AIASHI Logo
HEC Logo

Extracting Insights from Large Corpora for
Enhanced Decision Support

Project thesis presented by

Valentin ANGENOT

To obtain the degree of

MASTER IN BUSINESS ENGINEERING

with a specialization in

Science and Technology

Liège, 5 September 2025

Val Logo
AIASHI — UNIVERSITY OF LIÈGE
Slide 2 · Project Overview
AIASHI is a project of the University of Liège
Its aim is to better inform the allocation of research resources.
€0
invested in research last year
·
·
AIASHI — UNIVERSITY OF LIÈGE
Slide 3 · ORBi Goldmine
AIASHI focuses on ORBi's scientific articles
This corpus is a goldmine for the University of Liège
62,000 abstracts
Trends Analysis
Emerging & Declining Topics
Better Information
Better Allocation
Yet all this potential remains untapped.
·
·
AIASHI — UNIVERSITY OF LIÈGE
Slide 4 · Literature Review
Literature Review
LDA
BERTopic
LLM
AIASHI — UNIVERSITY OF LIÈGE
Slide 5 · Research Questions
AIASHI ⇔ RQs

Enhancing Decision Support

Reducing Insights Loss

Evaluating Reliability

Enhancing Navigability

AIASHI — UNIVERSITY OF LIÈGE
Slide 6 · Methodology
Methodology
LDA
BERTopic
LLM
·
·
AIASHI — UNIVERSITY OF LIÈGE
Slide 7 · Corpus Description
Corpus Description
62,077 publications
1839–2025
Temporal Distribution
< 1979 2.6%
1980–2000 12.2%
2000s 22.6%
2010s 38.9%
2020–2025 23.7%
Domains
Top disciplines
Biological & Health Sci. 51.0%
Earth & Planetary Sci. 16.0%
Engineering & Tech 9.8%
Natural Sci. 8.45%
Under-represented
Economics & Business 1.9%
Law · Pol. Sci. 2.3%
Humanities 4.3%
Social Sci. 6.3%


Languages
English 79.2%
French 19.4%
Other 1.4%

Limitation

Delays in ORBi registration (up to ~2 years)

AIASHI — UNIVERSITY OF LIÈGE
Slide 8 · Evaluation Framework
Evaluation Framework

Our Approach

1
Development review
2
Expert validation
·
·
AIASHI — UNIVERSITY OF LIÈGE
Slide 9 · Results
Results
LDA
BERTopic
LLM
Limitation

Many incoherent topics

Strength

Coherence Enhancement

Challenge

Outliers

30.2% unassigned documents

Limitation

LLM-Only non-scalable

Unsuitable for ORBi Corpus

Solution

UniversalTopic Development

A new hybrid pipeline reducing outliers by 75%

·
·
AIASHI — UNIVERSITY OF LIÈGE
Slide 10 · Universal Topic
Universal Topic
1
BERTopic
2
LLM-Labeling
3
LLM-Definition
4
LLM-Outliers Reclassification
5
LLM-Clustering · Lighthouse Method
6
LLM-Taxonomy
7
Metadata Enrichment
Before (raw)
Before
→
After (enriched)
After
·
·
AIASHI — UNIVERSITY OF LIÈGE
Slide 11 · Outcome
OUTCOME
Scope expansion — from “Governance Tool” to Institutional Asset
Governance Tool
Initial scope
Institutional Asset
External Valorisation

RISE

Partnerships • Experts • Visibility

Internal Collaboration

Interdisciplinarity

Researcher matching • Knowledge discovery • Cross-faculty projects

Research Governance

Decision Support

Resource allocation • Strategic planning

·
·
AIASHI — UNIVERSITY OF LIÈGE
Slide 12 · Future Work — Research Gaps
Future Work — Research Gaps Benchmarking
Merge Corpora Processing Gap Detection
AIASHI — UNIVERSITY OF LIÈGE
Slide 13 · Recommendations

Recommendations — 3 Pillars

Pillar Accuracy & Relevance over Time

Corpus Updates

Via outlier reclassification methods

Reducing Submission Delays

Reminders J-30 / J-15 &
Impact measurement vs baseline

Metadata

ORBi taxonomy &
Tailored coherence metric

LLM Upgrade

Mistral-Large

Pillar Institutional Integration

3 Disctinct Tools

Governance dashboard · External visibility platform · Matchmaking platform

Co-development

Need assessment · Pilots · Iterative feedback

Gaps Partnerships

Start early · Concentric approach for quick wins

Vision

ULiège-only vs Commercialisation

Pillar Risk Management

Rectorate Change 2026

Launch in 2025–26

Fragmentation of Effort & Scope Creep

Dual structure · Dedicated budget lines

Legitimacy of the tool

Establishing an ethical charter

Legal & Data

Legal department engaged from day-1 for partnerships

Scalability

Shown not to be a concern

·


·
AIASHI — UNIVERSITY OF LIÈGE
Thank You
Thank You
Questions & Discussion
Valentin ANGENOT

Master in Business Engineering · Science & Technology
End of Presentation
Defense Complete