Security and the Role of Resilience: A Foundation for Operational Metrics.” Transport Policy 18 (2). Elsevier: 307–17. [21] Devanandham, Henry, and Emmanuel Ramirez-Marquez Jose. 2012. “Generic Metrics and Quantitative Approaches for System Resilience as a Function of Time.” Reliability Engineering & System Safety 99: 114– 122. Page 30

SLIDE 35

! "!

!"#$%&'()*'%+(,$-.&/"&0'(1.)$0'(',.0.(%$ !"#$%&'(")*!"!!"#$%&'!()*+,"&-.&!!!"#$%#&'% !"#$%&'()!*!"#$%"&'(&%)#"'*"!"#$%$&'!! !""#$%&'(%!!

#$%&'()!*$)%&'&+! ,-.&/012$/2!#3%4(!5$)+631)!3)7!8/3)9$12!:/$-22&+! ;%$'&!7&2!61)&2!7<='>2!?!@ABC*DE5=!F&2&3/%G!H&)+&/I! *3/%!2%1&)+1J1K-&!AL!M&22&I!NOONP!QR6&2I!8/3)%&! E631'S!TJ1/2+)36&L)36&UV61)&2W3'&2LJ/!

!"#$%&'(#)%"*

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`L!! ! 5(! *Ga! 2-Xb&%+! +34&2! .'3%&2! 1)! +G&! /&+31'! [$/'7^! 1)7&&7! :F8! F&+31'! 12! 3! %$6.-+1)Y! 2&/01%&! %$6.3)(!J$/!/&+31'&/2L!:G&!X321%!17&3!12!+$!+/3)2'3+&!.&/J$/63)%&2!63)3Y&6&)+!+&%G)1K-&2! 3'/&37(!3../$0&7!1)!1)7-2+/13'!%$)+&Z+2!1)+$!+G&!/&+31'![$/'7L!:G&!3/&32!$J!16./$0&6&)+!J$/! /&+31'&/2! 3/&! 6-'+1.'&S! '$[&/1)Y! +G&! 2+$%42I! $.+161c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a! +G&212! J1/2+! 3162! 3+! 7&21Y)1)Y! 3)7! 322&221)Y! +G&! %$)%&.+! $J! $.+161c&7! 322$/+6&)+! +G3+! 2G$-'7! 63Z161c&! +G&! %$)212+&)%(! X&+[&&)! +G&! X&2+! 2&''2! 3)7! +G&! X&2+! %-2+$6&/2^!+G&)I!+G&!17&3'!322$/+6&)+!2G$-'7!2-..$/+!2+$%42!$.+161c3+1$)L!! ! :G&! )$+1$)! $J! 17&3'! 322$/+6&)+! 2&&62! +$! X&! &22&)+13'! J$/! /&+31'&/2<! .&/J$/63)%&! 21)%&! 1+! 2()+G&+1c&2! +G&! &JJ1%1&)%(! $J! &Z%G3)Y&2! X&+[&&)! +G&! %-2+$6&/2<! 7&63)7I! +G&! /&+31'! 3)7! 1+2! 2-..'(! %G31)L! ! :G12! %G3''&)Y&! %$)212+2! 1)! 17&)+1J(1)Y! +G&! 6$2+! .&/J$/63)+! 3/+1%'&2! J$/! 3! G(.&/63/4&+!$)!$)&!G3)7I!3)7!+G&!X&2+!%-2+$6&/2!$)!+G&!$+G&/!G3)7L!:G&2&!)$+1$)2!$J!dX&2+e! 3/+1%'&2!$/!dX&2+e!%-2+$6&/2!X$+G!7&.&)7S! W $)! +G&! %$66&/%13'! .$'1%(! $J! +G&! /&+31'&/! +G3+! %$//&2.$)72! +$! G12! 2-Xb&%+10&! 03'-&! 2(2+&6!3)7!2G$-'7!X&!6$7&'&7!32!3!./&J&/&)%&!6$7&'^! W $)!+G&!2+$%4!3)7!2-..'(!%G31)!/&3'!'1J&!%$)2+/31)+2L!!

Page 31

SLIDE 36

! "! ! #$%!&'%()!(**+,-.%/-!&**0%!/+-!+/)1!2+/2%,/*!&3!&'%/-&41&/5!6%,4+,.(/-!6,+'02-*7!&&3!'%-%2-&/5! 8('!+/%*!-$(-!*$+0)'!8%!,%.+9%'!80-!()*+!&&&3!&'%/-&41&/5!6%,4+,.(/-!6,+'02-*!+4!-$%!8,(/'!

$(-!(,%!/+-!6,+6+*%'!&/!-$%!20,,%/-!(**+,-.%/-!+4!-$%!$16%,.(,:%-!(/'!*$+0)'!8%!(''%'!-+!
$%!+44%,;!#$%*%!6&%2%*!+4!&/4+,.(-&+/!.(1!2+.%!4,+.!%<-%,/()!'(-(!=>(,:%-!?(-(7!@6%/!?(-(7!

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

!"#$%&'()'#*"'+,(-&".$#%/''

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

$%! -(<+/+.1! 8(*%'! *&.&)(,&-1! +4! 6,+'02-*! -+! '%-%,.&/%! -$%! 8%*-! (**+,-.%/-;! M)0*-%,&/5!

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

Page 32

SLIDE 37

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

!"##$%&'(&)&$'

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

Page 33

SLIDE 38

! "! #$%&'('&')'%#*! +$%,-).,#! /'0%! 1%'23)%4! 5%-67! +,4%,%4! 1%'23)%4! 5%-6! -64! 83.9:%)! ;6)%2,-/!

,%!<%'62!#):4'%4!).!'4%6)'(=!)3%!<%#)!>.4%/!(.,!.:,!$:,$.#%*!!

!"#$%&'("#)*+,'-+$.(/+0

! 1%!3-?%!&.>$-,%4!)3%!,%$-,)')'.6!.(!$,.4:&)#!<%)@%%6!#).,%#!(,.>!)3%!#->%!<,-64!:#'62! A->%,-:BC%?%6#3)%'6D#!4'#)-6&%*!;)!3-#!<%%6!:#%4!).!'4%6)'(=!4'#&,%$-6&'%#!<%)@%%6!#).,%#D! <%3-?'.,#!-64! >'##'62! $,.4:&)#! )3-)! &.:/4! <%! ,%/%?-6)! ).! '>$,.?%! )3%! -##.,)>%6)! .(! .)3%,! #).,%#!.(!)3%!<,-64*!1%!3-?%!$,.$.#%4!&:#).>%,#D!-64!$,.4:&)#D!&/:#)%,'62!(:6&)'.6#!:#'62!

$$,.E'>-)%!,%-#.6'62!<-#%4!.6!-!)-E.6.>=!.(!$,.4:&)#*!F//!)3-)!@.,0!'#!-/,%-4=!-?-'/-</%!'6!

GHID#!-$$/'&-)'.6#*!! ! 1%! 3-?%! (.,>-/'J%4! )3%! >-)3%>-)'&-/! $,.</%>! ).! '4%6)'(=! -#! @%//! -#! $.##'</%! ,%)-'/%,#D! ,%9:',%>%6)#! ).! 2'?%! )3%>! -4-$)%4! #./:)'.6#! ).! )3%! '4%-/! -##.,)>%6)! $,.</%>*! G3'#! '##:%! ,%9:',%#!/.)#!.(!/')%,-):,%!'6!/.)#!.(!4.>-'6#!K&/:#)%,'62!-64!&/-##'('&-)'.6!)%&36'9:%#7!.6)./.2=!

64! #'>'/-,')=! >%-#:,%#7! $,%(%,%6&%! >.4%/#! -64! >:/)'! &,')%,'-! )%&36'9:%#7! %)&*L*! 1%! -,%!

&:,,%6)/=! @.,0'62! .6! 3.@! ).! '6),.4:&%! $,%(%,%6&%! >.4%/#!

$,.4:&)#M#).,%#M#:$$/'%,#M<,-64!&.6#),-'6)#!'6).!)3%!'4%-/!-##.,)>%6)!'4%6)'('&-)'.6!$,.&%##*!

1+$.&,+'0)02+3+,+#$+'0

8/:#)%,'62! N! O-@-6! C'62,-#*! PQ6#:$%,?'#%4! H.:23! R%)! 8/-##'('&-)'.6! Q#'62! S-#T! KU.:,6-/! .(! ;6)%//'2%6)! ;6(.,>-)'.6!

R=#)%>#7!F:2:#)!VWWX7!Y./*!!"7!;##:%!Z7!$*!VX[\VV]L!

N! F#>-! F>>-,! %)! -/*! PHO^5_! G3%! H.:23! O.##'<'/'#)'&! ^B5.4%#T! K5;1F;! VWXZ! O,.&%%4'62#! .(! )3%! `)3!

;6)%,6-)'.6-/!1.,0#3.$!'6!F,)'('&'-/!;6)%//'2%6&%!\!Y./*!#$%!7$*!XZBV"L!

N! O-@-6! C'62,-#! %)! -/*! P8/:#)%,'62! a=$%,>-,0%)! 8:#).>%,#! Q#'62! H.:23! R%)! b-#%4! ^.3.6%6! c%)@.,0#T!

