Tracking Player Sentiment Across Local APAC Languages

Reading thousands of posts in Traditional Chinese, Tagalog, and other regional APAC languages with Natural Language Processing |Weekly sentiment/mood signals for publishing strategy

Riot Games

The problem

Community mood over our games was dispersed over niche local forums and sites which were not tracked by enterprise sentiment trackers, and spread across many languages over 8 APAC territories (Vietnam, Taiwan, Indonesia, Malaysia, Singapore, Japan, INSA, Australia and New Zealand). Player sentiment was therefore not a reliable, known input to publishing decisions.

The approach

Scrape posts from forums & social media sites (copyright/privacy considered) that players actually use, with added weight on local-language ones. Score every post with a multilingual transformer tuned for Traditional Chinese, Tagalog, Bahasa Indonesia and so on. Then cluster into themes, and aggregate to a weekly signal report scannable in minutes.

What it revealed

Sentiment moves ahead of the lagging metrics like Monthly Actives, Gaming Hours, and Player Spend. Community mood about Game patch updates, New Character/Skin drops, Publishing & Content Marketing Beats, Esports, and other key events works as an early-warning layer.

Also, local language discussion forums often contain a richer spectrum of discussion, and the Natural Language Processor model needs to be fine-tuned to pick up on current slang, hyper-local symbology, and sarcasm.

This was achieved in conjunction with local Publishing office teams who helped identify existing model effectiveness (available on HuggingFace); as well as train models to pick up on relevant terms/symbols.

Under the hood

Python · BeautifulSoup · XLM-RoBERTa · Natural Language Processing · pandas