A Multistage Extraction Pipeline for Long Scanned Financial Documents: An Empirical Study in Industrial KYC Workflows
This paper presents a multistage extraction pipeline that decouples page localization from multimodal reasoning to significantly improve structured information extraction accuracy in noisy, multilingual, and lengthy industrial KYC documents, outperforming direct end-to-end vision-language model baselines by up to 31.9 percentage points.
Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
Imagine you are a bank employee trying to read a massive, messy stack of old financial reports to check if a customer is trustworthy. These aren't neat, typed documents; they are scanned photocopies, some are crooked, some are blurry, and they are written in different languages. Worse, a single report might be 80 pages long, but the one number you need (like "Net Profit") is hidden on just page 42.
This paper describes a new, smarter way to handle this mountain of paperwork. Instead of trying to read the whole stack at once, the authors built a multi-stage assembly line that acts like a team of specialized workers.
Here is how their system works, using simple analogies:
1. The Problem: The "Blind Giant"
The authors tried using the newest, most powerful AI models (called Vision-Language Models or VLMs) to read these documents. They treated the AI like a giant who tries to swallow the entire 80-page book in one bite.
- The Result: The giant gets confused. It chokes on the noise, misses the tiny details, and often gives the wrong answer. It's also very slow and expensive to feed the giant that much food at once.
2. The Solution: The "Specialized Team"
Instead of one giant, the authors created a pipeline with four distinct steps, like a well-organized factory:
Step 1: The Janitor (Image Preprocessing)
Before anyone reads the document, a "Janitor" cleans it up. This step straightens crooked pages, removes coffee stains and shadows, and makes the text sharp. It's like ironing a wrinkled shirt before you try to read the label on it.Step 2: The Translator (Multilingual OCR)
The cleaned pages are fed to a translator (OCR) that turns the pictures of text into actual digital words. Since these documents are in English, Chinese, Indonesian, and more, this translator is fluent in many languages and can even read messy handwriting.Step 3: The Librarian (Page-Level Retrieval)
This is the most important step. Imagine you need to find a specific fact in a 100-page book. Instead of reading every single page, you hire a Librarian. The Librarian quickly scans the table of contents and the text to say, "Hey, the answer is on pages 12, 15, and 42. Ignore the other 95 pages."- Why this matters: The paper found this step is the "secret sauce." By filtering out the irrelevant pages, the system saves massive amounts of time and money, and it stops the AI from getting distracted by useless information.
Step 4: The Expert (Compact VLM Extraction)
Now, the system only sends those few relevant pages to the "Expert" AI. Because the Expert isn't overwhelmed by 80 pages of junk, it can focus perfectly on the specific numbers you need. The paper uses a "compact" (smaller, faster) AI for this job, which works surprisingly well when given clean, focused data.Step 5: The Human Check (Human-in-the-Loop)
Finally, a human analyst briefly checks the AI's work. The authors note that in banking, humans already have to check these documents for safety regulations. This system just makes the human's job easier by highlighting the relevant parts and flagging anything the AI was unsure about.
The Results: A Big Win
The team tested this on 120 real-world financial documents (about 3,000 pages total).
- Accuracy: Their new pipeline was 31.9% more accurate than just feeding the whole document directly to the AI. The best setup got the answer right about 87% of the time.
- Speed: Even though they added extra steps, the system wasn't slower. Because the "Librarian" filtered out so much junk, the "Expert" AI had less work to do, keeping the total time the same.
- The "Why": The study showed that for long, messy financial reports, finding the right page (the Librarian step) was the most critical factor. If you skip this, the system fails, no matter how smart the AI is.
In Summary
The paper argues that for long, messy financial documents, you shouldn't just throw a powerful AI at the whole problem. Instead, you should clean the data, translate it, filter out the noise, and then let a focused AI do the final extraction. It's the difference between asking a student to memorize an entire library versus giving them a specific book and a highlighter.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.