K8.>$:)%,!R&'%6&%7!Y./*!$#%!7!$*!XdeBX`Z!f!+&).<%,!VWWZL!

N! O-@-6! C'62,-#7! 83-4! 1%#)*! P;6)%,?-/! #%)! &/:#)%,'62! .(! @%<! :#%,#! @')3! ,.:23! ^B>%-6#T! KU.:,6-/! .(!

;6)%//'2%6)!;6(.,>-)'.6!R=#)%>#7!U:/=!VWW"7!Y./*!$&7!;##:%!X7!$*![\XdL!

N!O-@-6!C'62,-#7!%)!-/"!P;)%,-)'?%!>%)-B&/:#)%,'62!)3,.:23!2,-6:/-,!3'%,-,&3=!.(!3=$%,>-,0%)!&:#).>%,#!-64!

$,.4:&)#T!K;6(.,>-)'.6!R&'%6&%#7!Y./*!$'%7$*X"\ZXL!

! R%>-6)'&!/'60#!>%-#:,%>%6)! N! H->-0,'#36-! b! b-','! %)! -/*! PR:>>-,'J-)'.6! .(! 5:/)'BA.&:>%6)! G.$'&! a'%,-,&3'%#! :#'62! R:<>.4:/-,!

5'E):,%#T!KO,.&%%4'62#!.(!)3%![Z,4!F66:-/!5%%)'62!.(!)3%!F##.&'-)'.6!(.,!8.>$:B!)-)'.6-/!C'62:'#)'&#7!VWX[7!$*! [[Z\[dZL!

N! R*! a-,'#$%7! U*! 5.6)>-'67! 5*! 5%4g0.:6%"! PR:>>-,'J'62! 8.6&%$):-/! A%#&,'$)'.6#! :#'62! ^6.@/%42%!

H%$,%#%6)-)'.6#T!K8.6(%,%6&%_!;hhh!R=>$.#':>!R%,'%#!.6!8.>$:)-)'.6-/!;6)%//'2%6&%!VWXd7!A%&%><%,!VWXdL!

N!c*!I'.,'6'!%)!-/*!PR%>-6)'&!#'>'/-,')'%#!-)!)3%!&.,%!.(!2%6%,'&!'64%E'62!-64!&/:#)%,'62!-$$,.-&3%T! ! +)3%,! N!F6:$->!U.#3'7!H-23:!^,'#36-$:,->*!PH.<:#)!I:JJ=!8/:#)%,'62!5%)3.4#!).!R:$$.,)!1%<!5'6'62T! N! A-?'4! +$')J7! H'&3-,4! 5-&/'6*! PO.$:/-,! h6#%></%! 5%)3.4#_! F6! h>$','&-/! R):4=T! KU.:,6-/! .(! F,)'('&'-/!

;6)%//'2%6&%!H%#%-,&3!!!7!A%&%><%,!XeeeL!

(b%,,-37!C*7!5-:,'#7!S*7!5.6)>-'67!U*!KVWW]L!A'-26.#'#!-64!'>$,.?%>%6)!'64%E%#!(.,!-!>:/)'B&,')%,'-!!

'64:#),'-/!$%,(.,>-6&%!#=6)3%#'J%4!<=!-!83.9:%)!'6)%2,-/!-22,%2-)'.67!#$%!&'(%)'*(+,'*-!.,/)'*-!,0! 1*'*2%3%'(!45+%'5%!617897!&"!KZL7!$$*!Z"WBZ[X*!

b')')&'7!Q*!R*!KXee[L*!5.4%//'62!.(!$%,(.,>-6&%!>%-#:,%>%6)!#=#)%>#!'6!>-6:(-&):,'62!%6)%,$,'#%#*!

;6)%,6-)'.6-/!U.:,6-/!.(!O,.4:&)'.6!h&.6.>'&#*!($7!KVL!$$*!XZ`BX"`*!

Page 34

SLIDE 39

Second year PhD students

Page 35

SLIDE 40

Page 36

SLIDE 41

Evaluation of data veracity using value confidence and source trustworthiness

Valentina Beretta1, S´ ebastien Harispe1, Sylvie Ranwez1, and Isabelle Mougenot2

1 Centre de Recherche LGI2P, ´

Ecole des mines d’Al` es, Parc Scientifique Georges Besse, F-30 035 Nˆ ımes cedex 1, France {valentina.beretta, sebastien.harispe, sylvie.ranwez}@mines-ales.fr,

2 UMR Espace-Dev, Universit´

e de Montpellier, Montpellier, France isabelle.mougenot@umontpellier.fr

Abstract. Identifying the most reliable and trustworthy data is the

main goal of truth discovery approaches. Following semantic Web formalism, only pairs < data item, value > are considered here; a data item being composed of a subject and a predicate. Existing methods can be improved taking into account additional information. During my thesis both dependencies among values and dependencies among data items have been considered in order to evaluate the confidence associate to certain data. This synthesis intends to provide an overview of how this kind of information has been integrated into an existing truth discovery model. The empirical experiments, conducted on DBpedia, show the potential of the proposed approaches. Keywords: Source Trustworthiness, Value Confidence, Truth Discov- ery, Ontology, Rules

1 Introduction

Nowadays, Web data is used to reach a growing number of different targets. For instance, applications and services may use this data to enrich knowledge bases, assist decision making, etc. However, Web data is not controlled and filtered before its publication; the process able to distinguish true information from false

ne is therefore needed. It always requires the same series of steps independently

from the predefined target. The first step consists in monitoring the Web. The second one regards how interesting information can be identified from this stream of data. Note that, in this report, we will consider, as interesting information, claims in the form of triples such as < subject, predicate, value >, e.g. < PabloPicasso, bornIn, Malaga >. Then, the third step consists in checking the veracity of these claims. Finally, during the last step, the true claims, i.e. facts, can be used to reach the established goal. My PhD thesis focus on the third step, i.e. the analysis of the claims in order to identify the true ones. More precisely, the aim of the project is to enhance the performance of existing Truth Discovery (TD) models taking into account Page 37

SLIDE 42

additional information such as the dependencies that may exist among values or among data items (a data item d is a pair < subject, predicate >). For instance, if a source provides the claim Pablo Picasso was born in Spain, considering the dependencies of values, then it should support also all claims containing more general values than Spain, e.g. Europe. These claims should not be considered false a priori, even if only a single true value is admitted for the given pred-

icate. Indeed, it is well known that Spain is an European country. Moreover,

considering that dependencies may exist among data items, observing that a lot

f people that speak Spanish were born in Spain and thatPablo Picasso speaks

Spanish should support the claim Pablo Picasso was born in Spain. In this report I briefly summarize how the dependencies among values and data items have been incorporated into TD models in order to enhance their performances. Readers will refer to [1, 2] for technical details.

2 Related work

The existing truth discovery models can be divided into three main classes: baseline, basic and extended approaches [1]. In the baseline method (voting), the true values are the ones that are provided most frequently by sources. The basic approaches are able to identify the truth estimating value confidence and source trustworthiness. Indeed, they assume that reliable sources give true information and true information is provided by reliables sources [4]. These iterative approaches do not use external knowledge to improve their results. This is done by extended models. Among them, there are methods that take into account dependencies that may exist among sources (e.g. [6]), data items (e.g. [7]) and values(e.g. [8]). The existing methods take into account value dependencies evaluating string distance or numerical difference. Moreover they use spatial and temporal analysis to identify dependencies among data items. The models we propose consider that values are not independent, as well as data items given, respectively, a partial order and frequent patterns that may exist among them. Our aim is to incorporate in the truth discovery process this new knowledge to improve both confidence and trustworthiness estimations.

3 Proposed approach

During the first year of my thesis, we proposed an approach that considers the dependencies that may exist among values in form of partial order. The rationale we use is that more general values than a true one are true as well. Since the partial order indicates which values are more general than others, it can be used to propagate the confidence associated to a provided value (a source claims a value when it considers the value as true) to all its more general ones. In

ur work, we adapted the existing Sums method [5]; an iterative procedure in

which the estimations of source trustworthiness and fact confidence are alternate until convergence. At each step, the source trustworthiness of a source s ∈ S is evaluated summing up all confidences of values that are provided by s. Similarly, the confidence of a value v ∈ V is computed summing up all trustworthiness of sources that claim v. The proposed approach that considers the partial ordering Page 38

