← All projects

Copyright metadata extraction

A pipeline that reads scanned Korean copyright contracts and consent forms and extracts the metadata needed to register the work. Our Soongsil University team built it for two public agencies: 한국문화정보원 (Korea Culture Information Service Agency) and 한국저작권위원회 (Korea Copyright Commission).

Where
Soongsil University, Seoul · April 2025 – December 2026
Project
공유저작물의 글로벌 확산을 위한 콘텐츠 분석 및 유형정보 판단 기술 개발, a 2025 Ministry of Culture, Sports and Tourism R&D project. Soongsil is one of three organizations, with two companies.
Team
About 7 people at Soongsil, including 3 professors
My role
Built the LLM pipeline and integrated it into the service, built the multimodal pipeline for scanned pages, images and video, and integrated the OCR models
Stack
Qwen3.5-122B-A10B (extraction and arbiter), Qwen3.7-Plus and Qwen3.8-Max (OCR; Mistral and Naver OCR as fallbacks), Qwen3.8-Omni-Flash and Gemma 4 (images and video), KLUE-RoBERTa-Large NER, PyMuPDF, ffmpeg, FastAPI, Caddy · lab server

Demo and evidence

Try it

Pick a contract or a work, and the pipeline runs: each stage with its real timing, the record it returns and, for contracts, how the arbiter settled each field where the LLM and the NER model were compared. I wrote the three contracts for this page, and everything in them is made up; contract C is a scan with no text layer, so OCR has to read it first. The image and the video are public-domain works: a painting by Kim Hong-do and NASA's Apollo 11 launch film.

This replays runs recorded on 17 and 18 November 2026: each file was sent once to the pipeline's API (/api/llm-extract), and the page plays back the answer. Every run calls paid model APIs, so the page does not call the service itself. The NER model makes more mistakes on these contracts than in our tests, because I used my own layout rather than the ones it was trained on. Catching those mistakes is the arbiter's job.

Evidence

The examiner console scoring the files from the demo above, at 2× speed: Te-1 on the three made-up contracts (macro-F1 93.3%), then Te-2 on the painting and the Apollo film (83.3% on the six fields that have an answer). Recorded on 20 November 2026; a local file path is blurred.

Consider updating the public repo, which still shows the first version.

The problem

The system

Copyright metadata pipeline: an upload goes through a router by file type. PDFs: text layer or OCR (Qwen3.7-Plus, Qwen3.8-Max, Mistral, Naver), then LLM extraction with Qwen3.5-122B and a KLUE-RoBERTa-Large NER model in parallel, an LLM arbiter and contract rules. Images: a vision LLM chain (Qwen3.8-Omni-Flash, Qwen3.7-Plus, Gemma 4). Videos: keyframes, then a vision LLM chain. Format, size, resolution and hash are measured from every file. Output: one JSON response with 87 unified fields, the TTA standard form and the 22 scored fields. Runs on the lab server behind Caddy with HTTPS and API keys.
The system as it runs today. For contracts, two extractors read the text and an LLM arbiter compares them field by field; fixed rules then read the title, the ticked work type, the rights boxes and the parties. Works go to a vision model, and a work can inherit its rights fields from its contract. Our EEECS 2026 paper describes the first version of the arbiter. Click to enlarge.

What I did

Results

The project measures this part by metadata accuracy: at least 80% in the first year and 85% in the second, on 1,000 test items for each type of work (expired, donated, Creative Commons, Korea Open Government License). An accredited test lab checks the second-year result. Before that test, we rehearse it on 4,000 sets against our own answer key.

100.0%macro-F1 on contract fields (Te-1, 15 fields)4,000 generated contracts, 14 Nov 2026, no failures
93.6%macro-F1 on work attributes (Te-2, 22 fields)4,000 works (2,000 images, 2,000 videos, 500 per licence type), 15–16 Nov 2026; 95% CI 93.6–93.7; target 85%
76.8%F1 on generated descriptions, the hardest fieldSame run: hashtags 47.9%, keywords 35.5%; every other work field is 100%
33 smedian time per contractImages 22 s, videos 27 s; same runs, with 32–50 files in parallel

These are internal rehearsals on test contracts we generated, which all have a text layer, so OCR was not exercised. In Te-2 the 15 contract fields come from the matching Te-1 contract. The official test by the accredited lab is still to come.

Examiner console, results tab of the Te-2 rehearsal: macro precision 93.6%, recall 93.7%, macro-F1 93.6% against an 85% target on 4,000 sets, and a per-field table of precision, recall and F1 for the 22 fields.
The examiner console after the Te-2 rehearsal on 4,000 works. The per-set table below it, which names the test works, is left out. Click to enlarge.
← Thai Fruit Garden agent All projects →