BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//pretalx//pretalx.com//compute-paris-2026//talk//YR9B3N
BEGIN:VTIMEZONE
TZID:Europe/Paris
BEGIN:STANDARD
DTSTART:20251126T000000
TZNAME:CET
TZOFFSETFROM:+0100
TZOFFSETTO:+0100
END:STANDARD
BEGIN:DAYLIGHT
DTSTART:20260329T030000
RDATE:20270328T030000
TZNAME:CEST
TZOFFSETFROM:+0100
TZOFFSETTO:+0200
END:DAYLIGHT
BEGIN:STANDARD
DTSTART:20261025T030000
RDATE:20271031T030000
TZNAME:CET
TZOFFSETFROM:+0200
TZOFFSETTO:+0100
END:STANDARD
END:VTIMEZONE
BEGIN:VEVENT
SUMMARY:Measuring Knowledge in AI Systems:   Automated domain specific Kno
 wledge Graph Construction from Scientific Literature - Feryal Batoul TALBI
DTSTART;TZID=Europe/Paris:20261126T120000
DTEND;TZID=Europe/Paris:20261126T123000
DTSTAMP:20260930T112449Z
UID:pretalx-compute-paris-2026-YR9B3N@pretalx.com
DESCRIPTION:What is knowledge\, and how do we measure it? Extracted knowle
 dge is more than a true statement\; it carries justification\, provenance
 \, and a context of use. When an expert builds a knowledge graph from scie
 ntific literature\, every node and edge reflects a justified interpretatio
 n: they know why a relation holds\, which paper supports it\, and under wh
 at conditions it applies. This raises a fundamental challenge for AI-based
  information retrieval systems. Retrieval-augmented generation (RAG) can n
 ow produce knowledge graphs that resemble those assembled by hand\, yet th
 ese graphs are generated through statistical pattern recognition rather th
 an contextual understanding of the scientific processes they describe. The
  central question of this work is therefore: how can we measure whether su
 ch systems are epistemic agents or only convincing imitators of one? \n\nT
 o explore this problem\, we developed OntoGeoRAG\, an open-source framewor
 k that combines document retrieval\, LLM-based information extraction\, an
 d entity–relation normalization to transform scientific PDFs into proven
 ance-aware semantic knowledge graphs. Using mass-transport deposits in sub
 surface geophysics as a challenging test case\, the framework achieves 77%
  recovery against an expert-curated benchmark. Rather than treating the re
 maining 23% as simple error\, we interpret these failures as evidence of t
 he current epistemic limitations of retrieval-based AI systems. Although d
 emonstrated in geophysics\, the methodology is domain-agnostic and transfe
 rable to any scientific field with structured terminology and complex conc
 eptual relationships.
LOCATION:Auditorium
URL:https://pretalx.com/compute-paris-2026/talk/YR9B3N/
END:VEVENT
END:VCALENDAR