an adjustment function for dealing with similarities€¦ · introduction adjustment function...
TRANSCRIPT
![Page 1: An Adjustment Function For Dealing With Similarities€¦ · Introduction Adjustment Function Properties Related Work Conclusions and Perspectives Weighted Relations For instance:](https://reader033.vdocuments.net/reader033/viewer/2022060907/60a1aa3bf282421c8e13cf5b/html5/thumbnails/1.jpg)
An Adjustment Function For Dealing WithSimilarities
Leila Amgoud Victor David
IRIT, CNRS, Universite de Toulouse31062 Toulouse Cedex 9, France
September 2020
![Page 2: An Adjustment Function For Dealing With Similarities€¦ · Introduction Adjustment Function Properties Related Work Conclusions and Perspectives Weighted Relations For instance:](https://reader033.vdocuments.net/reader033/viewer/2022060907/60a1aa3bf282421c8e13cf5b/html5/thumbnails/2.jpg)
Introduction Adjustment Function Properties Related Work Conclusions and Perspectives
Outline
Introduction
Adjustment Function
Properties
Related Work
Conclusions and Perspectives
![Page 3: An Adjustment Function For Dealing With Similarities€¦ · Introduction Adjustment Function Properties Related Work Conclusions and Perspectives Weighted Relations For instance:](https://reader033.vdocuments.net/reader033/viewer/2022060907/60a1aa3bf282421c8e13cf5b/html5/thumbnails/3.jpg)
Introduction Adjustment Function Properties Related Work Conclusions and Perspectives
Introduction
Argumentation is a reasoning approach based on interactingarguments.[a] My salary is quite high then I can rent the apartment.[b1] If the bank approves the loan to buy a car, I will not haveenough money to rent the apartment.[b2] I may adopt a dog, however the owner forbids pets, thus I willnot rent the apartment.[b3] I may adopt a cat, however the owner forbids pets, thus I willnot rent the apartment.
a
b1 b2 b3
3/25
![Page 4: An Adjustment Function For Dealing With Similarities€¦ · Introduction Adjustment Function Properties Related Work Conclusions and Perspectives Weighted Relations For instance:](https://reader033.vdocuments.net/reader033/viewer/2022060907/60a1aa3bf282421c8e13cf5b/html5/thumbnails/4.jpg)
Introduction Adjustment Function Properties Related Work Conclusions and Perspectives
Weighted Arguments
For instance: certainty degree.[a] My salary is quite high then I can rent the apartment.[b1] If the bank approves the loan to buy a car, I will not haveenough money to rent the apartment.[b2] I may adopt a dog, however the owner forbids pets, thus I willnot rent the apartment.[b3] I may adopt a cat, however the owner forbids pets, thus I willnot rent the apartment.
a
1
b1
0.8
b2
0.6
b3
0.4
4/25
![Page 5: An Adjustment Function For Dealing With Similarities€¦ · Introduction Adjustment Function Properties Related Work Conclusions and Perspectives Weighted Relations For instance:](https://reader033.vdocuments.net/reader033/viewer/2022060907/60a1aa3bf282421c8e13cf5b/html5/thumbnails/5.jpg)
Introduction Adjustment Function Properties Related Work Conclusions and Perspectives
Weighted Relations
For instance: relevance degree.[a] My salary is quite high then I can rent the apartment.[b1] If the bank approves the loan to buy a car, I will not haveenough money to rent the apartment.[b2] I may adopt a dog, however the owner forbids pets, thus I willnot rent the apartment.[b3] I may adopt a cat, however the owner forbids pets, thus I willnot rent the apartment.