SLIDE 43

f values uses the following formula to estimate confidences (the formula to

compute the trustworthiness scores remains the same): ci(v) = β X

s2Sv+

ti(s) where ti and ci denote the estimated source trustworthiness and value confidence respectively at iteration i, Sv+ is the new set of sources that is considered. It is composed of either sources providing the value under examination or sources claiming values that implie v considering the partial ordering of values, i.e., Sv+ = {s ⇧ S|s ⇧ Sv0 : v0 ⇧ V ⌥ v0 ⇤ v}. Finally, α and β are two normalization factors respectively equal to 1/maxs2S(ti(s)) and 1/maxf2F (ci(f)). They avoid the degeneration of the estimated quantities. Note that considering the partial

rder, confidence estimations are monotonically increasing from the leaves to

the root of the poset. Formally, ⌃(x, y) ⇧ V 2 : x ⇤ y, c(x) ⇥ c(y). This is in accordance with our rationale. We cannot be less confident in the correctness of the value Europe than Spain. Indeed the confidence of Europe has to be at least the same than the confidence of Spain. During the second years of my thesis we proposed another approach that considers the dependencies that may exist among data items detected thanks to rules. Indeed rules are able to capture the frequent patterns contained in a KB and, moreover, they are easily interpretable. We used AMIE+ [3] to extract rules from DBpedia. In this case, the rationale we use is that observing recurrent pattern can enhance our confidence in a specific value. The proposed approach uses the following formula to estimate confidences (the formula to compute the trustworthiness scores remains the same): ci(v) = β0[(1 γ)cbasic(v) + γ boost(v)] where β is a normalization factor, cbasic : V ⌅ [0, 1] is the confidence score computed with the original formula used in Sums (the sum of all sources that provide v), γ ⇧ [0, 1] is a weight that represents the importance assigned to the

bservation provided by sources and to the KB in order to estimate the value
confidence. Moreover, boost : V ⌅ [0, 1] is a function that computes, for each

value, a score indicating the degree of the value correctness w.r.t. the instances contained in the external KB. In particular, it captures the amount of rules that confirm the value under examination among all the rules in which the body is verified by at least one instantiation considering the current subject contained in the data item. We also proposed a model that takes into account both depen-

dencies. In this case we cannot use the previous boosting factor since it does not

permit to obtain confidences monotonically increasing w.r.t. the partial order. Indeed a rule can be extracted for a value, but not for its more general ones. The new boosting factor should be able to propagate the information given by the set of rules associated to a value to its more general ones, see [2] for further

details. We generated synthetic datasets to evaluate our approach (see [1] for

details regarding the protocol we proposed). Empirical experiments performed

n these datasets extracted from DBpedia (using the predicate birthPlace) show

Page 39

SLIDE 44

that both the partial ordering of values and the frequent patterns extracted in an external KB are valuable resources to the truth discovery process. The results show that the best model is the one considering both dependencies. It obtains an average precision of 73% (percentage of returned expected true values), while Sums only achieves 28%, Sums incorporating data item dependencies obtains 32% and Sums incorporating value dependencies reaches 69%.

4 Conclusion and perspectives

We have proposed two different adaptations of truth discovery models, in order to take into account that a partial order may exist among values and frequent patterns may exist among different data items. We tested this framework adapting the Sums model showing that incorporating a priori information permits to improve the results. Next steps will consist in extending experiments on rule settings in order to generalize our findings, adapting other models to incorporate this additional information and proposing an application scenario that can benefit from the results we have obtained.

References

1. V. Beretta, S. Harispe, S. Ranwez, and I. Mougenot. How can ontologies give you

clue for truth-discovery? an exploratory study. WIMS ’16.

2. V. Beretta, S. Harispe, S. Ranwez, and I. Mougenot.

Identification de r` egles pour renforcer la d´ etection de v´ erit´ e. In IC 2017 : 28es Journ´ ees francophones d’Ing´ enierie des Connaissances, Caen, France, 3-7 juillet., 2017.

3. L. Gal´

arraga, C. Teflioudi, K. Hose, and F. M. Suchanek. Fast rule mining in

ntological knowledge bases with amie+. The VLDB Journal, 24(6):707–730, 2015.
4. Y. Li, J. Gao, C. Meng, Q. Li, L. Su, B. Zhao, W. Fan, and J. Han. A survey on

truth discovery. CoRR, abs/1505.02463, 2015.

5. J. Pasternack and D. Roth.

Knowing what to believe (when you already know something). In Proceedings of the 23rd International Conference on Computational Linguistics, COLING ’10, pages 877–885, 2010.

6. G.-J. Qi, C. C. Aggarwal, J. Han, and T. Huang.

Mining collective intelligence in diverse groups. In Proceedings of the 22Nd International Conference on World Wide Web, WWW ’13, pages 1041–1052, Switzerland, 2013. International World Wide Web Conferences Steering Committee.

7. D. Wang, T. Abdelzaher, and L. Kaplan. Social sensing: building reliable systems
n unreliable data. Morgan Kaufmann, 2015.
8. X. Yin, J. Han, and S. Y. Philip. Truth discovery with multiple conflicting informa-

tion providers on the web. IEEE Transactions on Knowledge and Data Engineering, 20(6):796–808, 2008.

Page 40

SLIDE 45

Extracting information from court decisions

Gildas Tagny Ngompe1,2, Sébastien Harispe1, Jacky Montmain1, Stéphane Mussard2, Guillaume Zambrano2

1 LGI2P, École des mines d’Alès, 69 Rue Georges Besse, Nîmes, France 2 CHROME, Université de Nîmes, Rue du Dr Georges Salan, Nîmes, France

Abstract. We present our recent progress on the analysis of the French

judicial corpus; we discuss in particular (i) results of our detailed study of HMM and CRF probabilistic models for detecting sections and entities in court decisions, and (ii) primilary work on the extraction of claim information through key-phase extraction and decision classification. We also discuss faced open challenges like dealing with multiple claims of similar or different types in the same decision, as well as lines of thought for solving these challenges based on discourse analysis. Keywords: Natural Language Processing, probabilistic models, text classification, court decisions sectioning, entities and claims extraction

1 Introduction

A court decision is a document containing the description of a case, i.e. the decision of the judges as well as their motivations. In our aim to provide meaningful insights of court decisions corpora, we are designing automatic information extraction approaches enabling to extract relevant information for characterizing decisions and further analyse them. In particular, we have been working on (1) document sectioning considering three distinct parts (header, body, conclusion), (2) legal named entities detection (city, type of court, judges, parties, date, norm, lawyers, ...), and (3) claims extraction (i.e. for each claim made by parties, we aim at extracting its category, types of involved parties, requested quantum, result meaning, and granted quantum). More precisely, we have studied how well Hidden Markov Model (HMM) [1] and Conditional Random Fields (CRF) [2] can solve the two first tasks, and how results can be improved by taking into account additional labeled training data, as well as some particular design aspects like feature subset and segment representation selection. Ongoing work also focuses

n experimenting a simple approach combining some weighting and classification

methods to identify categories of claims and the meaning of the corresponding result; we also study how characteristic terms of a category can be used to locate the amount of money (quantum) requested and finally granted. This paper briefly provides details on our work and suggest some ways of improvement. Page 41

SLIDE 46

2 Sections and entities detection using HMM and CRF

HMM and CRF are well-known probabilistic models particularly adapted for extracting information from texts. In our work, we have studied their performance in detecting sections and entities in court decisions. To understand how HMM and CRF work, let’s consider a text T as a sequence of observations or "tokens" t1:n. Each ti is a segment of text - in our case a word for entities and a line for sections. Considering a collection of labels L, labeling T consists of assigning the appropriate labels to each ti. Our detection tasks are both segmentation tasks of text contents T, i.e. the aim is to split T into partitions such that the elements of a partition necessarily form a subsequence of T. While HMM assigns a joined probability P(T, L) = Q

P(li|li−1)P(T|li) to couples of sequences of observations T = t1:n and labels L = l1:n, the CRF rather assigns a conditional probability P(L|T) =

1 Z exp n

i=1 F

j=1

λjfj(li−1, li, t1:n, i) ! to them; where Z is the normalization factor and f(·) are the features functions that handle observation representations in CRF. However, after been trained on labeled text examples, both models are applied on an unlabeled text by using the Viterbi algorithm - this algorithm simply infers the label sequence having the highest probability. To help those models to get better results, some candidate descriptors of tokens have been designed. Our results related to the definition

f some features, showing how they improve HMM and CRF performances are

summarized in [3]. A lot of features have been defined; that are either based on the writing style

f court decisions (e.g. line length, line words, neighbors words, word proper-

ties), or extended properties of tokens (e.g. part-of-speech tags, word topic). To

ptimize the feature sets, two feature selection algorithms have been tested for

CRF: the bidirectional search (BDS) and Sequential Floating Forward Selection (SFFS).Those algorithms add or remove features over iterations as the results improve or not. Our actual experiments show that those algorithms are able to reduce significantly the quantity of features but only slightly improve the results. Their disadvantage is the duration of the selection phase - up to ten hours. We also studied four different segment representations (IO, BIO, IEO, and BIEO) that are different ways of tagging tokens [4]. BIO, IEO, and BIEO help to detect successive entities, as the beginning or the end of each entity is differently

labeled. We also observed that those three representations actually improve re-

sults at the cost of a training time increase. Beside the segment representation and feature selections, labeling more data also obviously improve result within a

limit. This suggests to correctly select the examples in order to cover the different

variants of text structure.

