Creating a High Quality Abstract Meaning Representation Dataset Automatically

Abstract

As only a few gold training datasets are available today, Abstract Meaning Representation (AMR) parsers are mainly trained on AMR 3.0, the largest dataset (Knight et al., 2020) which contains 55k sentences for training. Even if great progress has been made, leading to parsers that can reach Smatch scores higher than 83% on the AMR 3.0 test dataset, this is not accurate enough to be used in real world application pipelines. More data could help improve performance, but manually annotating sentences is costly. So, we have investigated an approach to automatically create synthetic data using different existing tools and models trained on AMR 3.0. This leads to better parsing performance with Smatch scores increased by 1 to 2 points (depending on the 3 gold test datasets used) with models trained on the augmented data.

Resources

Details