Information Extraction

Information extraction is the job of reading through a thick document from start to finish and pulling out only the items you need, organized into a table. Picture an intern flipping through a hundred contracts, highlighting company names, contract terms, and penalty clauses, then typing them into a spreadsheet. Information extraction is what happens when a computer takes over that repetitive work.

The real question is how flexibly the system can find things. An approach that quickly locates only predefined fields is cheap, but it breaks down the moment a new kind of field shows up. On the other hand, using a large language model that can answer almost anything is flexible, but the cost adds up with every document you process. There's another trap too: length. Short items like names or dates are usually found easily, but long stretches of text—like a forty-word liability waiver clause—are easy to miss entirely.

To ease this dilemma, recent approaches try scanning the whole document once while only marking where each item starts and ends. Instead of listing out every candidate one by one, marking just the boundaries cuts down on computation and removes the need to cap the length of what can be extracted.

Articles describe it this way: "pulling information such as names, dates, or clauses out of a document is called information extraction." A common point of confusion is that information extraction is not the same as summarizing a document or answering questions about it. Information extraction pulls specific items out in their exact form (names, dates, clauses, etc.) and turns them into structured data—it doesn't explain or interpret the meaning of the document.

Try asking a chatbot the following to get a feel for how information extraction actually works.

Find the person's name, the date, and the amount of money in the sentence below, and organize them into a table.

"CEO Kim Min-su agreed to pay a deposit of 50 million won on March 15, 2024."

After you get the result, try rephrasing the sentence to make it longer and more complex, then ask again. You'll see for yourself that short items are found easily, while long, drawn-out clauses tend to get missed.