3 A simple approach for claims extraction

During a case, each party expresses claims willing to be granted by the judges. Court decisions summarize, beside facts, claims and arguments of the parties, as Page 42

SLIDE 47

well as answers and reasons of judges. There are a lot of diverse styles to express claims and results: explicitly (in a dedicated paragraph precising the party to condemn, the corresponding fault, and eventually how much we claim), implicitly (need interpretation of texts), through reference to previous decisions (in appeal decisions, judges partially or totally confirm or reverse previous judgments). Hence the expression of claims is not standard and moreover, categories of claims are not all yet known. This makes claim extraction a difficult task. That is why

ur first approach is fully supervised and iterative, i.e. by trying to solve the

problem considering the specificity of each category. Firstly, a set of characteristic terms (n-grams) of a category is extracted by computing a score evaluating the likelihood for terms to occur in texts of that

category. Several supervised global weighting methods have been tested: relative

frequency [5], correlation coefficient [6], ... All these methods measure how much a term is relevant for a category by using labeled examples, i.e. positive texts (within the category) and negative ones (outside the category). Since, it may be hard to build a large set of labeled negative samples, we used a very large set

f unlabeled data by assuming that the considered category only covers a small

proportion of this large set. In practice, this postulate is true for the majority of the categories but fails for some categories covering a large subset of decisions (e.g. article 700 du code de procédure civile). Secondly, to determine whether the decision refers to a category, we represent texts as vectors and then train a classifier on labeled examples. The dimensions

f the vector are the terms learned from training sets that are the most relevant

to the category; with the weight of a term w∗ for a text t defined as: weight(w∗, t) = weightlocal(w∗, t) ∗ weightglobal(w∗) ∗ factornormalization The local weight is generally a term frequency TF-based method like term presence TP, or the logarithm of the term frequency. In our experiment, we have tried different configurations with different sizes of n-grams (1 ≤ n ≤ 5) fixed or not, global and local weighting methods, vector sizes, classifiers (SVM, Naive Bayesian, ...). Classifications on different categories have been tested (e.g. dommages-intérêts pour procédure abusive, trouble de voisinage, article 700, prestation compensatoire). Tests show that this classification task is actually straightforward since some configurations reach a 100 % accuracy even by using the simple Inverse Document Frequency (IDF) schema on all words. Indeed, very few terms can lead to very accurate classification since the vocabulary associated to the categories is very specific, short and common in justice. However, our expectation was that the optimal subset of terms (for classification) would help locate the amount of money claimed and eventually granted

by evaluating distances between those terms and the mentioned quanta. Note

however that the quantum can be mentioned before or after the terms - and sometimes the closest quantum to the terms is not the targeted one; in addition a quantum can also refer to a past decision or a claim. To tackle the problem

f quantum/claim linking we assumed that it might help to zone the different

contexts to further match claims and results using parties’ names mentioned in Page 43

SLIDE 48

that zone. In the setting considered so far, the problem has been simplified by assuming that a decision contains at most one claim of a given category. By training a classifier on all the decisions, we expect to distinguish decisions with granted claims and decisions with rejected claims for a particular category of

claim. Results on the datasets analyzed show the modest but interesting per-

formance of the approach: maximal F1-measures at 75.52 for article 700, and at 86.66 for trouble de voisinage). To improve those results, instead of reduc- ing features by selection, we are currently evaluating techniques of reduction by projection - in the style of the principal component analysis.

4 Conclusion

We have summarized our result on the study of sectioning and detection of legal named entities using probabilistic graphical models (HMM & CRF) - showing the interesting performance of those models in our application context. We are currently working on one of the more difficult tasks of the project: claims extrac-

tion. In particular, we are experimenting a simple method with which promising

results have been obtained so far. As a schedule for the rest of the project, we also propose to study four additional tasks: (1) claims text segments detection using semantic and discourse analysis to deal with dependencies between statements; (2) comparing unsupervised and supervised claims categorization; (3) formalize as much as possible the information extracted so far in a knowledge base by applying disambiguation and resolution methods on extracted information; (4) analyze which factors and situations (contexts) are correlated with judges decision making - paving the way to predictive analysis.

References

1. Lawrence R. Rabiner. A tutorial on hidden markov models and selected applications

in speech recognition. Proceedings of the IEEE, 77(2):257–286, 1989.

2. John Lafferty, Andrew McCallum, and Fernando CN Pereira. Conditional random

fields: Probabilistic models for segmenting and labeling sequence data. International Conference on Machine Learning, 2001.

3. Gildas Tagny Ngompé, Sébastien Harispe, Guillaume Zambrano, Jacky Montmain,

and Stéphane Mussard. Reconnaissance de sections et d’entités dans les décisions de justice: application des modèles probabilistes HMM et CRF. In In Extraction et Gestion des Connaissances - EGC 2017, Revue des Nouvelles Technologies de l’Information, Grenoble, France, January 2017.

4. Michal Konkol and Miloslav Konopík.

Segment representations in named entity

recognition. In International Conference on Text, Speech, and Dialogue, pages 61–
70. Springer, 2015.
5. Man Lan, Chew Lim Tan, Jian Su, and Yue Lu. Supervised and traditional term

weighting methods for automatic text categorization. IEEE transactions on pattern analysis and machine intelligence, 31(4):721–735, 2009.

6. Hwee Tou Ng, Wei Boon Goh, and Kok Leong Low. Feature selection, perceptron

learning, and a usability case study for text categorization. In ACM SIGIR Forum, volume 31, pages 67–73. ACM, 1997.

Page 44

SLIDE 49

Third year PhD students

Page 45

SLIDE 50

Page 46

SLIDE 51

On the definition of a Knowledge Inference Model from relation extraction

Pierre-Antoine Jean1, Sébastien Harispe1, Sylvie Ranwez1, Patrice Bellot2, and Jacky Montmain1

LGI2P, Laboratoire de Génie Informatique et d’Ingénierie de Production, ENS Mines Alès, Parc Scientifique G.Besse, 30035 Nîmes cedex 1, France {firstname.lastname}@mines-ales.fr1 & patrice.bellot@lsis.org2 Abstract This article introduces an automated knowledge inference approach taking advantage of relationships extracted from texts. It is based

n a novel framework making possible to exploit (i) a generated partial
rdering of studied objects (e.g. noun phrases), and (ii) prior knowledge

defined into ontologies. This framework is particularly suited for defining information propagation rules based on evidence theory in order to infer new knowledge. The proposed approach is illustrated and evaluated through the definition of a system performing question answering by analyzing texts available on the Web. This case study shows the benefits of structuring processed information (e.g. using prior knowledge) for inferring new knowledge. Keywords: Knowledge inference, information extraction, evidence the-

ry, ontology, natural language processing

1 Context and positioning

Nowadays, we can easily access large text corpora that offer a real opportunity in artificial intelligence. Moreover, natural language processing methods are more and more efficient and allow us to extract explicit information from texts. The aim of our pipeline is to mix these approaches with reasoning techniques in order to infer new potentially implicit relations. A large diversity of methods could benefit of this automatic process e.g. knowledge base population (KBP), question answering or knowledge discovery. The pipeline described in the following is based on a specific KBP method. Traditionally, these methods are classified according to different criteria: how to build the KB (manually, automatically), data nature (structured, semi-structured, non-structured) and presence or not

f an initial ontological schema [1]. The method we used, named Reverb [2] is

based on a automatic process that takes as input non-structured data and that do not use initial ontology. We coupled it with a reasoning module in order to realise induction process on extracted relations. In other words, new relations are deduced with the generalisation of our observations. The general architecture of

ur approach reminds the Watson method [3].

Page 47

SLIDE 52

2 Overall of the approach

Our pipeline allows to extract textual information in open domain in order to evaluate and infer knowledge while considering syntactic and taxonomic implic- ations from extracted relations. Its aim can be resumed into four main steps: i) extract relations from non-structured texts, ii) generate new relations from extracted relations, iii) compute criteria to evaluate relations and iv) assess relations according to their criteria. We can illustrate it through an example that takes into account two sentences, and extracts two relations <subject, predicate,

bject> using Reverb.

Marfan syndrome is caused by FBN-1 gene mutation. Cystic fibrosis is caused by CFTR gene mutation. ↓ <Marfan syndrome, is_caused_by, FBN-1 gene mutation> <Cystic fibrosis, is_caused_by, CFTR gene mutation> 2.1 Structuring and enrichment of extracted relations The observation of nominal phrases “FBN-1 gene mutation” and “CFTR gene mutation” involves more general concepts such as: “gene mutation” and “mutation”. Hence, from this syntactic decomposition, we can build a partial order between these concepts. This partial order allows to transpose implicit extra knowledge from the extraction step like <Marfan syndrome, is_caused_by, gene mutation> and <Marfan syndrome, is_caused_by, mutation>. The main interest of this conceptual organisation is the possibility to use prior knowledge expressed in taxonomies in order to enrich our extractions. For instance, from a specific medical taxonomy, we can add “genetic disease” as a more general concept compared to ”Marfan syndrom” and ”Cystic fibrosis”. However, a disambiguation step must be used to map terms with concepts in the taxonomy. In

ur pipeline, we use the similarity distance of Wu & Palmer.

2.2 Generation and inference of knowledge As we have seen, we can leverage the partial order on concepts to generate new information from texts. To make this, we use a cartesian product between subject and object ascendants for a given predicate in a relation (cf. Figure 1). After this generation, the aim is to evaluate generated and extracted rela-

tions. Therefore, we modelled two main criteria to discriminate relations: belief

and specificity. In our work, belief is tackled by the evidence theory. So to obtain these criteria, we gathered and structured relations together in a specific graph. We defined manually a set of rules to organize relations according to their hierarchical links between theirs subjects and objects. Figure 2 presents an example

f this graph.

