=Paper=
{{Paper
|id=Vol-2528/3_Ferilli_et_al_AI3_2019
|storemode=property
|title=An analysis of the Avetrana Murder Case through Abstract Argumentation
|pdfUrl=https://ceur-ws.org/Vol-2528/3_Ferilli_et_al_AI3_2019.pdf
|volume=Vol-2528
|authors=Stefano Ferilli,Fabio Leuzzi
|dblpUrl=https://dblp.org/rec/conf/aiia/FerilliL19
}}
==An analysis of the Avetrana Murder Case through Abstract Argumentation==
https://ceur-ws.org/Vol-2528/3_Ferilli_et_al_AI3_2019.pdf
An analysis of the Avetrana Murder Case
through Abstract Argumentation
Stefano Ferilli1 and Fabio Leuzzi2
1
Dipartimento di Informatica – Università di Bari, Bari, Italy
stefano.ferilli@uniba.it
2
Italian National Police, Rome, Italy
fabio.leuzzi@poliziadistato.it
Abstract. Argumentation is the activity by which one justifies his own
position in a discussion. It plays a fundamental role in court trials, for
faithfully reconstructing the course of the events from evidence and tes-
timonies. So, the availability of automated techniques for carrying out
argumentation would be extremely useful in that domain. Abstract ar-
gumentation focuses on the inter-relationships among the available ar-
guments, neglecting their internal structure or specific interpretation. In
this paper we report about the application of abstract argumentation
techniques to a very famous criminal case happened in Italy. Our aim
is, on one hand, showing a practical and relevant application of abstract
argumentation, and, on the other, comparing the outcome of the auto-
mated approach to that of the judges in the real court.
Keywords: Machine Reasoning, Argumentation, Forensics
1 Introduction & Related works
Argumentation is ubiquitous in our everyday life, for trivial issues as well as for
important ones. It is the activity by which one justifies his own position in a
discussion. When such a position is not agreed upon by others, there is a need
to settle the disagreement among di↵erent, contrasting positions. This is done by
carrying out a debate in which each participant tries to support his own position
with suitable arguments, also attacking the arguments put forward by others to
support their position, and of course defending his position from the attacks of
the others. More formally, argumentation is an inferential strategy for practical
and uncertain reasoning able to cope with partial and inconsistent knowledge.
Its goals are to identify, analyze and evaluate arguments.
One traditional application for argumentation is for settling legal disputes.
In court trials, judges must make their decisions after considering both a body of
evidence and a set of testimonies given by witnesses. While evidence, especially
scientific one, is the main foundation for a decision, it is often insufficient, which
calls for the evaluation of testimonies in order to obtain a reconstruction of the
events of the case as faithful as possible. Sometimes the best available explana-
tion is sufficient for the judge to take his decision, sometimes a proof beyond
Copyright c 2019 for this paper by its authors. Use permitted under Creative Commons
License Attribution 4.0 International (CC BY 4.0).
any reasonable doubt is required. Identifying the truth among all testimonies is
not easy for several reasons:
– there are often many di↵erent positions, and the relationships of the tes-
timonies among each other, and among them and the positions, are often
indirect and unclear;
– some testimonies are purposely deceptive in order to favor one of the posi-
tions;
– even when eye witnesses are truly committed to faithfully reporting what
they saw, human memory tends to be inherently unreliable when recalling
past events.
For these reasons, and given the relevance of taking the most faithful position,
automated techniques for carrying out argumentation would be extremely useful
both in general, and in court trials in particular.
Indeed, a lot of literature is available about application of argumentation to
legal reasoning. Just to mention a few examples, [1] shows how it can be more
appropriate than deduction for handling the need to support and justify some
interpretations of a case against others, and review several systems aimed at
implementing di↵erent aspects of argumentative reasoning. [2] presents a set of
argument schemes that encapsulate preferences between social values revealed
in past decisions to decide new cases. [3] proposes the use of abstract argumen-
tation in case-based reasoning, resulting in a methodology to produce decision
and explanations for new cases. [4] proposes a method to encapsulate the knowl-
edge of bodies of case law using Abstract Dialectical Frameworks (ADFs), and
for producing programs that decide the cases. [5] develops and discusses an Ar-
gumentation Framework to represent case law. [6] investigates natural-language
argumentation in the case law domain, aimed at formalizing the structure of
argumentation as a context-free grammar.
This paper reports about the application of abstract argumentation tech-
niques to analyze a very famous criminal case happened in Italy. In the case,
most of the available knowledge of the events was in the form of testimonies, in
which deception and falsity were widespread. Everybody in Italy knows about
this case, because it was emotionally enthralling since it concerned the murder
of a young girl, and got wide coverage by the media. Actually, while the formal
decision for the Italian law is definitive, many people are not fully convinced of
the final outcome, nor about the reconstruction of the events that was considered
as the most reliable to take that decision. Our aim is, on one hand, showing a
practical and very relevant application of abstract argumentation, and, on the
other, trying to explain on a formal basis the decision of the judges in the Court.
