The people and research behind extract-text

Publications

The work it comes from

The Cityscape paper

Many states have no process to count, digitize, or release information on evictions, and the filings sit in court records as page images. The paper describes how the Eviction Research Network read those images. It downloaded the court records, turned the pages into text, found and mapped the addresses, and estimated demographics from names and locations. For Washington State it processed millions of pages in more than 110,000 eviction filings from 2004 to 2017. The paper closes with a strategy for collecting eviction data across the country.

extract-text replaces the software that paper used. The Network scored both against the same hand-labeled Washington addresses, on a scale where 1 means the address read is identical to the label. The paper's software had a mean similarity of 0.82. The deterministic engine had 0.87 to 0.92 in the July 2026 runs.

Cheng Ren's dissertation

The dissertation asks what happens after an eviction is filed. It used natural language processing and computer vision to read 772,629 PDF files from 56,070 Pierce County cases, 2004 to 2022. It then followed each case to a dismissal or a judgment and, after a judgment, to an eviction by the sheriff.

extract-text takes three things from the dissertation. The rules that read a whole case file and decide whether a case was dismissed, went to judgment, or ended with the sheriff follow the dissertation's. The form templates are a rule-based version of its method, which learned the layout of each court form from 10 to 20 labeled pages. And Cheng Ren first tested the AI engine's model, on 300 Pierce County summonses.

The card game's design

Address Check, the card game that produces hand-checked address labels, takes two ideas from braindr.

A player makes one judgment on one image with a single tap. And each player's answers are weighted by how often that player is right on cards whose answer is already known.

Terms of use

What the project supports, and what it does not

The code is open source under the AGPL-3.0-or-later license. Commercial use is permitted. Building a closed-source product or a hosted service on this code is not.

The project is for research, litigation, journalism, policy analysis, and direct legal services that describe or reduce displacement. It is also for tenant defense, court administration and self-help, and public accountability for how eviction is practiced.

The project is against treating a filing as a fact about a person. An eviction filing is not a finding of wrongdoing. We do not want this code, or anything derived from it, used to

• produce tenant screening reports or any product that scores or ranks rental applicants,
• build or supply landlord blacklists,
• target people for debt collection,
• locate people for immigration enforcement or any other enforcement action,
• resell extracted records or feed them to data brokers, or
• train models whose purpose is any of the above.

The full statement is ETHICS.md. The license covers the software, not the records you read with it. DATA-USE.md sets the terms for any dataset the project supplies.

If you are unsure whether a use fits, ask before you build. Write to evictions@berkeley.edu.