Thanks to the relation graph, we can propagate bottom-up the belief mass to obtain the belief value, bel(s), of each relation s and use the maximal depth Page 48

SLIDE 53

Figure 1. Generation of new relations. Dashes represent possible generated relations. > genetic diseases Marfan synd. Cystic fibrosis mutation gene mutation FBN-1 gene mut. CFTR gene mut. Figure 2. Implicit hierarchy between extracted and generated relations. is_caused_by(genetic diseases, gene mutation) is_caused_by(Marfan syndrome, FBN-1 gene mutation) is_caused_by(Cystic fibrosis, CFTR gene mutation)

f relations to represent their specificity value, depth(s). Then, to exploit these

criteria, we implemented some inference models M. According to our validation, the best model was: SM = {s ∈ S | bel(s) ≥ beldepth(s)} where SM is the relevant relations set, S the generated and extracted relations set and belx is described in equation 1. belx = P

{s∈S|depth(s)=x}

bel(s) | {s ∈ S | depth(s) = x} | (1) 2.3 Validation To evaluate our pipeline, we used and modified a specific protocol to automatically generate Multiple Choices Question (MCQ). This protocol exploits a taxonomic structure to generate questions, false answers and true answers. In our case, we used it to generate our own quiz based on the MeSH taxonomy and Diseases/Symptoms relations from the OMIM database. In this experiment, our baseline is the relation extraction method (Reverb) without the reasoning mod-

ule. Results resumed in Table 1.

Page 49

SLIDE 54

Table 1. Results obtained on questionnaires using sentences gathered on Web. “Propagation” means the deployment of relation generation process described in sub- section 2.2 and “Inference” means the deployment of the inference model. Propagation Inference Precision F-measure No No 0.96 0.21 Yes No 0.95 0.40 Yes Yes 0.97 0.35

Many observations can be realised from these results. The first one con- cerns the interest to use a generation relation protocol. Indeed, thanks to it, our pipeline allows to significantly improve results compared to the baseline. The second one is the low F-measure due to a low recall. It can be explained by the nature of the initial corpora used to find answers. In fact, to build the corpora we used Google with requests containing the diseases name from subjects OMIM

relations. However, we have no certainty that these texts contain the expected
relations. Nevertheless, the best result of our pipeline, for us, is the relation dis-
covery. Indeed, we observe some false-positives in results but a manual analysis

showed that was not false answers. So, we discovered some unknown relations regarding to OMIM base.

3 Work in progress

We focus on two main works. The first one is to modify the pipeline to take into account linguistic uncertainty [4]. Indeed, we saw that our approach can be used in knowledge discovery and in this domain, linguistic uncertainty can probably play an important role in the aggregation of low signals. To take into account this uncertainty, we can modify the mass function e.g. decrease the importance

f an uncertain relation. Finally, we work on specific user profiles. In fact, users

can have different aims and different knowledge levels about a domain, so their expectation can be really different. To tackle this, we try to define user profiles in order to distinguish: neophyte vs. expert and the context of their search: knowledge discovery vs. KBP.

References

1. Murphy, K., Dong, X., Gabrilovich, E., Heitz, G., Horn, W., Lao, N., Zhang, W.:

Knowledge vault: A web-scale approach to probabilistic knowledge fusion. 20th ACM SIGKDD international conference on Knowledge discovery and data mining (2014) 601–610

2. Etzioni, O., Fader, A., Christensen, J., Soderland, S., Mausam, M.: Open informa-

tion extraction: The second generation. IJCAI 11 (2011) 3–10

3. Ferrucci, D.A.:

Introduction to “this is watson”. IBM Journal of Research and Development 56(3.4) (2012) 1–1

4. Jean, P.A., Harispe, S., Ranwez, S., Bellot, P., Montmain, J.: Uncertainty detection

in natural language: a probabilist model. WIMS (2016) 1–10

Page 50

SLIDE 55

Towards a non-Oriented Approach for the Evaluation of Odor Quality

Massissilia Medjkoune1, 2, Sébastien Harispe1, Jacky Montmain1, Stéphane Cariou2 and Jean-Louis Fanlo2 1- Scientifique Georges Besse, F30035 Nîmes cedex 5, France 2- de Clavières, 30319 Alès cedex, France firstname.name@mines-ales.fr

Abstract. In this article, we present a non-expert approach to evaluate the odor quality. Charac-

terizing odour quality consists in identifying a set of descriptors that best synthesizes the olfac- tory perception. Two issues are then related to odour characterization. The first one is how translating free natural language (NL) descriptions into structured descriptors; the second one is how summarizing a set of individual characterizations into a consistent and synthetic unique characterization for professional purposes. This paper will propose an approach to automatize both translation of bags of terms into sets of descriptors and summarization of sets of structured descriptors. Keywords: Odour Quality, Natural Language, Information Fusion, Taxonomy.

1. Introductory example
expert evaluators to describe the odour of a given tested

wine by using their own terms. If for example, we consider that we get the following conceptual descriptions and after translating the free NL descriptions into structured descriptors: . The idea is to merge the information expressed by these two conceptual annotations and to formally characterize the odour of this wine by a unique synthetic set of concepts. Intuitive- ly, abstracting these two descriptions by summarizing the information provided by the descriptors relies on our knowledge of the organization of the descriptors. Because there is no formal definition to what a relevant summary is, several summaries can be proposed, e.g. Indeed, the idea beyond this choice is that and are sorts of (i.e., ), and are sorts of and is a generalization of and . This reasoning is based on taxonomic knowledge partially ordering concepts ( ) (descriptors ordered by the sort-of relationships). The next section presents the proposed approach: we first compute the individual conceptual annotations from the free descriptions and then we propose a formal method to produce the conceptual summary that best approximates all the annotations.

Page 51

SLIDE 56

2. Automated approach for assessment of odor quality

The first step of our approach is to define a mapping between terms and concepts of the related

taxonomy. In this way, we compute the degree to which a term evokes a concept of the taxon-
my. Then, each term is associated to the concept that evokes it the best. Finally, the free de-

scription (here, bag of terms) is transformed into a conceptual annotation (set of ordered concepts). The second step of our approach is to synthesize conceptual annotations into a limited set of concepts. The various definitions which will be used are listed below: Notations:

The notion of Information Content refers to the degree of specificity of con-

cepts: [1].

We denote by and respectively the inclusive ancestors and inclusive

descendants of the set of concepts .

,
We denote by the mass function, , that corresponds to the

number of observations of concept and

. In our application,

represents the number of times concept has been proposed by evaluators.

The belief and plausibility functions and proposed in the

DempsterShafer theory are next defined such as [2]:

2.1 Individual conceptual annotation

In this section, we try to associate to each term of a NL description, the concept that is the most likely evoked by the term. We intuitively propose to measure the intensity of evocation a term has with a concept. This evocation can be expressed as the semantic proximity between the term and the concept denoted by with the set of terms that constitutes the

vocabulary. We consider that measuring the similarity between a term and a concept con-

sists in calculating the similarity between the term and the labels associated to the concept . This semantic proximity between and can be estimated, for example, as the maximal similarity between the term and the labels associated to the concept : with the set of terms associated to the concept and . Numerous measures for comparing terms based on distributional representations of terms have already been proposed in the literature (co-occurrence, pointwise mutual information, etc.) [3]. The conceptual annotation associated to the evaluator can be computed as follows:

2.2 Collective conceptual annotation (summary) The purpose of this section is to provide a synthesis that summarizes the information provided by evaluators . We consider that each evaluator provides a set of terms

to describe the

dor. Using the model proposed in the previous section (Eq. (1)), to each evaluator is

now associated a set of concepts (individual conceptual annotation) denoted . We denote by =( ) the sequence of annotations to be summarized and the set of

Page 52

SLIDE 57

concepts of evaluators such that:

. We formally define by the function summariz-

ing a sequence of individual conceptual annotations by a single summary from . We suppose that each synthesis respects the following properties: 1) Summarizing: 2) Fidelity: such as x 3) Non-total-redundancy: The search of best summary for any can be expressed as an optimization

problem. The objective function is defined as:
The function

is used to estimate the amount of information conveyed by which is summarized by whereas estimates the loss of information using . We first evaluate the common information conveyed by the ancestors of and . Intuitively this quantity could be defined as follows:

The function is introduced to weigh the importance of concepts.

For criticizing the relevance of a summary it is also important to discuss penalties regarding loss, addition and distortion of information. In the following, we detail the different factors of penalty . We define the penalty of abstraction by:

models the amount of exact information conveyed by

which is not conveyed by .

models the amount of plausible information conveyed by which is not conveyed by
(penalty regarding addition of plausible information)
models the amount of plausible information conveyed by

which is not conveyed by (penalty regarding loss of plausible information).

penalizes the distortion which is made considering a specific choice among partially

covering summaries. This penalty should be a function of i.e., all the elements of that have not been summarized. We propose the following model to estimate the distortion.

Page 53

SLIDE 58

The parameter is used to weigh the importance of a specific uncovering. Finally, the penalty

f abstraction

is defined as follow: with ,, , weight parameters used to balance the importance of each abstraction penalty factor. Finally, is a penalty used to evaluate the conciseness of the summary by penalizing redundant information implicitly conveyed by a summary.