This paper is organized as follows. After recalling the basics of the formal
approach to argumentation we will use, and some other works concerned with
application of automated argumentation to real court cases, we will present
the story, then we will formalize it in an abstract argumentation framework,
and finally we will apply several di↵erent kinds of strategies to determine the
winning arguments, discussing their features and outcomes.
31
2 Background
One of the most influential computational models of argument, developed by re-
search in Artificial Intelligence (AI), is Dung’s Abstract Argumentation Frame-
work [7], where the internal structure or specific interpretation of arguments is
ignored, and only their inter-relationships are taken into account, representing
the exchange of arguments and counter-arguments in a dispute.
Definition 1 An Argumentation Framework (AF) is a pair F = hA, Ri, where
A is a finite set of arguments and R ✓ A ⇥ A is the attack relation.
↵R means that ↵ attacks , or is attacked by ↵. A set S ✓ A attacks
2 A, denoted SR , if 9↵ 2 S : ↵R .
Basically, an AF is a directed graph where the nodes are abstract arguments
and the directed edges correspond to attacks between them.
An argumentation semantics is a formal criterion that determines which
subsets of arguments in an AF, called extensions in extension-based semantics [7],
are ‘collectively acceptable’.
Definition 2 Let F = hA, Ri be an AF, an extension-based semantics S asso-
ciates F with a subset of 2A , denoted by ES (F ).
So, many alternative extensions may exist for a given AF, as well as no extension
at all (i.e., ES (F ) = ;). DS denotes the set of AFs where an extension for
semantics S exists.
A central concept is acceptability, or defense: a (set of) argument(s) defends
an argument if it attacks all of its attackers.
Definition 3 Given an AF F = hA, Ri, S ✓ A and ↵ 2 A, ↵ is defended by S
i↵ 8 2 A : R↵ ) SR .
The basic concept shared by all argumentation semantics in the literature is
conflict-freeness: if an argument attacks another argument, then they cannot be
both in the same extension.
Definition 4 Given an AF F = hA, Ri, S ✓ A is conflict-free i↵ @↵, 2 S s.t.
↵R . The collection of all conflict-free sets for F is denoted by cf(F ).
; is conflict-free. A further requirement is admissibility: an extension should
withstand the attacks it receives from other arguments by replying with other
attacks.
Definition 5 Given an AF F = hA, Ri, S ✓ A is admissible if S is conflict-
free and S is defended by itself, i.e. 8 2 A s.t. R↵ : 9 2 S s.t. R . The
collection of all admissible sets is denoted by adm(F ).
Note that ; is not only conflict-free, but also defended by itself.
32
Definition 6 Given an AF F = hA, Ri, a semantics S satisfies the admissibil-
ity principle i↵ 8F 2 DS : ES (F ) ✓ adm(F ), namely
8E 2 ES (F ) : ↵ 2 E ) (8 2 A : R↵ ) ER ).
Another essential property is reinstatement: if the attackers of an argument
↵ are in turn attacked by an extension E, one may assume that they have no
e↵ect on ↵: then ↵ should be, in a sense, reinstated, therefore it should belong
to E.
Definition 7 A semantics S satisfies the reinstatement principle i↵
8F 2 DS , 8E 2 ES (F ) : (8 2 A : R↵ ) ER ) ) ↵ 2 E.
The justification state of an argument ↵ can be conceived in terms of its
extension membership. A basic classification [8] encompasses only two possible
states for an argument (justified or not justified). In this respect, two alternative
types of justification, skeptical or credulous, can be considered.
Definition 8 Given a semantics S and an AF F 2 DS , an argument ↵ is:
– skeptically justified i↵ 8E 2 ES (F ) : ↵ 2 E;
– credulously justified i↵ 9E 2 ES (F ) : ↵ 2 E.
Let us now introduce the four ‘traditional’ semantics, considered in Dung’s
original paper [7], that we will use in this work. All of them are built on the
concept of admissible sets.
Complete Semantics The notion of complete extension [7] is based on the
principles of admissibility and reinstatement: a complete extension is a set which
is able to defend itself and includes all arguments it defends.
Definition 9 Given an AF F = hA, Ri, S ✓ A is a complete extension i↵ S
is admissible and every argument of A which is defended by S belongs to S. The
collection of all complete extensions is denoted by comp(F ).
Note that the empty set is admissible and initial arguments are acceptable
with respect to the empty set. Due to reinstatement, any complete extension
includes not only the initial arguments, but also the arguments they defend,
those which are in turn defended by them, and so on. By definition, any complete
extension is admissible.
Grounded Semantics Considering each complete extension as a reasonable
solution, the grounded extension is the set of arguments shared by all these
reasonable solutions.
