Sentence mining from games — the toolkit, and how the pieces fit together

Sentence mining from games — the toolkit, and how the pieces fit together

7 min read

If you learn a language by playing games, sooner or later you end up mining: saving the sentence you just read, with the word you didn’t know, so you can review it later. The community that does this seriously has built a toolkit for it, and most of it is open source. The trouble is that the toolkit is a set of parts rather than a product, and guides tend to describe one particular assembly.

This page is a parts catalog. It sorts the tools by the job each one does, gives the official facts for each (checked against their GitHub pages and sites on 2 September 2026), and ends with a choose-by-situation guide. There is no ranking, and no “best” — the right assembly depends on your game, your platform, and how much you want to set up.

The four jobs in a mining pipeline

  1. Get the text out of the game. Two mechanisms: a text hook pulls the string from inside the game process before it is drawn (exact, instant, only where a hook exists), or OCR reads the rendered screen (works anywhere, accuracy depends on the font).
  2. Look words up. A pop-up dictionary over the extracted text.
  3. Make the card. Sentence, word, definition, plus ideally the audio of the line and a screenshot of the moment.
  4. Review. A spaced-repetition system that schedules the cards.

Some tools do one job; a few span several. Playto’s own place is described at the end.

Getting text out: hooks

Textractor (Artikash, GPL-3.0, free). The classic text hooker: it injects a hook into the running game process, intercepts the text-output functions, and passes the text on to other tools. Many engines hook automatically; some need a manually supplied hook code. Windows 7 and later, plus Wine. The last release is v5.2.0 from January 2022, so it is stable rather than evolving. GitHub

Agent (0xDC00, free). A script-based text hooker built on Frida: instead of binary hook codes, it uses JavaScript hook scripts, and it reaches console emulators — Yuzu, Vita3k, PPSSPP, RPCS3 are named in the README — as well as Unity and PC engines. A game needs a script to exist, and the scripts repository has an auto-updater. GitHub

LunaTranslator (HIllya51, GPLv3, free). Primarily a hooker for visual novels, with emulator support (Switch, PSP, Vita, PS2 listed), a built-in OCR fallback, and a full set of translation engines including offline ones. It is a complete reader on its own; in a mining setup it is usually used as the text source. Windows 10/11 64-bit recommended; releases are frequent (v10.16.5.4, September 2026). GitHub

Getting text out: OCR

YomiNinja (matt-m-o, GPL-3.0, free). An OCR overlay: it reads the whole screen or a chosen window with one of several engines — PaddleOCR, Google Cloud Vision, Google Lens, Manga OCR, Apple Vision — and renders the recognized text as a transparent layer, so a pop-up dictionary like Yomitan can work on top of a game. English, Japanese, Chinese, Korean. Windows, Linux (X11 only), macOS. Last release v0.9.3, August 2025. GitHub

OwOCR (AuroraWright, GPL-3.0, free). A command-line OCR backend rather than an app: one interface over about fifteen engines, local (OneOCR, Windows OCR, Apple Live Text, Manga OCR, EasyOCR, RapidOCR and others) and cloud (Google Lens, Bing, Google Vision and others), with output to the clipboard, files or a websocket. Its focus is Japanese but it works for other languages. Needs Python and each engine installed separately. Last release 1.26.8, March 2026. Game Sentence Miner ships a fork of it. GitHub

Looking words up

Yomitan (yomidevs, GPL-3.0, free). The successor to Yomichan: a browser extension that shows a pop-up dictionary over any selectable text, with audio, kanji information, and one-click card creation through AnkiConnect. It supports dozens of languages. Because it reads browser text, it cannot look at a game window directly — that is why YomiNinja and Game Sentence Miner’s overlay exist, to put game text where Yomitan can reach it. Last release July 2026. GitHub

Making the card

Game Sentence Miner (bpwhelan, GPL-3.0, free). The orchestrator most current guides are built around. It takes text from an external source — Textractor, Agent, LunaTranslator, or its own built-in OCR — via clipboard or websocket, records the game through OBS’s replay buffer, uses voice-activity detection to trim the audio of the line, grabs a screenshot at that moment, and sends the finished card to Anki through AnkiConnect. It also has an in-game overlay with furigana for Yomitan lookups and a stats dashboard. Japanese, Chinese, Korean, Spanish are listed. Windows primarily; a Linux setup page exists. Requirements as documented: OBS with the replay buffer enabled (the installer bundles it), Anki running, and one text source. It is the most actively maintained tool here (v2026.8.2, August 2026). GitHub · Wiki

Reviewing

Anki (Ankitects, AGPL, free on desktop and Android; the iOS app is paid). The spaced-repetition program every open-source tool above targets. AnkiConnect is the add-on that exposes Anki’s local API on port 8765 so other programs can create cards; Anki has to be running for it to work. Its GitHub repository was archived in November 2025 and development moved to SourceHut. Anki · AnkiConnect

Migaku (commercial, subscription with a free trial). A browser extension plus its own review system: lookups, card creation, and spaced repetition under one account, without Anki. Eleven languages are listed. It works on web pages and subtitles (Netflix, YouTube), not on a game window, so for games it pairs with an OCR overlay or is used for the review half only. migaku.com

Where Playto fits

Playto covers the OCR, lookup, card, and review jobs in one desktop app, aimed at people who want the loop without assembling it. It reads the screen of any Windows game with local OCR and a local translation model, shows the meaning as an overlay while keeping the original in front of you, and saves the words and sentences you tap into a word book with flashcard review and spaced repetition inside the app. It can also export a ready-made Anki deck, screenshot included, if Anki is where you review.

What it does not do: it has no text hook, so on a game where a hook exists Textractor or LunaTranslator will give you exact text and Playto will give you OCR. It also does not record game audio into cards the way Game Sentence Miner does. Playto Lite is free with a limit on saved words; the full version removes the limit and adds cursor-follow reading for scattered UI text.

Choosing by situation

A visual novel with a working hook, and you already use Anki. Textractor or LunaTranslator as the text source, Game Sentence Miner to build cards with audio and screenshots, Yomitan for lookups on GSM’s overlay, Anki to review. This is the assembly the community guides describe, and it produces the richest cards.

A game with no hook — 3D, action, strategy, or an engine nobody wrote a hook for. The text has to come from OCR. Options: Game Sentence Miner’s built-in OCR (keeps the rest of the pipeline the same), YomiNinja if you mainly want lookups over the game, or Playto if you want the overlay and the review in one app with nothing to wire.

A console game in an emulator. Agent’s scripts cover several emulators, and LunaTranslator lists emulator support. OCR tools work on the emulator window like any other window.

You want the least setup. Playto (one install, no Anki, no OBS) or Migaku for the review side if you already pay for it. You give up hooks and recorded audio.

You do not use Anki and do not want to. Migaku or Playto, both of which review inside the app. The open-source chain assumes Anki.

macOS or Linux. YomiNinja and OwOCR run on both; Game Sentence Miner is Windows-first with a Linux page; Textractor is Windows and Wine; Playto is Windows only.

Before you start

  • Check the game’s display mode. OCR tools read most cleanly with the game in borderless or windowed mode.
  • Check what is still maintained. Textractor’s last release is from 2022; Game Sentence Miner, LunaTranslator, and OwOCR shipped releases in 2026.
  • Decide where you review first. Everything else in the pipeline is chosen to feed that.

Specifications change. The version numbers and dates above are what the official pages showed on 2 September 2026 — confirm on each project’s page before installing.

Share: X
← All posts