a
1
b1
0.8
b2
0.6
b3
0.4
0.9 0.3 0.3
5/25
![Page 6: An Adjustment Function For Dealing With Similarities€¦ · Introduction Adjustment Function Properties Related Work Conclusions and Perspectives Weighted Relations For instance:](https://reader033.vdocuments.net/reader033/viewer/2022060907/60a1aa3bf282421c8e13cf5b/html5/thumbnails/6.jpg)
Introduction Adjustment Function Properties Related Work Conclusions and Perspectives
Similarity between Arguments
[b1] If the bank approves the loan to buy a car, I will not haveenough money to rent the apartment.[b2] I may adopt a dog, however the owner forbids pets, thus I willnot rent the apartment.[b3] I may adopt a cat, however the owner forbids pets, thus I willnot rent the apartment..Similarity between arguments:
s(b1, b2) = s(b1, b3) = 0.2
s(b2, b3) = 0.9
6/25
![Page 7: An Adjustment Function For Dealing With Similarities€¦ · Introduction Adjustment Function Properties Related Work Conclusions and Perspectives Weighted Relations For instance:](https://reader033.vdocuments.net/reader033/viewer/2022060907/60a1aa3bf282421c8e13cf5b/html5/thumbnails/7.jpg)
Introduction Adjustment Function Properties Related Work Conclusions and Perspectives
Argumentation Framework
Definition (AF)
An argumentation framework (AF) is a tuple G = 〈A,w,R, σ, s〉,where
A ⊆f Argsa
w : A → [0, 1]
R ⊆ A×A (Attack relation)
σ : R → [0, 1]
s : A×A → [0, 1] (Similarity measure)
aArgs denotes the universe of all possible arguments
7/25
![Page 8: An Adjustment Function For Dealing With Similarities€¦ · Introduction Adjustment Function Properties Related Work Conclusions and Perspectives Weighted Relations For instance:](https://reader033.vdocuments.net/reader033/viewer/2022060907/60a1aa3bf282421c8e13cf5b/html5/thumbnails/8.jpg)
Introduction Adjustment Function Properties Related Work Conclusions and Perspectives
Evaluation of Argument Strength
Gradual Semantics
A degree (strength) of each argument
Evaluation Method = 〈f, g,h,n〉
characterizes
computes
8/25
![Page 9: An Adjustment Function For Dealing With Similarities€¦ · Introduction Adjustment Function Properties Related Work Conclusions and Perspectives Weighted Relations For instance:](https://reader033.vdocuments.net/reader033/viewer/2022060907/60a1aa3bf282421c8e13cf5b/html5/thumbnails/9.jpg)
Introduction Adjustment Function Properties Related Work Conclusions and Perspectives
Evaluation MethodM = 〈f, g,h,n〉
a
1
b1
0.8
b2
0.6
b3
0.4
0.9 0.3 0.3 s(b1, b2) = 0.2
s(b1, b3) = 0.2
s(b2, b3) = 0.9
1 Assess the strenght of every attack (bi , a),α1 = h(0.8, 0.9), α2 = h(0.6, 0.3), α3 = h(0.4, 0.3)
2 Adjust the strenght of every attack w.r.t. similarity,β = n((α1, b1), (α2, b2), (α3, b3)) = (β1, β2, β3)
3 Assess the strength of the group of attacks on a,δ = g(β1, β2, β3)
4 Evaluate the impact of attacks on the initial weight of a,λ = f(1, δ)
9/25
![Page 10: An Adjustment Function For Dealing With Similarities€¦ · Introduction Adjustment Function Properties Related Work Conclusions and Perspectives Weighted Relations For instance:](https://reader033.vdocuments.net/reader033/viewer/2022060907/60a1aa3bf282421c8e13cf5b/html5/thumbnails/10.jpg)
Introduction Adjustment Function Properties Related Work Conclusions and Perspectives
Evaluation MethodM = 〈f, g,h,n〉
Definition (EM)
An evaluation method (EM) is a tuple M = 〈f, g,h,n〉 such that:
f : [0, 1]× Range(g) a → [0, 1],
g :⋃+∞
k=0[0, 1]k → [0,+∞[,
h : [0, 1]× [0, 1]→ [0, 1],
n :⋃+∞
k=0 ([0, 1]× Args)k → [0, 1]k .
aRange(g) denotes the co-domain of g
10/25
![Page 11: An Adjustment Function For Dealing With Similarities€¦ · Introduction Adjustment Function Properties Related Work Conclusions and Perspectives Weighted Relations For instance:](https://reader033.vdocuments.net/reader033/viewer/2022060907/60a1aa3bf282421c8e13cf5b/html5/thumbnails/11.jpg)
Introduction Adjustment Function Properties Related Work Conclusions and Perspectives
Gradual Semantics
Definition (Gradual Semantics - S)
S based on an evaluation method M = 〈f, g,h,n〉 is a functionassigning to every AF G = 〈A,w,R, σ, s〉, a weightingDegSG : A → [0, 1] such that ∀a ∈ A,
DegSG(a) = f
(w(a), g
(n(
(h(DegSG(b1), σ(b1, a)), b1), · · · ,
(h(DegSG(bk), σ(bk , a)), bk))))
,
where {b1, · · · , bk} = Att(a)a.
