Iterative Span Selection: Self-Emergence of Resolving Orders in Semantic Role Labeling

Proceedings of the 29th International Conference on Computational Linguistics, pages 5383–5397 October 12–17, 2022. 5383 Iterative Span Selection: Self-Emergence of Resolving Orders in Semantic Role Labeling Shuhei Kurita1,2 Hiroki Ouchi3,1 Kentaro Inui4,1 Satoshi Sekine1 1RIKEN 2JST PRESTO 3Nara Institute of Science and Technology 4Tohoku University shuhei.kurita@riken.jp hiroki.ouchi@is.naist.jp inui@tohoku.ac.jp satoshi.sekine@riken.jp Abstract Semantic Role Labeling (SRL) is the task of labeling semantic arguments for marked seman- tic predicates. Semantic arguments and their predicates are related in various distinct man- ners, of which certain semantic arguments are a necessity while others serve as an auxiliary to their predicates. To consider such roles and relations of the arguments in the labeling order, we introduce iterative argument identification (IAI), which combines global decoding and iterative identification for the semantic argu- ments. In experiments, we first realize that the model with random argument labeling orders outperforms other heuristic orders such as the conventional left-to-right labeling order. Com- bined with simple reinforcement learning, the proposed model spontaneously learns the op- timized labeling orders that are different from existing heuristic orders. The proposed model with the IAI algorithm achieves competitive or outperforming results from the existing models in the standard benchmark datasets of span- based SRL: CoNLL-2005 and CoNLL-2012. 1 Introduction Semantic role labeling (Carreras and Màrquez, 2004, 2005) is the task of identifying and resolv- ing the relations between semantic predicates and their arguments based on the PropBank (Kingsbury and Palmer, 2002) predicate-argument structure. In span-based SRL, semantic predicates comprise several semantic arguments that are expressed as spans of tokens in the sentence. Recent span-based SRL models incorporate neural networks into a global decoding approach. Here syntactic features are injected into the neural network model (Strubell et al., 2018), span-based scoring for semantic ar- gument is adapted (Ouchi et al., 2018), and the unified representations for both span-based and dependency-based SRL are applied (Li et al., 2019; Zhou et al., 2020). However, in these approaches, the labeling order is not determined inside the neu- They can directly look at the agreement with us because (V) They can directly look at the agreement with us because A0 (V) A1 They can directly look at the agreement with us because (V) A1 They can directly look at the agreement with us because A0 AM-NMR (V) A1 AM-LOC AM-CAU They can directly look at the agreement with us because A0 AM-MOD AM-NMR (V) A1 τ = 0 τ = 1 τ = 2 τ = 4 τ = 6 AM-MOD Figure 1: Example of the iterative argument identifi- cation with the given predicate “look”. At the bottom of the final state τ = 6, all semantic arguments are presented with spans and labels. The time step τ corre- sponds the iterative process of the proposed model. ral network and hence models require some extrap- olated graph decoding procedures, similar to the graph-based approaches that rely on external graph decoding (Lewis et al., 2015). Such external decod- ing procedures are typically not trained during the model training and hence hinder accurate …

Content truncated to 500 words