project origin

context and challenges

illustration Gestion des informations
Information management

spreadsheets, relational databases

illustration Gestion des informations
Information extraction

turning text into structured data

illustration Gestion des informations
Desire for automation

AI boom raised organizational awareness

images chip and robo-hand

outstanding needs

  • Lack of internal AI expertise
    • Difficulty in evaluating or negotiating services
  • Solutions not tailored to specific needs
    • Uncertain cost-benefit analysis
  • Lack of trust in LLMs
    • Hallucinations, opacity

interface between two canonical approaches

Top-down Symbol Top-down

symbolic rules

explicit specification of target regularities

Plus Icon
  • High precision
  • Low recall
  • Explainable
Plus Icon
  • Rigid
  • Requires few
    examples

Top-down Symbol Bottom-up

deep learning, LLMs

implicitly captures regularities with numerous annotated examples

Plus Icon
  • High recall
  • Enhanced handling
    of context
Plus Icon
  • Black box
  • Costly training
  • Limited ability to
    explore data

hyperquest

best of both worlds

background image - hexagone bottom up

HyperQuest
reveals key patterns

icon - bottom up

Allows domain experts without AI expertise to select the most relevant cases

icon bottom up icon top down

Suggests/co-constructs targeted information extraction rules with few examples

icon bottom up

Expands assisted extraction through neuro-symbolic patterns

successive refinement of sub-corpora and increasingly relevant patterns

proposed solution

application ready to be used by non-specialists

icon Hypergraphes semantique Semantic Hypergraphs

generalize semantic graphs, great flexibility in defining regularities to extract

icon App Ergonomic / No-code App

democratizes interactive extraction of relevant structured information

App Screenshot