aAtt(a) denotes the set of attackers of a
DegSG(a) represents the strength of a
11/25
![Page 12: An Adjustment Function For Dealing With Similarities€¦ · Introduction Adjustment Function Properties Related Work Conclusions and Perspectives Weighted Relations For instance:](https://reader033.vdocuments.net/reader033/viewer/2022060907/60a1aa3bf282421c8e13cf5b/html5/thumbnails/12.jpg)
Introduction Adjustment Function Properties Related Work Conclusions and Perspectives
Outline
Introduction
Adjustment Function
Properties
Related Work
Conclusions and Perspectives
12/25
![Page 13: An Adjustment Function For Dealing With Similarities€¦ · Introduction Adjustment Function Properties Related Work Conclusions and Perspectives Weighted Relations For instance:](https://reader033.vdocuments.net/reader033/viewer/2022060907/60a1aa3bf282421c8e13cf5b/html5/thumbnails/13.jpg)
Introduction Adjustment Function Properties Related Work Conclusions and Perspectives
Weighted h-Categorizer
Definition (Weighted h-Categorizer)
The function Swh transforms any AF G = 〈A,w,R, σ, s ≡ 0〉, intoa weighting DegSwhG : A → [0, 1] s.t. ∀a ∈ A,
DegSwhG (a) =
w(a) iff Att(a) = ∅
w(a)
1+∑
b∈Att(a)
DegSwhG (b)×σ(b,a)
else
Swh is based on M = 〈ffrac, gsum,hprod〉 such that:ffrac(x1, x2) = x1
1+x2
gsum(x1, · · · , xn) =n∑
i=1xi
hprod(x1, x2) = x1 × x2
13/25
![Page 14: An Adjustment Function For Dealing With Similarities€¦ · Introduction Adjustment Function Properties Related Work Conclusions and Perspectives Weighted Relations For instance:](https://reader033.vdocuments.net/reader033/viewer/2022060907/60a1aa3bf282421c8e13cf5b/html5/thumbnails/14.jpg)
Introduction Adjustment Function Properties Related Work Conclusions and Perspectives
Adjustment function nwh
Definition (nwh)
LetG = 〈A,w,R, σ, s〉 be an AF, x1, · · · , xk ∈ [0, 1], and b1, · · · ,bk ∈ A. We define nwh as follows:
nwh((x1, b1), · · · , (xk , bk)) = (DegSwhG′ (b1), · · · , DegSwhG′ (bk))
where G′ = 〈A′,w′,R′, σ′, s′〉, such that:
A′ = {b1, · · · , bk},w′(b1) = x1, · · · ,w′(bk) = xk ,
R′ = {(b1, b2), · · · , (b1, bk), · · · , (bk , b1), · · · , (bk , bk−1)},∀(bi , bj) ∈ R′, σ′(bi , bj) = s(bi , bj),
s′ ≡ 0.