Definition 10 Given an AF F = hA, Ri, a set S ✓ A is a grounded extension
i↵ S 2 comp(F ) ^ @T 2 comp(F ) s.t. T ⇢ S. The set of grounded extension is
denoted by grd(F ).
33
For each AF there is exactly one grounded extension. As the set of com-
plete extensions is non-empty and the grounded extension is just the ✓-minimal
complete extension, the grounded extension always exists. Also, any grounded
extension is a complete extension.
Preferred Semantics Preferred semantics seek to maximize the accepted ar-
guments, taking into account maximal admissible sets.
Definition 11 Given an AF F = hA, Ri, S ✓ A is a preferred extension i↵
S is admissible and @T 2 adm(F ) s.t. T S. The collection of all preferred
extensions is denoted by pref(F ).
Any preferred extension E is also a complete extension. Indeed, preferred
extensions may Tbe equivalently defined as maximal complete extensions. Note
that grd(F ) = E2comp(F ) E for any AF F . However, the grounded extension
does not coincide with the intersection of all preferred extensions in general.
Stable Semantics Stable semantics relies on the intuition that an extension
should be able to attack all arguments not included in it and leave no argument
“undecided”, i.e., an argument must either be in the extension or be attacked
by an argument of the extensions.
Definition 12 Given an AF F = hA, Ri, S ✓ A is a stable extension i↵ S is
conflict-free and 8↵ 2 A : ↵ 2
/ S ) SR↵. The collection of all stable extensions
is denoted by stb(F ).
Any stable extension E is also a complete extension and a maximal conflict-
free set of F . It follows that any stable extension is also a preferred extension, but
not vice-versa. There are argumentation frameworks where no stable extension
exists.
The classical semantics proposed above su↵er from significant limitations in
expressive power, due to their being able to express only unweighted attacks
between arguments. Among the extended frameworks proposed in the literature
to overcome this limitation, we recall here:
– the bipolar one [9], allowing to express also a support relation between ar-
guments, in addition to attacks;
– the weighted one [10], allowing to set weights on arcs, expressing the strenght
of the attack;
– the value-based one [11], allowing to set a precedence ordering on arguments
depending on the values they express;
– the bipolar weighted one (BWAF) [12], allowing to express both attacks and
supports between arguments, plus weights on them;
– the trust-based BWAF (T-BWAF) [13], which adds to BWAFs the possibility
of expressing the strength of the single arguments, by placing weights also
on the nodes of the argumentation graph.
34
3 The Avetrana Case and its Formalization
For the purposes of this paper, we chose to apply abstract argumentation to
a very famous criminal case happened in Avetrana, a town in the province of
Taranto, Italy, from which it drew its name. The case generated much ado in
Italy, for several reasons: the peculiarities of the story, the quite confusing in-
formation available about the events, the wide coverage it received from the
media, and, last but not least, the strong debate among common people about
the responsibility of the crime.
In a nutshell, the case started on August 26th, 2010, when a young girl, Sarah
Scazzi, was missing. At first, her parents and relatives were questioned, in order
to collect as much information as possible, and devise an inquiry strategy to
understand what happened and hopefully find her. Very early, Sarah’s mother,
Concetta, suggested to focus on the Misseri family: uncle Michele, aunt Cosima
(Concetta’s sister), and their daughter, Sabrina. Indeed, the phone logs indicated
that Sarah was at their place before disappearing. Since the very beginning, their
account of the events was strange, lacking details and full of inconsistencies,
which added to the suspects about their being involved in the case. After several
weeks of interrogatories, lies and deceptions, the corpse of Sarah was found in a
well in the country nearby. Many debates, tests and counter evidence followed
this tragic event, until, on February 21st, 2017, after 3 court levels, the case was
definitely closed condemning Cosima and Sabrina for the murder, and Michele
for concealing the corpse. The Court determined that the reason that triggered
Sabrina to kill Sarah, aided by her mother, had to do with her jealousy for a
guy named Ivano.
The very short account we just provided for this case cannot express, let
alone suggest, the intricacy of the case, due to inconsistency among di↵erent
testimonies, sometimes even among testimonies given by the same person (es-
pecially uncle Michele), and between testimonies and evidence found during the
inquiry. Indeed, especially the Misseris, tried to protect each other from being
charged of the murder, but without devising a consistent, albeit fake, version of
the events. So, they just tried to mess up the inquiry by declaring something,
then retracting, or changing, or even contradicting their own declarations.
The following paragraphs describe how we took the most relevant excerpts
from the documentation publicly available about the intermediate and final pro-
cesses, decomposed it into a series of atomic claims, and identified attack relation-
ships among these claims. Adopting a divide et impera approach, we partitioned
the claims by topic, each of which we called a panel. The 5 resulting panels are:
1. the relationship between Sabrina and Ivano;
2. Sabrina’s jealousy for Ivano;
3. the morning of Sarah’s disappearance;
4. the days immediately following Sarah’s disappearance;
5. Michele’s claims.
Table 1 summarizes the extracted claims (column Argument), reporting for each
a unique identifier (#), its source (Source), the element of the partition to which
35
it belongs (Panel ), and the other arguments it attacks (Attacks). Figure 1 shows
the corresponding argumentation graph Avetrana-case. Note the independence
among panels.
