Enhancing MOBA Game Commentary Generation with Fine-Grained Prototype Retrieval.

With the development of the esports industry, more and more people are immersing themselves in watching various competitive matches, such as MOBA (Multiplayer Online Battle Arena) matches. Although MOBA games are attractive, the complexity of the games themselves also makes it difficult for many audiences to enjoy them easily without the assistance of professional commentators. This work studies using AI techniques to generate game commentaries automatically. Compared to human commentators, AI commentators can be more objective and work at any time and place at a low cost. Following the previous MOBA-E2C framework, we first use event handlers to extract various highlight events from the game metadata and organize them as event tables; then, this task can be regarded as a table-to-text task. Subsequently, this work proposes a BART-based MOBA-FPBART framework for further improving the generation quality of MOBA game commentaries by retrieving the human-written prototypes as guidance. On the one hand, in few-shot scenarios, we use a Fine-Grained Prototype Retrieval method to retrieve more relevant prototypes based on the characteristics of event tables. On the other hand, we also use a Corse-Grained Prototype Retrieval method in zero-shot scenarios. Experimental results on Dota2-Commentary have demonstrated our approach can notably outperform previous SOTA MOBA-FuseGPT in various metrics.