14/25
![Page 15: An Adjustment Function For Dealing With Similarities€¦ · Introduction Adjustment Function Properties Related Work Conclusions and Perspectives Weighted Relations For instance:](https://reader033.vdocuments.net/reader033/viewer/2022060907/60a1aa3bf282421c8e13cf5b/html5/thumbnails/15.jpg)
Introduction Adjustment Function Properties Related Work Conclusions and Perspectives
Example of nwh
a
1
b1
0.8
b2
0.6
b3
0.4
0.9 0.3 0.3 s(b1, b2) = 0.2
s(b1, b3) = 0.2
s(b2, b3) = 0.9G:
Assume that we use hprod(x1, x2) = x1 × x2
b1
0.72
b2
0.18
b3
0.12
0.2
0.2 0.9
G′: nwh((0.72, b1), (0.18, b2), (0.12, b3))
= (DegSwhG′ (b1), DegSwhG′ (b2), DegSwhG′ (b3))
= (0.687, 0.147, 0.094)
15/25
![Page 16: An Adjustment Function For Dealing With Similarities€¦ · Introduction Adjustment Function Properties Related Work Conclusions and Perspectives Weighted Relations For instance:](https://reader033.vdocuments.net/reader033/viewer/2022060907/60a1aa3bf282421c8e13cf5b/html5/thumbnails/16.jpg)
Introduction Adjustment Function Properties Related Work Conclusions and Perspectives
Example of EMwh = 〈ffrac, gsum,hprod,nwh〉
a
1
b1
0.8
b2
0.6
b3
0.4
0.9 0.3 0.3 s(b1, b2) = 0.2
s(b1, b3) = 0.2
s(b2, b3) = 0.9
1 Strenght of every attack, α1 = hprod(0.8, 0.9) = 0.72,α2 = hprod(0.6, 0.3) = 0.18, α3 = hprod(0.4, 0.3) = 0.12
2 Strenght of every attack w.r.t. similarity, (β1, β2, β3) =nwh((α1, b1), (α2, b2), (α3, b3)) = (0.687, 0.147, 0.094)
3 Strength of the group of attacks on a,δ = gsum(β1, β2, β3) = 0.928
4 Strength of a, λ = ffrac(1, δ) = 0.51916/25
![Page 17: An Adjustment Function For Dealing With Similarities€¦ · Introduction Adjustment Function Properties Related Work Conclusions and Perspectives Weighted Relations For instance:](https://reader033.vdocuments.net/reader033/viewer/2022060907/60a1aa3bf282421c8e13cf5b/html5/thumbnails/17.jpg)
Introduction Adjustment Function Properties Related Work Conclusions and Perspectives
Outline
Introduction
Adjustment Function
Properties
Related Work
Conclusions and Perspectives
17/25
![Page 18: An Adjustment Function For Dealing With Similarities€¦ · Introduction Adjustment Function Properties Related Work Conclusions and Perspectives Weighted Relations For instance:](https://reader033.vdocuments.net/reader033/viewer/2022060907/60a1aa3bf282421c8e13cf5b/html5/thumbnails/18.jpg)
Introduction Adjustment Function Properties Related Work Conclusions and Perspectives
Some Properties on nwh
nwh can be used by a gradual semantics
Theorem
There exists a unique semantics that is based on the evaluationmethod 〈ffrac, gsum,hprod,nwh〉.
18/25
![Page 19: An Adjustment Function For Dealing With Similarities€¦ · Introduction Adjustment Function Properties Related Work Conclusions and Perspectives Weighted Relations For instance:](https://reader033.vdocuments.net/reader033/viewer/2022060907/60a1aa3bf282421c8e13cf5b/html5/thumbnails/19.jpg)
Introduction Adjustment Function Properties Related Work Conclusions and Perspectives
Some Properties of nwh
When all the arguments are dissimilar, the adjustment functiondoesn’t alter their initial values
Property (P1)
For any AF G = 〈A,w,R, σ, s〉, for all a1, · · · , an ∈ A, for allx1, · · · , xn ∈ [0, 1], if ∀i , j ∈ {1, · · · , n}, i 6= j , s(ai , aj) = 0, then
nwh((x1, a1), · · · , (xn, an)) = (x1, · · · , xn).
nwh can only reduce the value of an argument
Property (P2)
For any AF G = 〈A,w,R, σ, s〉, for all a1, · · · , an ∈ A, for allx1, · · · , xn ∈ [0, 1], if nwh((x1, a1), · · · , (xn, an)) = (x ′1, · · · , x ′n),then ∀i ∈ {1, · · · , n}, x ′i ≤ xi .
19/25
![Page 20: An Adjustment Function For Dealing With Similarities€¦ · Introduction Adjustment Function Properties Related Work Conclusions and Perspectives Weighted Relations For instance:](https://reader033.vdocuments.net/reader033/viewer/2022060907/60a1aa3bf282421c8e13cf5b/html5/thumbnails/20.jpg)
Introduction Adjustment Function Properties Related Work Conclusions and Perspectives
Some Properties of nwh
If an argument is dissimilar to all other arguments and its initialvalue is 0, then it will not have any impact on the readjustedvalues of the other arguments
Property (P3)
For any AF G = 〈A,w,R, σ, s〉, for all a1, · · · , an, b ∈ A, for allx1, · · · , xn, y ∈ [0, 1], if
∀i ∈ {1, · · · , n}, s(ai , b) = 0,
y = 0,
thennwh((x1, a1), · · · , (xn, an), (y , b)) = (nwh((x1, a1), · · · , (xn, an)), 0).
