Learning Kernel-Smoothed Machine Translation with Retrieved Examples

Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 742–750 November 7–11, 2021. c⃝2021 Association for Computational Linguistics 742 Smelting Gold and Silver for Improved Multilingual AMR-to-Text Generation Leonardo F. R. Ribeiro†, Jonas Pfeiffer†, Yue Zhang‡ and Iryna Gurevych† †Ubiquitous Knowledge Processing Lab, Technical University of Darmstadt ‡School of Engineering, Westlake University ribeiro@aiphes.tu-darmstadt.de Abstract Recent work on multilingual AMR-to-text gen- eration has exclusively focused on data aug- mentation strategies that utilize silver AMR. However, this assumes a high quality of gen- erated AMRs, potentially limiting the trans- ferability to the target task. In this paper, we investigate different techniques for auto- matically generating AMR annotations, where we aim to study which source of information yields better multilingual results. Our mod- els trained on gold AMR with silver (machine translated) sentences outperform approaches which leverage generated silver AMR. We find that combining both complementary sources of information further improves multilingual AMR-to-text generation. Our models surpass the previous state of the art for German, Italian, Spanish, and Chinese by a large margin.1 1 Introduction AMR-to-text generation is the task of recover- ing a text with the same meaning as a given Ab- stract Meaning Representation (AMR) (Banarescu et al., 2013), and has recently received much re- search interest (Ribeiro et al., 2019; Wang et al., 2020; Mager et al., 2020; Harkous et al., 2020; Fu et al., 2021). AMR has applications to a range of NLP tasks, including summarization (Hardy and Vlachos, 2018) and spoken language understand- ing (Damonte et al., 2019), and has the potential power of acting as an interlingua that allows the generation of text in many different languages (Da- monte and Cohen, 2018; Zhu et al., 2019). While previous work has predominantly focused on monolingual English settings (Cai and Lam, 2020b; Bevilacqua et al., 2021), recent work has also studied multilinguality in meaning represen- tations (Blloshmi et al., 2020; Sheth et al., 2021). Whereas Damonte and Cohen (2018) demonstrate 1Our code and checkpoints are available at https://github.com/UKPLab/m-AMR2Text. :ARG1 :ARG0 glorious :ARG0 look-02 live-01 they Their life looks glorious. Su vida parece gloriosa. En Es It La loro vita sembra gloriosa. AMR-to Ihr Leben sieht herrlich aus. 他们的⽇⼦看起来很光鲜. De Zh Figure 1: A generation example from English AMR to multiple different languages. that parsers can be effectively trained to transform multilingual text into English AMR, Mille et al. (2018, 2019) and Fan and Gardent (2020) discuss the reverse task, turning meaning representations into multilingual text, as shown in Figure 1. How- ever, gold-standard multilingual AMR training data is currently scarce, and previous work (Fan and Gardent, 2020) while discussing the feasibility of multilingual AMR-to-text generation, has inves- tigated synthetically generated AMR as the only source of silver training data. In this paper, we aim to close this gap by provid- ing an extensive analysis of different augmentation techniques to cheaply acquire silver-standard mul- tilingual AMR-to-text data: (1) Following Fan and Gardent (2020), we parse English sentences into silver AMRs from parallel multilingual corpora (SILVERAMR), resulting …

Content truncated to 500 words