IndoNLG: Benchmark and Resources for Evaluating Indonesian Natural Language Generation

Natural language generation (NLG) benchmarks provide an important avenue to\nmeasure progress and develop better NLG systems. Unfortunately, the lack of\npublicly available NLG benchmarks for low-resource languages poses a\nchallenging barrier for building NLG systems that work well for languages with\nlimited amounts of data. Here we introduce IndoNLG, the first benchmark to\nmeasure natural language generation (NLG) progress in three low-resource -- yet\nwidely spoken -- languages of Indonesia: Indonesian, Javanese, and Sundanese.\nAltogether, these languages are spoken by more than 100 million native\nspeakers, and hence constitute an important use case of NLG systems today.\nConcretely, IndoNLG covers six tasks: summarization, question answering,\nchit-chat, and three different pairs of machine translation (MT) tasks. We\ncollate a clean pretraining corpus of Indonesian, Sundanese, and Javanese\ndatasets, Indo4B-Plus, which is used to pretrain our models: IndoBART and\nIndoGPT. We show that IndoBART and IndoGPT achieve competitive performance on\nall tasks -- despite using only one-fifth the parameters of a larger\nmultilingual model, mBART-LARGE (Liu et al., 2020). This finding emphasizes the\nimportance of pretraining on closely related, local languages to achieve more\nefficient learning and faster inference for very low-resource languages like\nJavanese and Sundanese.\n