# Source Panel Argument attacks
1 Sabrina 1 Ivano is just a friend of her -
2 Sabrina’s friends 1 Sabrina is in love with Ivano 1
3 Sabrina’s friends 1 Sabrina was obsessed by Ivano so much that she spied him 1
4 Sabrina’s friends 1 Sabrina and Ivano were close 1
5 Technical report 1 Sabrina and Ivano sent about 4500 SMS each other -
6 Sabrina 2 The evening of August 25th (the day before Sarah disappearance) she didn’t -
fight with Sarah
7 Sabrina’s friends 2 The evening of August 25th Sabrina had a fight with Sarah because of Sab- 6
rina’s jealousy for Ivano
8 Sarah’s diary 2 Sabrina was angry with her due to Ivano’s attentions for her 6
9 Sabrina 2 She admits that the relationship with Sarah was not relaxed. 6
10 Sabrina 3 On August 26th (Sarah’s disappearance day) she had no appointment with -
Sarah to go to the beach
11 Technical report 3 The SMS log provides evidence that there was an appointment among Sab- 10
rina, Sarah and their friend Mariangela to go to the beach on August 26th
12 Sabrina 3 She claims that she was inside the home when Mariangela arrived at Sab- -
rina’s home to go to the beach
13 Mariangela 3 She claims that when arrived found Sabrina outside her home 12
14 Sabrina 3 In presence of Mariangela, because of Sarah’s delay, she called Sarah’s home, -
without success
15 Sabrina 3 During the call of argument 14, her father Michele is standing on the entrance -
of the garage
16 Mariangela 3 During the call of argument 14, Sabrina’s father Michele was not present 15
17 Michele 3 He was into the garage with Sarah’s corpse, and he heard Sarah’s phone 15
ringing because of Sabrina’s calls
18 Michele 3 He saw Sabrina only after her return home from Sarah’s home, where she 15
went in search of Sarah
19 Concetta 3 Sabrina came to her home looking for Sarah, but Sarah had already left to -
go to Sabrina’s home, hence she suggested to call her parents Michele and
Cosima to stop there Sarah in case of arrival
20 Concetta 3 Sabrina claimed that nobody was at home -
21 Cosima 3 During the facts reported by these arguments she was sleeping 20
22 Cosima 3 On August 26th she worked 21
23 Michele 3 He heard his wife and daughter fighting during the morning 22
24 Cosima’s colleague 3 On August 26th Cosima and her were unable to go at work because the public 22
transports were crowded
25 Cosima’s neighbour 3 Cosima’s car was not on the roadside 22
26 Cosima 3 After returning home from work, she went to sleep 23,24
27 Donato 3 Between 2:00 PM and 2:20 PM, he saw Cosima’s car in the neighbourhood 26
of via per Mare at high speed and followed by a blue van
28 Giovanni 3 He saw Cosima intimating Sarah to get in her car 26
29 Technical report 3 There is evidence that a call from Cosima’s phone to Michele’s phone is -
geolocalized in the Misseris’ garage
30 Technical report 4 Phone calls happened between 10:26 AM and 10:40 AM on August 27th 31
(Sarah’s disappearance day) confirm that Cosima and Sabrina were in a zone
compatible with the well in Contrada Mosca where the corpse has been found
31 Cosima 4 During the morning of August 27th she went to check the wine strength in -
a Misseri family’s property, far away from Contrada Mosca
32 Concetta 4 In the morning of August 27th, at 11:00 AM, Cosima and Sabrina went to -
visit her
33 Michele 4 He found Sarah’s phone in a farm several km away from Avetrana on Septem- -
ber 29th
34 Michele 5 He killed Sarah, indicating exactly the place where he, alone, concealed the 35,36
corpse (declared on October 6th)
35 Michele 5 Sabrina murdered Sarah after a joke that became a fight (declared on Octo- 34,37
ber 15th)
36 Michele 5 Sabrina murdered Sarah and called him to ask for help to conceal her corpse 34,37
37 Michele 5 He wrongly accused his daughter, he confesses to be the murderer taking the 35,36
blame for the crime (declared on December 23th)
Table 1. Arguments and attacks in our formalization of the Avetrana case
Panel 1 While Sabrina always denied that she was close to Ivano (1), SMS logs
reveal that they sent each other more than 4.500 messages (5), and indeed their
friends declared that she was in love with him (2). Her obsession made her even
occasionally spy him (3). Furthermore they had intercourse (4). So arguments 2,
3 and 4 attack argument 1. Argument 5 is not necessarily a proof of sentimental
involvement between the two, anyway for some it could be. The choice has been
to leave argument 5 without arcs.
