Simple and Effective Graph-to-Graph Annotation Conversion

Proceedings of the 29th International Conference on Computational Linguistics, pages 5450–5460 October 12–17, 2022. 5450 Simple and Effective Graph-to-Graph Annotation Conversion Yuxuan Wang♦∗, Zhilin Lei♡∗, Yuqiu Ji♡, Wanxiang Che♡† ♦Zhejiang Lab ♡Research Center for Social Computing and Information Retrieval, Harbin Institute of Technology yxwang@zhejianglab.com, {zllei, yqji, car}@ir.hit.edu.cn Abstract Annotation conversion is an effective way to construct datasets under new annotation guide- lines based on existing datasets with little hu- man labour. Previous work has been limited in conversion between tree-structured datasets and mainly focused on feature-based models which are not easily applicable to new conver- sions. In this paper, we propose two simple and effective graph-to-graph annotation conver- sion approaches, namely Label Switching and Graph2Graph Linear Transformation, which use pseudo data and inherit parameters to guide graph conversions respectively. These methods are able to deal with conversion between graph- structured annotations and require no manually designed features. To verify their effectiveness, we manually construct a graph-structured par- allel annotated dataset and evaluate the pro- posed approaches on it as well as other exist- ing parallel annotated datasets. Experimental results show that the proposed approaches out- perform strong baselines with higher conver- sion score. To further validate the quality of converted graphs, we utilize them to train the target parser and find graphs generated by our approaches lead to higher parsing score than those generated by the baselines.1 1 Introduction While tree-structured representations have domi- nated parsing for the last decade, graph-structured datasets are receiving growing interest in recent years (Oepen et al., 2019, 2020). Over the last few years, an increasing number of graph-structured datasets have become available. Some of them, such as DM corpora from the SemEval 2015 task 18 dataset (Oepen et al., 2015) and AMRBank (Ba- narescu et al., 2013), are manually annotated. Some ∗Equal Contribution. †Email Corresponding. 1Our code and dataset are available at https:// github.com/WangYuxuan93/G2GConversion. You make me want to smile ROOT AGT PAT eRESU EXP dCONT AGT mDEPD root nsubj obj xcomp xcomp mark Figure 1: Example of annotation conversion from Uni- versal Dependency Tree (bottom) to our Semantic De- pendency Graph (top). others, such as the Enhanced English Universal De- pendencies dataset (Schuster and Manning, 2016), are converted from existing dataset with manually designed rules. As illustrated in Figure 1, the Se- mantic Dependency Graph at the top is converted from the Universal Dependecy Tree at the bottom. Obviously, in the dataset construction process under a new annotation guideline, it would be ex- tremely expensive to annotate the whole dataset manually. Although rule-based conversion needs no human labour for annotation, it requires exper- tise to design the rules, which could be difficult if the new guideline is vastly different from the old one. Therefore, it would be efficient and attractive to exploit existing datasets and learn a transduc- tion that converts them into a new guideline. The converted dataset under the new guideline could be used in model training or further refined by human annotators to construct a high-quality dataset. Such conversion has …

Content truncated to 500 words