Data-to-Text Generation for Games: Factual Verbalization and Character-Conditioned Commentary
Master’s thesis · CentraleSupélec · August 2023
BDMA programme · Research conducted at Ubisoft Paris Studio
Advisor: Mélanie Lopez Malet
Read the full thesis (PDF, 4.66 MB)
Abstract
Generating language from game data requires both preserving event information and expressing it in a voice appropriate to the speaker. This thesis investigates these requirements through two complementary tasks: factual verbalization of game-event records and tone-controlled or character-conditioned commentary. For factual verbalization, we implement a modular pipeline that serializes records, generates individual sentences, and combines them through sentence fusion. For commentary, we condition generation on tone descriptions and character profiles, then construct a synthetic corpus through API-generated seeds, human review, few-shot expansion with Llama 2, model-based filtering, and text cleaning. The resulting corpus contains 8,881 training, 426 validation, and 304 test examples covering seven characters. We use this corpus to train character-specific low-rank adapters for a FLAN-T5-based model. Evaluation of 30,000 verbalized records finds that 94.5% retain every input value under a value-matching criterion. Held-out commentary evaluation combines MAUVE, ROUGE-L, and BLEU-4 with qualitative analysis of shared-scene outputs. The examples exhibit character-associated expressions, while the different metric rankings distinguish distributional similarity from lexical overlap. Together, the study provides a documented workflow connecting structured game data, explicit style conditioning, and compact model adaptation, with an empirical account of the resulting descriptions and commentary.
Keywords: data-to-text generation; game commentary; character conditioning; synthetic data; parameter-efficient fine-tuning.
Suggested citation
Zhang, Yanjian. 2023. Data-to-Text Generation for Games: Factual Verbalization and Character-Conditioned Commentary. Master’s thesis, CentraleSupélec.
This thesis reports work completed in August 2023. The publicly available PDF is a revised edition prepared in October 2026.