36
Fig. 1. Our Avetrana case argumentation graph
Panel 2 Sabrina initially denied that she argued with Sarah the evening before
her disappearance (6), but her friend declared the opposite (7), and the morning
she disappeared Sarah wrote in her diary: “[...] Yesterday evening I went out
with Sabrina and Mariangela. We were in a brasserie for a fast Red Bull, then
we came back home and Sabrina, as usual, got angry because, in her opinion,
when Ivano is with us, I spend a lot of time with him. Obviously. At least he,
unlike her, is gentle to me. I would love to have a boyfriend like him! Anyway,
it doesn’t matter, I’m used to.” (8) Later, Sabrina will admit that she was not
at peace with Sarah (9). So arguments 7, 8 and 9 attack argument 6.
Panel 3 Regarding the day Sarah disappeared, Sabrina declared that they had
no appointment to go to the beach (10). Conversely, an appointment had been
made the previous night with Mariangela (11), and further confirmed in the
morning. Then, argument 11 attacks argument 10.
Sabrina always denied that, at the moment Mariangela arrived, she was al-
ready outside her home (12). Mariangela declared the opposite (13). Argument
13 attacks argument 12.
Sabrina declared that, after Mariangela arrived, she called Sarah’s home to
check if whether she was there, given that she was late, but nobody answered
to the phone (14). According to her declarations, during the calls her father
Michele was near the garage entrance (15). Conversely, Mariangela declared that
she never saw Michele there (16). Michele, in one of his inconsistent versions,
declared that he was in the garage with the corpse, that he heard Sarah’s phone
ringing several times, and that those were Sabrina’s calls (17). According to
Michele’s declarations, he will see Sabrina only after she came back home from
Sarah’s home, where she went to check the presence of Sarah (18). So, arguments
16, 17 and 18 attack argument 15, defeating Sabrina’s claims.
37
Sarah’s mother, Concetta, claimed that Sabrina went at her home looking
for Sarah, and she informed Sabrina that Sarah had already left to her home,
and that it could be a good idea to warn her parents that Sarah, maybe soon,
should arrive, and that they should stop Sarah there (19). Sabrina answered that
nobody was at home (20), but Cosima (Sabrina’s mother) will declare that she
was sleeping at home (21). Argument 21 attacks 20.
Anyway, Cosima, in other declarations, asserted that she went at work the
morning Sarah disappeared (22). So, argument 22 attacks 21.
In a further version provided by Michele, Sabrina and Cosima had a fight that
morning (23). In general, the presence at home of Cosima is supported by the
declaration of a colleague of her, saying that public transportation was packed,
so they were unable to go to work, and they went back home (24). However, a
neighbor of Cosima declared that her car was missing from the place where it
was used to be (25). Arguments 23, 24 and 25 attack argument 22.
Cosima declared that when she came back from work, she went to sleep (26),
which is in contrast with 23 and 24. Anyway, this declaration has been refuted
by Donato, who saw Cosima’s car in Via per Mare street, between 2:00 PM and
2:20 PM, keeping a high speed and joined with a blue van (27).
The version of Cosima sleeping is contrasted from Giovanni the florist, who
declared he saw Cosima following Sarah and ordering her to get on her car (28).
However, a couple of day later he retracted this version, claiming that it was
just a dream. Standing on such arguments, 27 and 28 attack 26.
A milestone is represented by the technical report of the ROS (the Special
Operating Department of the Carabinieri corps, specialized in investigations),
that localized a call from Cosima’s phone to her husband at 3:25:04 PM (29).
Panel 4 There are other phone calls that have a central role in the case evalu-
ation, happened the day after Sarah’s disappearance, between 10:26 and 10:40
AM. These calls locate Sabrina and Cosima near the well in which Sarah’s corpse
was concealed (30). About that morning, Cosima declared that she went to check
the strength of the wine that her family produces, in a completely di↵erent place
than Contrada Mosca, where the well is located (31). Note that Cosima never
mentioned what Sabrina was doing in the meanwhile. Argument 30 is a fact, so
30 attacks 31.
In the morning of the day after Sarah’s disappearance, at 11:00 AM, both
Cosima and Sabrina went to Concetta’s home (32). No contradictions emerged
about this.
Panel 5 Let us turn to Michele’s declarations. He firstly claimed that he found
Sarah’s phone, three days after Sarah’s disappearance, in a farm few kilometers
away from Avetrana (33). This clumsy attempt to mislead investigations made
the police insist on having a deeper understanding of the dynamics of the case.
Under such a pressure, Michele admitted the murder, indicating exactly the lo-
cation of the corpse, which was soon verified (34). The lack of precision and
coherence of Michele’s declarations always caused suspects about his honesty.
