BiBL: AMR Parsing and Generation with Bidirectional Bayesian Learning

Proceedings of the 29th International Conference on Computational Linguistics, pages 5461–5475 October 12–17, 2022. 5461 BiBL: AMR Parsing and Generation with Bidirectional Bayesian Learning Ziming Cheng1,2, Zuchao Li1,2,∗, and Hai Zhao1,2,∗ 1Department of Computer Science and Engineering, Shanghai Jiao Tong University 2MoE Key Lab of Artificial Intelligence, AI Institute, Shanghai Jiao Tong University {kk.cheng, charlee}@sjtu.edu.cn, zhaohai@cs.sjtu.edu.cn Abstract Abstract Meaning Representation (AMR) of- fers a unified semantic representation for nat- ural language sentences. Thus transformation between AMR and text yields two transition tasks in opposite directions, i.e., Text-to-AMR parsing and AMR-to-Text generation. Existing AMR studies only focus on one-side improve- ments despite the duality of the two tasks, and their improvements are greatly attributed to the inclusion of large extra training data or complex structure modifications which harm the infer- ence speed. Instead, we propose data-efficient Bidirectional Bayesian learning (BiBL) to fa- cilitate bidirectional information transition by adopting a single-stage multitasking strategy so that the resulting model may enjoy much lighter training at the same time. Evaluation on benchmark datasets shows that our proposed BiBL outperforms strong previous seq2seq re- finements without the help of extra data which is indispensable in existing counterpart mod- els. We release the codes of BiBL at: https: //github.com/KHAKhazeus/BiBL. 1 Introduction Abstract Meaning Representation (AMR) (Ba- narescu et al., 2013) is a formalism that could cap- ture the semantic meaning of a natural language sentence in a graph representation. An AMR graph example is shown in Figure 1. Leaves in the AMR graph denote concepts in the text sentence while other nodes within the graph are entities. Two en- tities could be connected with relations defined under a common standard, and instance relations connect entities and corresponding concepts. The flexibility and semantic invariance characteristics make AMR applicable to various natural language processing (NLP) downstream tasks and achieve solid performance. ∗Corresponding author. This work was supported by Key Projects of National Natural Science Foundation of China (U1836222 and 61733011). The boy wants to play. Text Format DFS-linearized Format (<R0> want-01 :ARG0 (<R1> boy) :ARG1 (<R2> play-01 :ARG0 <R1>)) Graph Format instance want-01 boy instance ARG0 ARG1 ARG0 instance play-01 Figure 1: AMR graph and DFS-linearized formats for representing a text sentence “The boy wants to play". In- stance relations are denoted with dotted red lines while other relations are denoted with black lines. Concepts in AMR graphs are enclosed by oval shapes. The inden- tation in the DFS-linearized format is only for clarity purposes. The transformation between AMR graphs and texts forms two basic tasks offering helpful AMR information for the downstream tasks. Traditional approaches treat AMR-to-Text generation and Text- to-AMR parsing separately. However, utilizing graph linearization and the power of pre-trained lan- guage models, SPRING (Bevilacqua et al., 2021) casts two tasks into a unified sequence-to-sequence (seq2seq) structure and outperforms all traditional approaches proposed before. Such progress consol- idates the prevalent position of seq2seq approaches tackling the AMR transition tasks. Recently, many methods have been proposed to further enhance the seq2seq model. In order to improve …

Content truncated to 500 words