The penalization is designed such as each abstracted notion that is repeated more than once will

be penalized by the number of times the redundant information appears. The parameter can therefore be used to control the number of concepts in . Harispe et al. have proposed algorithms enabling to use this model for searching for relevant summaries and discuss interesting properties of the search space [4].

3. Conclusion

In this article, we have proposed a model to automatically summarize the free NL evocations provided by neophyte evaluators by considering knowledge representations and embedding words models. In our current work, we still have to carry on further experiments to validate each of the two steps of our approach. The mapping between terms and concepts of the taxon-

my that has been presented here is a naïve one to one mapping from terms to concepts. An

extension of this mapping without cardinality constraint has been published at 11th International Conference on Scalable Uncertainty Management, Granada, Spain, 2017. It allows merging a set of terms into a unique non ambiguous entity.

4. References

[1]

ECAI, vol. 16, 2004.

[2]

S. Harispe, A. Imoussaten, T. François, and J. Montmain-to-

mind model for computing thIEEE

Int. Conf. Fuzzy Syst. (FUZZ IEEE), Istanbul, Turkey, 2015.

[3]

Synth. Lect. Hum. Lang. Technol., 8 (1), 2015.

[4] IEEE Symp. Ser. Comput. Intell., 2017.

[5] Harispe, S., Medjkoune, M., Montmain, J. (2017). Eliciting Implicit Evocations using Word

Embeddings and Knowledge Representation, 11th International Conference on Scalable Uncertainty Management, Granada, Spain, 2017

Page 54

SLIDE 59

A Possibilistic Framework for Identifying the Most Promising Performances for Improving Complex systems

Diadie Sow, Abdelhak Imoussaten, Pierre couturier, Jacky Montmain

!!!!!!!!!!!!!!!!!"#$%&'()%*'+!,$#"-!&$'+!,$#"!.#%,"'/$&"'-)*

Abstract. The question addressed in this research is: how to identify the characteristics of a

e- quences are only imprecisely known and assessment of products is multidimensional? Although some theoretical worth indexes have been proposed in the multiple criteria literature to estimate the expectable gains when improving changes are planned, they generally rely on non-realistic assumptions on the achievability of the expected improvements. Based on multi-criteria decision analysis techniques and uncertainty theory, this paper proposes an extension of the worth index concept when the likelihood of expected improvements is not precisely known as it is the case at the preliminary stages of design activities. Keywords: Performance evaluation, Imprecise assessment, Multi-criteria analysis, Possibility Theory.

1 Introduction

In a highly competitive and unstable environment, industrials have to constantly improve their products to remain competitive and satisfy their customers while minimizing incurred costs and risk taking. Many constraints must be taken into account when designing or when improving a product [1]. Decisional strategies have thus to be set

ut to define, compare and select potential improvement actions with respect to many
criteria. However, in practice, criteria interact between them and prefer-

ential interactions may make counter-intuitive the overall result of elementary im- multi-criteria decision analysis and on the family of fuzzy integrals like the Choquet integral (quantitative framework) and the Sugeno integral (qualitative framework) which allows modelling both the preferential interactions between performances and their relative importance through a fuzzy measure [2]. Recently, some works exploit the preference model of the multiple criteria decision framework, in particular the Choquet integral, to estimate the expected improvements of an alternative. The authors in [3] [4] associate to any coalition of criteria an index named worth index that measures, for any alternative, the mean of all the possible expected improvements w.r.t a subset of crite-

ria. These models assume that all the improvements are equiprobable whatever their
magnitude. This assumption can clearly be challenged in practice. In [5] an extension
f the worth index in a probability framework is proposed. Improvements magnitude
beys probability laws that are learned from data. However, such a framework relies

upon the data availability which is not often the case in the context the preliminary Page 55

SLIDE 60

stages where little information is available making the determination of such probability distributions illusory. This led us to orient this work towards the theory of possibil- ities, which is an adequate framework to represent imprecision and uncertainty. Thus, we propose an extension of the worth index to a possibilistic framework . Section 2 briefly reminds essential notions to define the worth index and gives possibility to understand our proposal. Section 3 introduces and discusses our possibilistic framework of the worth index.

2 Recall about possibility theory

2.1 Reminds and Notation be the set of criteria. preferences are modeled by a Choquet integral! C !!"#$%&'()'*#%#+%,%-.//0%1',(.&'! "! (see [2] for more details)% 2.2 Possibility theory A possibility distribution assigns to each element in a set alternatives a degree of possibility ( )

[0;1]

f being the correct description of a state of the

world.

means that is impossible ;

means that is totally possible ;

[ ; ]

{ : ( ) 1 } b c

is the kernel of ;
] ; [ { : ( )

0} a d

is the support of .

Given a possibility distribution two distribution functions are associated to and are named respectively upper distribution

F and lower distribution

F [6]:

Fig. 1. Notations

*( )

sup{ ( ), } F x r r x

;

*( )

inf{ 1 ( ), } F x r r x

Page 56

SLIDE 61

The mathematical expectation associated to is imprecise and is then defined in [6] by the interval:

* *

( ) [ ( ), ( )] E E E

where

* *( )

( ) E xdF x

* *

( ) ( ) E xdF x

3 Possibilistic extension of the worth index

3.1 Worth index Let

[0,1]n x

be an initial performance vector. The worth index associated to C for a given subset of criteria I

knowing the initial performance

x is given in

[4] by:

( , ) [ ((1 ) 1 , ) ( )]

C I I I

w x I C x x C x d

The main drawback in this approach is that all the performance profiles are supposed to be equally probable ; this is evidently not the case in practice. 3.2 Formulation of possibilistic worth index We consider that experts can provide criteria performance in form of possibility distributions

...

. Indeed from their experience they know the minimal and maxi-

mal performance they have already met providing the kernel limits of each

.The

value range of the support with a degree of possibility lower than 1 corresponds to possible values from the expert points of view although he has never met them before. For the sake of simplicity, only trapezoidal distributions are considered in this paper. Improving criteria in the coalition I means that only performances related to criteria

I are improved whereas performances in remain equal

x . In other words the

initial distributions

are revised such as:

for i

[0;1]

0, if ( ) 1 ( ) ( ), otherwise

i i i i i i x i i

x x x x x

\ i N I

1, si ( ) 0, otherwise

i i i i x

x x x

For each subset of criteria I , let

/ , C x I

be the resulting aggregated possibility

distribution of the revised corresponding possibility distributions using the Choquet Page 57

SLIDE 62

integral C with respect to a fuzzy measure . The possibilistic worth index is imprecise and can be defined using the expectation associated to

/ , C x I

* * / , / ,

( , ) ( ) ( ), ( ) ( )

C C x I C x I

w x I E C x E C x

By construction of the

/ , C x I

, one have

* / ,

( ) ( )

C x I

E C x

Therefore it is just enough to take

* / ,

( , ) 0, ( ) ( )

C C x I

w x I E C x

The computation of

/ , C x I

is based on the Zadeh:

[0;1] z

/ , [ ,1 ], / / ( )

( ) sup min( ( ))

I I I I I I

i C x I y x i x i I y x z C y

z y

where

[ ,1 ]

I I I

y x

stands for

[ ,1]

i i

y x

for all i

. The time needed to com-

pute (4) may be too long, but it can be dramatically reduced using [5] proposals.

4 Conclusion:

Identifying the criteria to be profitably improved is an important issue for innovation but presents important costs and risks. Because of insufficient knowledge about the performances that can be expected from improvement decisions, the only available information are often incomplete and imprecise, early engineering decisions difficult to make. We presented here a possibilistic framewok to identify the most promising criteria in uncertain context.

References

1. E.:Prediction and multi-objective optimization of high-strength

concrete parameters via soft computing approaches. In: Expert Systems with Applications pp -6145-6155, (2009).

2. Marichal, J. L.:An axiomatic approach of the discrete Choquet integral as a tool to aggre-

gate interacting criteria. In:IEEE Transactions on Fuzzy Systems 8.6 pp 800-807,(200).

3. Grabisch, M. ,Labreuche, C. : How to improve acts: an alternative representation of the

importance of criteria in MCDM. In: International Journal of Uncertainty, Fuzziness and Knowledge-Based Systems 9.02 , pp145-157. (2001)

4. Labreuche, C. : Determination of the criteria to be improved first in order to improve as

much as possible the overall evaluation. In: Int. Conf. on Information Processing and Management of Uncertainty in Knowledge-Based Systems (IPMU), Perugia (ITA) (2000).

5. Imoussaten, A., Duthil, B.,Trousset, F., Montmain, J. In: Identifying priority lines of im-
provement. application to tourism data, LFA 2016, La Rochelle, France, (2016)
6. Dubois, D., Prade, H. In: The mean value of a fuzzy number. Fuzzy sets and systems 24.3

(1 987)

Page 58

SLIDE 63

Fourth year PhD students

Page 59

SLIDE 64

Page 60

SLIDE 65

ORIENTED FILTERS FOR FEATURE EXTRACTION IN DIGITAL IMAGES: APPLICATION TO CONTOURS, CORNERS DETECTION, EVALUATION

PH.D. T HESIS SUMMARY REPORT − T HIRD YEAR

Hasan ABDULRAHMAN 1,2, Philippe MONTESINOS 2 and Baptiste MAGNIER 2

1 University of Montpellier, Montpellier 34090, Cedex 5, France. 2 Ecole des Mines d’Al`

es, LGI2P, Parc Scientifique G.Besse, F-30035 Nˆ ımes Cedex 1, France.

