Multi-Label Classification on Natural Language Sentences for Video Game Design

This paper investigates the effectiveness of multi-label classification (MLC) in classifying controlled natural language sentences that used to write video games. The performance of such multi-label classifying tasks can help us both understand the complexities of program synthesis as well as providing suggestions to the users whenever a sentence is not understood. Real-user data is used in the experiments. The proposed classification paradigm shows that classification and prediction in code generation are feasible, and can offer suggestions to the users when the sentence is unclear. The impact of label correlation is also discussed.