Christianvedel/OccCANINE_OCCICEM

OccCANINE: I-CeM Occupational Classification (OCCICEM)

Overview

OccCANINE_OCCICEM is a version of OccCANINE fine-tuned to automatically convert English occupational descriptions into I-CeM (Integrated Census Microdata) occupational codes. It uses a CANINE encoder with a sequential decoder trained using a mixed loss, fine-tuned from the OccCANINE_s2s_mix base model on IPUMS UK census data.

See more on: GitHub.com/christianvedels/OccCANINE

Read the paper on arXiv: https://arxiv.org/abs/2402.13604

Key Features

• English: Trained and evaluated on English occupational descriptions.
• Sequential decoding: Outputs I-CeM codes digit-by-digit.
• Mixed loss training: Combines sequence-level and flat classification losses.
• Fine-tuned: Initialized from OccCANINE_s2s_mix and fine-tuned on IPUMS UK I-CeM data.

Usage

from histocc import OccCANINE

model = OccCANINE(name="OccCANINE_OCCICEM", system="OCCICEM", hf=True)

result = model.predict("blacksmith", lang="en")

Contribution and Support

Developed at the University of Southern Denmark by Christian Møller Dahl, Torben Johansen and Christian Vedel.

---

Model Details:

• Task: Text Classification / Sequence Generation
• Base Model: CANINE (fine-tuned from OccCANINE_s2s_mix)
• Target system: I-CeM (OCCICEM)
• Language: English
• Framework: Transformers / PyTorch
• License: Apache 2.0
• Paper: arXiv 2402.13604