ww
A Con ex -Sensi i e Reasoning F amewo k o
Knowledge G aphs
Sa a Todo o ikj & Flo ian Hahn
T
Mo i a ion
• Real-wo ld knowledge is con adic o y and con ex -dependen
• T adi ional Seman ic Web sys ems ely on mono onic logic → limi s abili y o adap o con ex ual nuances
• We p opose a h ee-laye ed easoning amewo k o KGs ha suppo s con ex -sensi i e easoning
Ou look
• Ou p oposed amewo k easons o e claims, esol es con adic ions and adap s conclusions based on con ex .
• Allows o nonmono onic easoning and con ex ual il e ing.
• Modula easoning laye - beyond deduc i e o de easible easoning.
• Aligns wi h cogni i e science indings showing how humans ank and e ise belie s.
• Uni ica ion o seman ic ep esen a ion wi h “old-school-AI” easoning and con ex -awa eness.
• Applicable beyond schola ly domains, a ounda ion owa ds us wo hy, adap i e and explainable knowledge sys ems.
P o esso ship Da a Managemen
Uni e si y o Technology Chemni z
On ology Laye
• S uc u ed ep esen a ion o claims and e idence
• Cap u es con adic ions wi hou o cing consis -
ency
• Rela ions like con adic s, suppo s o e u es.
Reasoning Laye
• Suppo s di e en easoning ypes (deduc i e,
de easible, abduc i e, e c.)
• Encodes ule p io i ies and excep ions
• Suppo s e ision o en a i e in e ences
Con ex ual Adap a ion Laye
• Me ada a-awa e p io i y modi ica ion
• Con ex dynamically eshapes ule applica ion
• Mul iple con ex s suppo ed o e he same KG
Wha is s o ed?
S udies on Policy X:
• A (Eu ope, 2018, obse a ional) → Pos-
i i e E ec
• B (Asia, 2021, RCTs) → No E ec
• C (Ame ica, 2019, RCTs) → Nega i e E -
ec
Con lic ing e idence s o ed as-is.
Wha is in e ed by de aul ?
Apply ules:
• P e e RCTs
• P e e ecen s udies
De aul ou come:
S udy B ou anks A and C
→ Policy X: no e ec
P oduces he con ex - ee conclusion.
How does he conclusion change wi h
con ex ?
Con ex : E alua ing Policy X o Eu ope
• Inc ease p io i y o local e idence
• Downweigh o eign RCTs
Re ised p io i ies:
Local > RCT > Recency
Con ex ual ou come:
S udy A now ou anks B
→ Policy X: posi i e e ec in Eu ope
Same knowledge base, di e en con-
clusion because he con ex shi s he
ule p io i ies.
How do we explain con lic ing esul s?
Gene a e hypo heses consis en wi h
me ada a:
• Regional implemen a ion di e ences
• Popula ion di e ences
• Va ia ions in measu emen o design
Abduc i e easoning: P o ides expla-
na ion and hypo heses on con ex ual
me ada a
[1] S. Al Mani , J. Nies oy, M. A. Le inson, and T. Cla k (2021). E idence G aphs: Suppo ing T anspa en and FAIR Compu a ion, wi h De easible Reasoning on Da a, Me hods, and
Resul s. IPAW 2020-2021.
[2] I. Asi , I. Tiddi, and A. J. G. G ay (2021). Using Nanopublica ions o De ec and Explain Con adic o y Resea ch Claims. eScience 2021.
[3] T. Cla k, P., and C. Goble (2014). Mic opublica ions: a Seman ic Model o claims, e idence, A gumen s and Anno a ions in Biomedical Communica ions. Jou nal o Biomedical
Seman ics, ol. 5(28).
[4] N. Fanizzi and C. D’ama o (2025). The Blessing o Dimensionali y Pe spec i es o Reasoning and Lea ning on Hype dimensional compu ing/ ec o Symbolic A chi ec u es. Neu-
osymbolic A i icial In elligence, ol. 1.
[5] A. Hogan e al. (2022). Knowledge G aphs. ACM Compu ing Su eys, ol. 54(4).
[6] S. Todo o ikj, G. Ke n-Isbe ne , and M. Ragni (2021). On he Cogni i e Adequacy o Non-mono onic Logics. NMR 2021 @ KR 2021.
✉ sa a. odo o ikj@in o ma ik. u-chemni z.de
