Design Considerations in Low-Resource Archival OCR: A Case Study in Lakota
We investigate how small ($\leq$3B) vision-language models (VLMs) can be adapted as optical character recognition (OCR) tools for Lakota, a North American Indigenous language.
ProofPaper ↗
Key points
- Building and expanding access to digital archives supports language reclamation and creates lasting resources for community revitalization.
- Designing effective tools to support these efforts requires taking into account the community-, language-, and document-specific context.
- In addition to benchmarking modeling approaches, we inspect four practical facets of low-resource archival OCR that arise in previous work and conversations with community members: constrained training budgets, segmentation noise, scan degradation, and orthographic variance.
- We demonstrate that fine-tuning VLMs for OCR post-correction on 35 pages of transcribed data decreases character error rate (CER) from 17.1 to 3.9% on a bilingual corpus, and find that training only with nine pages increases CER by only 1.1% over this.
Sources (1)
- [1]Design Considerations in Low-Resource Archival OCR: A Case Study in LakotaarXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 18, 06:02 PM
We investigate how small ($\leq$3B) vision-language models (VLMs) can be adapted as optical character recognition (OCR) tools for Lakota, a North American Indigenous language.
Building and expanding access to digital archives supports language reclamation and creates lasting resources for community revitalization.
Extractive summary: sentences quoted from the sources.
Before this
- Oct 18, 2026ViCoR: Estimating Visual Necessity via Counterfactual Residuals for Multimodal Medical Data Selection
- Oct 18, 2026PACT: NVFP4 Data-Flow LoRA Fine-Tuning on Consumer GPUs
- Oct 17, 2026Defending Against Dormant Poisoning Attacks Across Language and Multimodal Agents
- Oct 16, 2026Behavioral Couplings: From Formation to Behavioral Emergence in Language Models
- Oct 16, 2026Minimax-optimal Scaling Laws in Ridge Regression via Transfer Learning
- Oct 16, 2026Compression Precedes, Content Decides: Emergent Misalignment in Reward Models