Indonesian AMR-to-Text Generation by Language Model Fine-tuning
Taufiq Husada Daryanto, Masayu Leylia Khodra · 2022
Research related to generating text from Abstract Meaning Representation (AMR) for Indonesian Language is still a few and only using simple method (e.g., a heuristic approach that only generates bag of words from AMR). Therefore, a better method is needed for Indonesian AMR-to-text generation. Methods based on fine-tuning pretrained language models have become state-of-the-art for AMR-to-text generation. However, most of AMR-to-text generation research only focus on generating English text from AMR. In this paper, we develop a model for Indonesian AMR-to-text generation by fine-tuning a pretrained language model. We evaluate IndoT5, IndoBART, and mT5 on Indonesian AMR-to-text generation. We also observe the effect of adding supervised task adaptation and tree-level embedding on the model performance. Finally, we use our AMR-to-text generation model to improve the previous Indonesian AMR-based summarization system.