The police questioned Michele several other times, until, nine days after Sarah’s
38
Semantics Extensions
stable {2 3 4 5}
admissible {} {5} {4} {4 5} {3} {3 5} {3 4} {3 4 5}
{2} {2 5} {2 4} {2 4 5} {2 3} {2 3 5}
{2 3 4} {2 3 4 5}
preferred {2 3 4 5}
complete {2 3 4 5}
grounded {2 3 4 5}
Table 2. Semantics for panel 1.
disappearance, he retracted the murder confession, accusing his daughter Sab-
rina, telling that the unfortunate event had been the consequence of a strong
quarrel between them (35). Not even this version will turn out to be completely
true, since Michele will add that he had been called by Sabrina after the murder
to conceal the corpse (36). But he will retract again after several days, undertak-
ing responsibility for the crime (37). Argument 34 attacks 35 and 36, argument
35 attacks 34 and 37, whereas 36 attacks 34 and 37, and finally 37 attacks 35
and 36.
4 Application of Abstract Argumentation to the Case
Exploiting standard available tools, we computed the extensions for several se-
mantics for the Avetrana case argumentation framework.
4.1 Classical Semantics
Let us quickly recall that:
– admissible extensions are conflict-free subsets, containing arguments that
defend each other;
– the preferred extension is the largest admissible extension;
– the stable extension is composed by the conflict-free subset that attack ar-
guments outside the extension;
– the complete extension contains the admissible extensions leaving out the
attacked or not defended arguments;
– the grounded extension is the minimal complete extension (so that it is
unique);
– nodes without attacks (incoming or outcoming) are not relevant for exten-
sions computation and semantically coherent with every extension considered
to analyze the case.
While reporting the results of all these semantics in Tables 2-6, due to space
constraints, in the following we will comment the grounded extension only, since
it can be considered as a reliable core of arguments that is likely to be true.
39
Semantics Extensions
stable {7 8 9}
admissible {} {9} {8} {8 9} {7} {7 9} {7 8} {7 8 9}
preferred {7 8 9}
complete {7 8 9}
grounded {7 8 9}
Table 3. Semantics for panel 2.
Semantics Extensions
stable {11 13 14 16 17 18 19 21 23 24 25 27 28 29}
admissible {} {29} {25} {25 29} {19} {19 29} {19 25}
{19 25 29} {18} {18 29} {18 25}
{18 25 29} {18 19} {18 19 29} {18 19 25}
...
{11 13 14 16 17 18 19 21 23 24 25 27 28}
{11 13 14 16 17 18 19 21 23 24 25 27 28 29}
There are totally 12288 admissible extensions
preferred {11 13 14 16 17 18 19 21 23 24 25 27 28 29}
complete {11 13 14 16 17 18 19 21 23 24 25 27 28 29}
grounded {11 13 14 16 17 18 19 21 23 24 25 27 28 29}
Table 4. Semantics for panel 3.
Table 2 summarizes the relationship between Sabrina and Ivano. Translating
the grounded extension we obtain: “Sabrina was in love with Ivano and ob-
sessed by him, up to follow and control him. Sabrina and Ivano sent each other
about 4500 SMS. They had an a↵air.” The grounded extension is also stable: its
arguments attack the arguments not belonging to it.
Table 3 summarizes the jealousy of Sabrina for Ivano. The grounded extension
says that: “The evening of August 25th, Sabrina was angry with Sarah because
of the attentions of Ivano for her. Such an interest of Ivano for Sarah, and the
fact that Sarah appreciated it, are confirmed by Sarah’s diary and by Sabrina’s
interrogatories.” Just like in panel 1, the grounded extension is also stable.
Table 4 concerns the morning of Sarah’s disappearance. The grounded ex-
tension says: “There was an appointment among Sabrina, Sarah and Mariangela
to go to the beach on August 26th. Mariangela went to Sabrina’s to go to the
beach and found her on the street. Sabrina told her she could not find Sarah,
and called Sarah’s home phone, getting no answer. At that moment, Mariangela
did not see Michele (that was in the garage with the corpse). Sarah’s mobile
phone could be heard ringing from the garage, given that both her corpse and
her phone were there, and Sabrina was calling to cheat Mariangela. Michele and
Sabrina met later, only after Sabrina came back from Sarah’s home. During the
visit of Sabrina to Sarah’s home, Concetta told Sabrina that Sarah already left
to reach Sabrina’s home, so Concetta suggested Sabrina to inform her parents
that Sarah could arrive to them at any moment. Cosima declared that, in the
40
Semantics Extensions
stable {30 32 33}
admissible {} {33} {32} {32 33} {30} {30 33} {30 32}
{30 32 33}
preferred {30 32 33}
complete {30 32 33}
grounded {30 32 33}
Table 5. Semantics for panel 4.