20/25
![Page 21: An Adjustment Function For Dealing With Similarities€¦ · Introduction Adjustment Function Properties Related Work Conclusions and Perspectives Weighted Relations For instance:](https://reader033.vdocuments.net/reader033/viewer/2022060907/60a1aa3bf282421c8e13cf5b/html5/thumbnails/21.jpg)
Introduction Adjustment Function Properties Related Work Conclusions and Perspectives
Outline
Introduction
Adjustment Function
Properties
Related Work
Conclusions and Perspectives
21/25
![Page 22: An Adjustment Function For Dealing With Similarities€¦ · Introduction Adjustment Function Properties Related Work Conclusions and Perspectives Weighted Relations For instance:](https://reader033.vdocuments.net/reader033/viewer/2022060907/60a1aa3bf282421c8e13cf5b/html5/thumbnails/22.jpg)
Introduction Adjustment Function Properties Related Work Conclusions and Perspectives
Readjusted Score - [Amgoud and al., KR 2018]
avg denotes the average operator
Definition ( Readjusted Score nrs )
Let a1, · · · , ak ∈ Args and x1, · · · , xk ∈ [0, 1].nrs((x1, a1), · · · , (xk , ak)) =(
avgxi∈{x1,··· ,xk}\{x1}
(avg(x1, xi )× (2− s(a1, ai ))
2
), · · · ,
avgxi∈{x1,··· ,xk}\{xk}
(avg(xk , xi )× (2− s(ak , ai ))
2
)).
nrs() = () and nrs((x1, a1)) = (x1) if k = 1.
22/25
![Page 23: An Adjustment Function For Dealing With Similarities€¦ · Introduction Adjustment Function Properties Related Work Conclusions and Perspectives Weighted Relations For instance:](https://reader033.vdocuments.net/reader033/viewer/2022060907/60a1aa3bf282421c8e13cf5b/html5/thumbnails/23.jpg)
Introduction Adjustment Function Properties Related Work Conclusions and Perspectives
Properties on nrs
Property
1 nrs violates P1, P2 and P3
2 Let gsum(x1, · · · , xn) =n∑
i=1xi , ∃a1, · · · , an, b ∈ Args and
x1, · · · , xn, y ∈ [0, 1] such that:
∀i ∈ {1, · · · , n}, s(ai , b) = 0,y = 0,gsum(nrs((x1, a1), · · · , (xn, an))) <gsum(nrs((x1, a1), · · · , (xn, an), (y , b)))
23/25
![Page 24: An Adjustment Function For Dealing With Similarities€¦ · Introduction Adjustment Function Properties Related Work Conclusions and Perspectives Weighted Relations For instance:](https://reader033.vdocuments.net/reader033/viewer/2022060907/60a1aa3bf282421c8e13cf5b/html5/thumbnails/24.jpg)
Introduction Adjustment Function Properties Related Work Conclusions and Perspectives
Conclusions and Perspectives
Conclusions:
Extending the notion of evaluation method by anadjustment function
Proposing a novel adjustment function nwh
Investigating the properties of the existing functions
————————
Perspectives:
Axiomatisation of adjustment functions
Investigating other adjustment functions
24/25
![Page 25: An Adjustment Function For Dealing With Similarities€¦ · Introduction Adjustment Function Properties Related Work Conclusions and Perspectives Weighted Relations For instance:](https://reader033.vdocuments.net/reader033/viewer/2022060907/60a1aa3bf282421c8e13cf5b/html5/thumbnails/25.jpg)
Introduction Adjustment Function Properties Related Work Conclusions and Perspectives
Reference
[Amgoud and al., KR 2018]: Amgoud L, Bonzon E, Delobelle J,Doder D, Konieczny S, Maudet N. Gradual Semantics Accountingfor Similarity between Arguments. In: KR; 2018. p. 88–97.
25/25