ABSTRACT A digital image contains different information from the scene, such as objects, colour, and orientation. The discrimination

f the objects from their background is the first problem that is performed before any further processing. In order to extract

the contour of an object, edges forming that object must be detected, and this fact reveals the constitutional importance of edge detection in computer vision and image processing. Edge detection features support wide range of applications such as recognition, compression, image enhancement, restoration, registration, retrieval, watermarking, steganography, steganalysis. This report, gives a summary for our work in this year. We proposed three methods: In the first proposed method [1], several referenced based boundary detection evaluations are detailed, pointing their advantages and disadvantages through concrete examples of edge images. Then, a new supervised edge map quality measure is proposed. In the second proposed method [2]. We detail several edge dissimilarity measures and present how to evaluate filtering edge detection technique involving these considerate measures. In a second time, we demonstrate how to build a new ground truth database which can be used in supervised contour detection evaluation. Finally, the third proposed method is oriented half kernels for corner detection [3]. In this method, corners are directly extracted involving only a combination of asymmetric oriented kernels. Index Terms— Edge detection, ground truth, supervised evaluation, distance measure, corner detection.

1. INTRODUCTION

Edge detection remains a crucial stage in numerous image processing applications. Thus, an edge detection technique needs to be assessed before use it in a computer vision task. As dissimilarity evaluations depend strongly of a ground truth edge map, an inaccurate datum in terms of localization could advantage inaccurate precise edge detectors or/and inappropriate a dissimilarity evaluation measure. We noticed in the recent years there has been considerable interest in techniques for evaluation of computer vision applications and methods. All systems, especially automated information processing structures, must be evaluated before being developed, principally for industrial applications or medical data. Different algorithms have been developed in the past, but few of them give an objective performance comparison. The evaluation process should produce a result that correlates with the perceived quality of the edge image, which is relied on human judgment. In other words, a reliable edge map should characterize all the relevant structures of an image. On the other hand, a minimum of spurious pixels or holes (oversights) must be created by the edge detector at the same time. Therefore, an efficient evaluation can be used to assess and improve an algorithm, or to optimize edge detector parameters [4]. On the other hand, in the literature several approaches have been developed to detect corners and junctions: (i) involving contour chains, (ii) using templates or, (iii) by image filtering techniques. Traditional contour based corners methods focus on the processing of binary edges, by searching points having curvature in contour chains or in line segment intersections [5] This report, is organized in four sections, Section 2 is describe the proposed methods. Experimental results an comparisons with its tables are given in Section 3. The final Section 4 talk about the conclusion and future work.

2. PROPOSED METHODS

2.1. A new normalized supervised edge detection evaluation In [6] is developed a normalized measure of the edge detection assessment, denoted Γ. This function represents an over- segmentation measure which depends also of FN and FP. As this measure is not sufficiently efficient concerning FNs because Page 61

SLIDE 66

it does not consider dDc. Thus, inspired by Sk, the new measure Ψ holds different coefficients changing the behavior of the measure:

Ψ(Gt, Dc) = FP + FN |Gt|2 · s X

p∈Gt

Dc(p) +

p∈Dc

Gt(p)

(1)

Compared to Γ, Ψ improves the measurement by combining both dGt and dDc. Authors of Γ have studied the influence of the coefficient in different concrete cases [6]. They concluded that such a formulation must take into consideration all observable cases and theoretically observable. In fact, a performing measure has to take into account all the following input parameters |Gt|, FN and FP whereas the image dimensions should not be considered. Thus, the parameter F P +F N

|Gt|2

seems a good compromise and has been introduced to the new formula of assessment Ψ. 2.2. From Contours to Ground Truth: How to Evaluate Edge Detectors by Filtering The purpose is to compute the minimal value of a dissimilarity measure by varying the threshold Th of the thin edges computed by an edge detector. Indeed, compared to a ground truth contour map, the ideal edge map for a measure corresponds to the desired contour at which the evaluation obtains the minimum score for the considered measure among the thresholded gradient images. Theoretically, this score corresponds to the threshold at which the edge detection represents the best edge map, compared to the ground truth contour map [7] [8]. Fig. 1 illustrates our proposed process.

Fig. 1: The most relevant edge map for a dissimilarity measure is indicated by its minimum score.

This work, also presents ground truth edge maps which are labeled in a semi-automatic way in order to evaluate the performance of filtering step/ramp edge detectors. Fig. 2 illustrates the steps to obtain new ground truth images. Using the [ −1 0 1 ] mask enables to capture the majority of edge points and corners without deforming small objects, contrary to edge detectors involving Gaussian filters.

(a) Original image (b) Thin edges with [−1 0 1] mask (c) Adjustment by hand (d) Image in (b) vs image in (c) (e) Legend

Fig. 2: Image of our database are built after an edge detection involving a [−1 0 1] mask and concluded by hand.

2.3. Oriented Half Kernels for Corner Detection The main idea of this new approach is to combine the HGK ( Half Gaussian Kernels) with an asymmetric filter computing the homogeneity along edges. On one hand, the maxima responses of the HGK indicate the directions (2π periodic) of the edges. On the other hand, the oriented variance determines if the directions of the maxima of the HGK corresponds to edges or other types of pixels (texture, homogeneous region etc.). The asymmetric IRON (Isotropic and Recursive Oriented Network) filter estimates the homogeneity in multiple local directions [9]. This filter consists of a network of several parallel lines in which a homogeneity is computed and enables an Page 62

SLIDE 67 0.5 1 90 270 180 HGK 0.5 1 90 270 180 endstop 0.5 1 90 270 180 * endstop 0.5 1 90 270 180 IRON 0.5 1 90 270 180 S HGK 0.5 1 90 270 180 HGK 0.5 1 90 270 180 endstop 0.5 1 90 270 180 HGK * endstop 0.5 1 90 270 180 IRON 0.5 1 90 270 180 S 0.5 1 90 270 180 HGK 0.5 1 90 270 180 endstop 0.5 1 90 270 180 HGK * endstop 0.5 1 90 270 180 IRON 0.5 1 90 270 180 S 0.5 1 90 270 180 HGK 0.5 1 90 270 180 endstop 0.5 1 90 270 180 * endstop 0.5 1 90 270 180 IRON 0.5 1 90 270 180 S HGK

Fig. 3: Modulus of the energy of the different oriented kernels and their combinations (in degrees and normalized signals).

estimation of edge directions modulo 2π. The variance for a pixel located at (x, y) on the network is computed by:

IRON(x, y) = 1 L

j=0

@ 1 P

i=0

I(i, j)2

− 1 P

i=0

I(i, j) !

A . (2)

Here, L represents the number of lines where the variance is computed and P the number of points per line. Fig. ??(e) represents an example of an asymmetric IRON filter. Computationally, the rotation of the image is applied at some discretized

rientations from 0 to 2π before applying the IRON filters. Some examples of IRON filter signals are available in Fig. 3 fifth

column, values of the IRON are close to 0 in the edges directions and it is shown in [9] that the detections of edges directions stay precise in the presence of noise.

3. EXPERIMENTS AND RESULTS

There are any experiments and table results, we select some of these results from each proposed method. All the images and

ther results are available on the website: http://hkaljaf.wixsite.com/hasanabdulrahman/corners-and-junction-detection.

Table 1: Comparison of scores of dissimilarity measures using a ground truth from [10] image and a constructed ground truth

by a semi-automatic way.

Meas. Sobel Canny SF1 [11] AGK [12] H-K [13]

Berkeley Gt

Our Gt

Berkeley Gt

Our Gt

Berkeley Gt

Our Gt

Berkeley Gt

Our Gt

Berkeley Gt

Our Gt Φ∗ 0.738 0.298 0.757 0.430 0.971 0.447 0.813 0.496 0.761 0.504 χ2∗ 0.979 0.635 0.975 0.725 0.983 0.712 0.982 0.759 0.973 0.502 P ∗

0.901 0.530 0.901 0.603 0.909 0.594 0.917 0.637 0.893 0.778 F ∗

0.820 0.360 0.819 0.432 0.834 0.422 0.847 0.468 0.808 0.483 FoM 0.303 0.168 0.310 0.147 0.309 0.164 0.299 0.154 0.277 0.146 F 0.592 0.346 0.579 0.352 0.572 0.310 0.589 0.337 0.589 0.367 d4 0.675 0.333 0.671 0.379 0.687 0.375 0.695 0.412 0.667 0.424 SFoM 0.297 0.145 0.289 0.134 0.270 0.111 0.271 0.119 0.268 0.128 DP 0.173 0.036 0.184 0.058 0.193 0.056 0.208 0.065 0.183 0.072 H 40.02 29.52 19.41 15.175 18.97 18.02 35.35 14.76 36.87 15.03 H5% 13.72 9.406 11.89 9.142 11.53 6.781 14.18 6.048 14.56 7.165 ∆k 6.632 4.094 5.039 3.000 4.844 2.462 6.044 2.040 6.562 2.576 f2d6 2.851 1.066 2.498 1.294 2.467 0.900 2.625 0.895 2.582 0.983 Sk

k=1

2.584 1.005 2.315 0.990 2.316 0.877 2.471 0.866 2.432 0.966 Sk

k=2

4.270 2.323 3.725 2.361 3.690 1.819 4.172 1.667 4.281 2.029 Ψ 0.213 0.041 0.181 0.044 0.173 0.032 0.224 0.032 0.222 0.038