Semantics Extensions
stable {35 36} {34 37}
admissible {} {36} {35} {35 36}
{37} {34} {34 37}
preferred {34 37} {35 36}
complete {} {35 36} {34 37}
grounded {}
Table 6. Semantics for panel 5.
meanwhile, she was sleeping since in the early morning she had been unable to
go to work. Michele declared that he heard his wife and daughter having an
argument. Cosima’s car was not parked outside her home. Between 14:00 and
14:20 Cosima’s car was seen near via per Mare, in Avetrana, going at high speed
together with a blue van. Giovanni saw Cosima following Sarah and ordering her
to get on the car. At 15:25 a call from Cosima’s phone to Michele started.” The
extension is grounded and stable.
Table 5 regards the days after Sarah’s disappearance. The grounded extension
says: “On August 27th, the day after Sarah’s disappearance, between 10:26 and
10:40, Cosima and Sabrina can be located near the well in Contrada Mosca,
where the corpse will be found. Around 11:00, Cosima and Sabrina went to
Concetta’s home for a visit. On September 29th, Michele will find Sarah’s phone
in a farm a few kilometers away from Avetrana.” The extension is grounded and
stable.
Table 6 regards Michele’s declarations. The grounded extension is empty, i.e.,
no subset of information is a candidate to be true. Anyway, there are two stable
extensions. The former is: “ Michele killed Sarah and knows the place where is
the corpse. In fact, he claims that his initial accusation against his daughter is
false. He undertakes all responsibilities.” The latter is: “Michele claims that the
killer is his daughter, and that she called him only after the murder to ask for
help to conceal the corpse.” Note that Michele’s contradictions make evident
some limitations of the basic approach of Dung’s argumentation framework.
In order to understand which subset of information can be considered less
likely than others, the need to weigh the attacks arises. So, we evaluated panel
5 exploiting the weighted argumentation framework.
41
Semantics Extensions
stable {35 36}
admissible {} {36} {35} {35 36}
preferred {35 36}
complete {} {35 36}
grounded {}
Table 7. Semantics for panel 5 using a weighted argumentation framework.
Extension Sets
grounded {2 3 4 5 7 8 9 11 13 14 16 17 18 19 21 23 24 25
27 28 29 30 32 33}
stable {2 3 4 5 7 8 9 11 13 14 16 17 18 19 21 23 24 25
27 28 29 30 32 33 35 36}
{2 3 4 5 7 8 9 11 13 14 16 17 18 19 21 23 24 25
27 28 29 30 32 33 34 37}
Table 8. Semantics for the entire graph using a value argumentation framework.
4.2 Weighted argumentation framework on panel 5
Argument 34 was initially taken by the detectives as a possible, but not very
reliable, confession (he never appeared to be a reliable person). Arguments 35
and 36 are successive in time to 34. Given the results of other panels, confirming
the jealousy of Sabrina for Ivano, and that Ivano had gentle attentions for Sarah,
we decided to set weights of 0.3 for the attacks of argument 34 against arguments
35 and 36, and of 0.9 for attacks of 35 and 36 against argument 34. Later, Michele
retracted again his version, going back to accuse his daughter, a version that is
quite unreliable, so we set the weights of its attacks to 35 and 36 at 0.1, but
the weights of the opposite attacks at 0.9. Weights are defined using a personal
interpretation of arguments reliability on the base of its sense, and the state of
the art of the case at the date of declaration, a set of information that, for lack
of space and for scopes of the paper, is omitted here.
Table 7 reports the extensions that have been computed. Given the weights,
the only stable extension is composed by arguments 35 and 36: “Sabrina killed
Sarah and called Michele to ask for help to conceal the corpse”.
4.3 Value Argumentation Framework applied to the whole case
The value argumentation framework allows us to assign a type to each argument
and to define an order of preference on the defined types. By doing so, we can
make explicit if an argument refers to an evidence, that is: a document, a forensic
report, a witness. So, knowing that typically the preference follows exactly such
an order, we can insert it in the rules evaluated in the framework.
Table 8 reports the grounded and the stable extensions computed on the
entire graph. The outcome is the union, respectively, of the grounded and stable
42
extensions computed using Dung’s argumentation framework. So, its interpreta-
tion can be obtained just appending the interpretations of the five panels.
However, we note that the two possible stable extensions have as di↵erence
just the two possible versions provided by Michele, which highlights the same
weaknesses as Dung’s argumentation framework, and calls for a strategy to clar-
ify such cases (e.g. the weighted argumentation framework).
5 Conclusions and Comments
Argumentation is the activity by which a position is defended in a discussion
with other people. It is fundamental in court trials, in order to have a faithful
reconstruction of the course of the events from evidence and testimonies. Among
the automated techniques for carrying out argumentation developed by research
in Artificial Intelligence developed, abstract argumentation considers the inter-
relationships among the available arguments, neglecting their internal structure
or specific interpretation.
