Multilingual Persuasion Detection: Video Games as an Invaluable Data Source for NLP

This study presents a multilingual persuasion detection dataset built from dialogue in three Bioware role-playing games, Neverwinter Nights, Knights of the Old Republic, and Knights of the Old Republic II, spanning English, Spanish, German, French, and Italian. The corpus draws its labels from in-game "[Persuade]" tags aligned across languages through string identifiers, yielding roughly 7,900 persuasion sentences and 33,600 non-persuasion sentences, an approximately 80/20 class split. The authors argue that video games are an underused source of semi-annotated NLP data and that this resource is likely the largest multilingual persuasion dataset to date, extending prior work that concentrated almost entirely on English.

The central thesis is that game dialogue is a viable and valuable resource for natural language processing. To test this, the authors fine-tuned eleven BERT-based classifiers. Monolingual models performed strongly, achieving macro F1 scores near 0.8 and weighted F1 scores near 0.9. Multilingual BERT, by contrast, performed poorly on the persuasive class, with macro F1 near 0.5. The authors attribute this gap to a domain mismatch, since Wikipedia-based pretraining poorly suits game dialogue, rather than to multilingualism itself.

Error analysis indicates that the models reliably detect explicit cues such as currency references, commands, and promises, but struggle with ambiguous or context-dependent lines where persuasive intent depends on surrounding narrative. The scope of the work is therefore cross-lingual, covering five European languages, and focused on a single game genre and developer, with its findings positioned as a proof of concept for using interactive fiction as training and evaluation data for persuasion detection.