4. CONCLUSION AND FUTURE WORK

This report, explained three proposed methods, In the first proposed method [1], several referenced-based boundary detection evaluations are detailed, pointing their advantages and disadvantages through concrete examples of edge images. A new normalized supervised edge map quality measure is proposed, comparing a ground truth contour image, the candidate contour Page 63

SLIDE 68 Measure Score Measure Score Measure Score Φ∗ 0.519 SFoM 0.702 θδT H=5 0.993 χ2∗ 0.946 MFoM 0.814 ΩδT H=1 0.587 P ∗ m 0.883 DP 0.117 ∆k 0.995 F ∗ α 0.790 Υ 0.618 f2d6 0.987 SSIM 0.355 H 0.999 Sk k=1 0.982 FoM 0.629 H5% 0.999 Sk k=2 0.992 F 0.800 Dk k=2 0.049 Γ 0.812 d4 0.690 θδT H=1 0.989 Ψ 0.814 Measure Score Measure Score Measure Score Φ∗ 0.563 SFoM 0.601 θδT H=5 0.995 χ2∗ 0.928 MFoM 0.770 ΩδT H=1 0.799 P ∗ m 0.847 DP 0.139 ∆k 0.995 F ∗ α 0.775 Υ 0.527 f2d6 0.986 SSIM 0.282 H 0.999 Sk k=1 0.978 FoM 0.540 H5% 0.999 Sk k=2 0.992 F 0.723 Dk k=2 0.0714 Γ 0.597 d4 0.645 θδT H=1 0.989 Ψ 0.602 (a) Gt / Original [4] (b) Sobel (c) Canny [?] Measure Score Measure Score Measure Score Φ∗ 0.571 SFoM 0.574 θδT H=5 0.995 χ2∗ 0.920 MFoM 0.762 ΩδT H=1 0.814 P ∗ m 0.835 DP 0.143 ∆k 0.995 F ∗ α 0.717 Υ 0.509 f2d6 0.986 SSIM 0.274 H 0.999 Sk k=1 0.977 FoM 0.523 H5% 0.999 Sk k=2 0.992 F 0.701 Dk k=2 0.081 Γ 0.532 d4 0.632 θδT H=1 0.990 Ψ 0.539 Measure Score Measure Score Measure Score Φ∗ 0.625 SFoM 0.468 θδT H=5 0.996 χ2∗ 0.907 MFoM 0.715 ΩδT H=1 0.979 P ∗ m 0.812 DP 0.187 ∆k 0.995 F ∗ α 0.683 Υ 0.421 f2d6 0.983 SSIM 0.219 H 0.999 Sk k=1 0.974 FoM 0.431 H5% 0.999 Sk k=2 0.992 F 0.654 Dk k=2 0.108 Γ 0.326 d4 0.599 θδT H=1 0.989 Ψ 0.363 Measure Score Measure Score Measure Score Φ∗ 0.669 SFoM 0.378 θδT H=5 0.995 χ2∗ 0.903 MFoM 0.686 ΩδT H=1 0.986 P ∗ m 0.805 DP 0.211 ∆k 0.996 F ∗ α 0.673 Υ 0.338 f2d6 0.978 SSIM 0.198 H 0.999 Sk k=1 0.970 FoM 0.371 H5% 0.999 Sk k=2 0.991 F 0.622 Dk k=2 0.120 Γ 0.203 d4 0.604 θδT H=1 0.988 Ψ 0.260 (d) Steerable filters [11] (e) Steerable filters [?] (f) Half Gaussian kernels [13]

Fig. 4: Comparison measures of different edge detections. A score close to 0 indicates a good edge map whereas a score 1

translates a poor segmentation.

(a) SNR= ∞ (b) SNR= 6 dB. (c) SNR= 5 dB. (d) SNR= 4 dB. RMSE ∞ 11 10 9 8 7 6 5 4 3.53 2.5 5 10 15 20 25 SNR Xia [5] Shui [?] Fast [?] Kitchen [?] Shi-Tomasi [?] F¨

rstner [?]

Harris [?] K¨

the [?]

Aach [?] New method ∞ 1 1 1 9 8 7 6 5 4 3 . 5 3 2 . 5 5 10 15 20 25 SNR (e) σ = ση = 1, ρ = 1, σξ = 3. (f) σ = ση = 2, ρ = 1, σξ = 3. (g) Original real image. (h) Our proposed method corner detection for real image.

Fig. 5: RMSE in function of the noise level. (a)-(d) contains the corner detected by the new algorithm with L=3 and p=5.

image and their associated spacial nearness. The strategy to normalize the evaluation enables to consider a score close to 0 as a good edge map, whereas a score 1 translates a poor segmentation. In the second proposed method [2], presents a review of supervised edge detection assessment methods in details. Moreover, based on the theory of these dissimilarity evaluations, a technique is proposed to evaluate filtering edge detection methods involving the minimum score of the considerate measures. Indeed, to evaluate an edge detection technique, the result which obtains the minimum score of a measure is considerate as the best one and represents an objective evaluation. For the proposed method [3], A novel method for corner detection based on the combination of directional derivative and homogeneity kernels has been proposed in this paper. The half Gaussian kernels (HGK) allow to detect a relevant strength of a corner while the IRON indicates the more homogeneous directions. A such combination enables to remove undesirable corners near contour areas which could be considered as a ideal feature by the HGK. We plan in a future study to compare the robustness several edge detection algorithms by adding noise and blur on real images presented in the supplementary material and then using the optimum threshold computed by the minimum of the evaluation. Page 64

SLIDE 69

REFERENCES [1] Hasan Abdulrahman, Baptiste Magnier and Philippe Montesinos. A new normalized supervised edge detection evaluation. In Proceedings of the 8th Iberian International Conference on Pattern Recognition and Image Analysis (IbPRIA), 10 pages, published in Springer Lecture Notes in Computer Science Series, Faro, Portugal, June 20-23, appeared in 2017. [2] Hasan Abdulrahman, Baptiste Magnier and Philippe Montesinos. From contours to ground truth: How to evaluate edge detectors by filtering. In Proceedings of the 25th International Conference in Central Europe on Computer Graphics, Visu- alization and Computer Vision (WSCG), 10 pages, Pilsen, Czech Republic, May 29 - June 2, appeared in 2017. [3] Hasan Abdulrahman, Baptiste Magnier and Philippe Montesinos. Oriented half kernels for corner detection. In Proc. of the 25th European Signal Processing Conf. (EUSIPCO), Kos Island, Greek, August 28 - September 2, appeared in 2017. [4] M. D. Heath, S. Sarkar, T. Sanocki, and K. W. Bowyer, robust visual method for assessing the relative performance of edge-detection algorithms. IEEE TPAMI, vol. 19, no. 12, pp. 1338-1359, 1997. [5] G.S. Xia, J. Delon, and Y. Gousseau, Accurate junction detection and characterization in natural images. IJCV, vol. 106,

no. 1, pp. 3156, Springer, 2014.

[6] B. Magnier, A. Le, and A. Zogo, quantitative error measure for the evaluation of roof edge detectors, in IEEE IST, pp. 429-434, 2016. [7] N.L. Fernandez-Garca, R. Medina-Carnicer, A. Carmona-Poyato, F.J. Madrid-Cuevas, and M. Prieto-Villegas. Character- ization of empirical discrepancy evaluation measures. Patt. Rec. Lett., 25(1): 3547, 2004. [8] S. Chabrier, H. Laurent, C. Rosenberger, and B. Emile. Comparative study of contour detection evaluation criteria based

n dissimilarity measures. EURASIP J. on Image and Video Processing, 2008.

[9] F. Michelet, J.-P. Da Costa, O. Lavialle, Y. Berthoumieu, P. Baylou, and C. Germain. Estimating local multiple orientations. Signal Proc., 87(7):16551669, 2007. [10] D. Martin, C. Fowlkes, D. Tal, and J. Malik. A database of human segmented natural images and its application to evaluating segmentation algorithms and measuring ecological statistics. In IEEE ICCV, volume 2, pages 416423, 2001. [11] William T. Freeman and Edward H Adelson. The design and use of steerable filters. IEEE Trans. on Pattern Analysis and Machine Intelligence, Vol.13(9):pp.891906, 1991. [12] J.M. Geusebroek, A. Smeulders, and J. van de Weijer. Fast anisotropic gauss filtering. ECCV 2002, pages 99112, 2002. [13] B. Magnier, P. Montesinos, and D. Diep. Fast Anisotropic Edge Detection Using Gamma Correction in Color Images. In IEEE ISPA, pages 212217, 2011. Page 65