In this paper, we considered a very famous criminal case happened in Italy,
concerning the murder of a young girl. It was interesting for several reasons: it
involved very ambiguous testimonies, and attracted much attention due to its
story and to the wide media coverage it got. So, we extracted from the process
deeds an abstract argumentation formalization which turned out to be consis-
tent with the process outcomes, and computed a number of semantics on it.
We specifically commented the grounded one, as a reliable core of arguments
that is likely to be true. The automated outcomes confirmed most of the inter-
pretations of facts made in the Courts, but also highlighted the limitations of
traditional semantics when tackling cases with many contradictions. In any case,
they confirm that abstract argumentation tools may represent a valid support
for quickly highlighting consistent subsets of evidences and testimonies on which
basing further investigations and final decisions.
Of course, the quantity and quality of the selected arguments, and of the
attacks among them, determine the quality of the outcome. This points out
another possible use of abstract argumentation: providing a formal explanation
or justification of the reasoning carried out by the judges, allowing this way other
people (e.g., the defendant’s lawyers) to understand, and possibly criticize, the
formalization, by adding arguments or changing the attacks, in order to check
whether other interpretations of the available information are possible.
For this reason, our future work on this case will be devoted to extract
further useful claims concerning the case, and to identify di↵erent strategies for
determining when and how to set the attacks among arguments. Also, we plan
to apply more semantics, and complex abstract argumentation frameworks (e.g,
the bipolar, bipolar-weighted, and trusted-bipolar-weighted ones), to the case,
trying to overcome the limitations of classical approaches by introducing more
information in the schema. If successful, we also plan to apply these approaches
to other well-known, ambiguous or unsolved, criminal cases.
43
References
[1] Bench-Capon, T., Prakken, H., Sartor, G.: Argumentation in legal reasoning. In
Rahwan, I., Simari, G., eds.: Argumentation in Artificial Intelligence - Chapter
18. Springer (2009) 363–382
[2] Bench-Capon, T., Prakken, H., Wyner, A., Atkinson, K.: Argument schemes for
reasoning with legal cases using values. In: Proceedings of the 14th International
Conference on Artificial Intelligence and Law. ICAIL ’13, ACM (2013) 13–22
[3] Čyras, K., Satoh, K., Toni, F.: Abstract argumentation for case-based reasoning.
In: Proceedings of the 15th International Conference on Principles of Knowledge
Representation and Reasoning. KR’16, AAAI Press (2016) 549–552
[4] Al-Abdulkarim, L.M., Atkinson, K., Bench-Capon, T.: A methodology for de-
signing systems to reason with legal cases using abstract dialectical frameworks.
Artificial Intelligence and Law 24 (2016) 51–91
[5] Bench-Capon, T.: Representation of case law as an argumentation framework.
In Daskalopoulu, A., Bench-Capon, T., Winkels, R., eds.: Proceedings of JURIX
2002, IOS Press (2002) 103–112
[6] Mochales, R., Moens, M.F.: Study on the structure of argumentation in case
law. In: Proceedings of the 21st Annual Conference on Legal Knowledge and
Information Systems: JURIX 2008, IOS Press (2008) 11–20
[7] Dung, P.M.: On the acceptability of arguments and its fundamental role in non-
monotonic reasoning, logic programming and n-person games. Artificial intelli-
gence 77(2) (1995) 321–357
[8] Baroni, P., Giacomin, M.: Skepticism relations for comparing argumentation se-
mantics. International Journal of Approximate Reasoning 50(6) (2009) 854–866
[9] Cayrol, C., Lagasquie-Schiex, M.: On the acceptability of arguments in bipolar
argumentation frameworks. In: European Conference on Symbolic and Quanti-
tative Approaches to Reasoning with Uncertainty (ECSQARU). Volume 3571 of
Lecture Notes in Computer Science. (2005) 378–389
[10] Dunne, P.E., Hunter, A., McBurney, P., Parsons, S., Wooldridge, M.: Weighted
argument systems: Basic definitions, algorithms, and complexity results. Artificial
Intelligence 175(2) (2011) 457 – 486
[11] Bench-Capon, T.J.M., Doutre, S., Dunne, P.E.: Value-based argumentation
frameworks. In: Artificial Intelligence. (2002) 444–453
[12] Pazienza, A., Ferilli, S., Esposito, F.: Constructing and evaluating bipolar
weighted argumentation frameworks for online debating systems. In: 1st Work-
shop on Advances In Argumentation In Artificial Intelligence (AI3@AIIA2017).
Volume 2012 of Central Europe (CEUR) Workshop Proceedings. (2017) 111–125
[13] Pazienza, A., Ferilli, S., Esposito, F.: On the gradual acceptability of arguments in
bipolar weighted argumentation frameworks with degrees of trust. In: Foundations
of Intelligent Systems (ISMIS 2017). (2017) 195–204
44