Extracting Structured Data from Legal Documents
summary
Uses a conditional random field with hand engineered features (spaCy POS tags, baby name lists, number and month detectors) on OCRed PDFs via Google Cloud OCR and page bounding boxes, deployed per document type (about six common types) with a hierarchical classifier that hits ~93% top1 doc type and a model accuracy around two thirds overall, raising throughput from 30 to 150 documents per hour; common failures come from OCR errors and unseen names, and humans remain in the loop.
AI-written summary · about
chapters (9)
- 0:00Atrium and legal automation
- 1:51Paralegal document intake
- 3:56Machine learning speeds processing
- 4:39Disturbia and hidden Markov models
- 17:45Calculating probabilities
- 19:10Weather to legal text with CRFs
- 23:24Labeling legal documents
- 25:54CRF model strengths and weaknesses
- 26:58Q&A
related talks
01h 00mComplex Data Ingestion with Open Source AI
29mConquering PDFs- document understanding beyond plain text
01h 19mLearn to Unlock Document Intelligence with Open-Source AI
29mEY Aleph: deep learning applied to jurimetrics practice
26mAdvanced Document Parsing for AI Workflows
01h 25mLearn to Unlock Document Intelligence with Open-Source AI
42mUsing Deep Learning to Understand Documents
15mFrom Image to Graph: Leveraging Tesseract OCR Engine for Document Chunking and GraphRAG