Classification of Business Activities by Machine Learning: The Case of France.

Nathan Randriamanana

30 April 2024

Context

  • Sirene is the French national company registry
  • When a company registers, an activity code is attributed
  • Challenges :
    • Refactoring of the Sirene information system
    • NACE revision to come in 2025
    • Teams still overwhelmed
  • End of March 2024:
    • Officially switch from Sirene 3 to Sirene 4
  • Consequences: Ideal moment to innovate (but under the constraint!)

The flow” of formalities

The administrative landscape

  • Siren number: company directory identification system
  • Principal activity code (APE)
  • Classification on a daily basis
  • Different administrations
  • Different information systems
  • Requirements: quick, responsive and flexible to updated instructions

Assign an activy code: different processes

Assign an activy code: two outcomes

  • Automatic coding
  • Human review

Near-ubiquity of ML

Automatic activity classification

Human review

Model

  • Text classification model which uses additional categorical variables
  • For now we use the fastText library
  • Originally trained on legacy data annotated partly by the coding engine and partly manually

FastText

  • a rapid and lightweight model

Calendar on new NACE adoption

  • 2025: statistical business register adopts new NACE
    • Dual Coding
      • major coding with NACE rev 2
      • minor coding with NACE rev 2.1
  • 2026: administrative business register adopts new NACE
    • Dual Coding
      • major coding with NACE rev 2.1
      • minor coding with NACE rev 2

1-to-1 correspondence

  • An easy and ideal case

1-to-many correspondence

  • An ultimately less desirable solution
  • Need expert decision based on activity description

Annotation campaign strategy

  • A one-shot operation spanning multiple months in 2024
  • manual labeling only on the 1-to-many codes
  • select data to annotate since the implementation of the one-stop shop
  • kill two birds with one stone
    • check NACE rev 2 coding quality on the 1-to-many codes
    • attribute a NACE rev 2.1 code

Labeling method

Reduce tasks to annotate

  • theoretical 1-to-many scenarios can eventually become 1-to-1 in practice

  • duplicated textual descriptions may occur due to shared practices among registrants.

  • How to reduce the annotation workload for our annotators ?

    • adapt correspondence table by considering real-world business rules
    • avoid giving the same textual descriptions to annotate