Word, Character prediction in any language in the browser (ppmpredictor) (willwade.github.io)

ppmpredictor is a new npm package and browser/Node.js demo that brings the classic Prediction by Partial Matching (PPM) character‑based language model into easy web use. Built on Google Research’s ppm_language_model.js from the JSLM project, the library extends the original PPM algorithm (Cleary & Witten, 1984) with word‑level prediction, error‑tolerance features, and integration with WorldAlphabets — a frequency list and keyboard layout collection for 100+ languages. The result is a lightweight, privacy‑friendly predictor that runs in the browser or server, supports multiple scripts, and can power assistive entry methods like Dasher’s continuous‑gesture text input. For the AI/ML community this is a useful reminder that classical models still have strong practical value: PPM uses variable‑length contexts to balance specificity and robustness, delivering efficient next‑character probabilities when the data distribution is well represented in training. Unlike large neural LLMs it won’t generalize from untrained patterns, but it’s far cheaper to run, simpler to deploy, and well suited for multilingual assistive tech, low‑resource environments, and privacy‑sensitive applications. The package’s word‑prediction and error‑tolerance additions make PPM more practical for real interfaces, enabling developers to add responsive, multilingual prediction without heavyweight model stacks.