Modeling Syntactic-Semantic Dependency Correlations in Semantic Role Labeling Using Mixture Models

Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 7959 - 7969 May 22-27, 2022 c⃝2022 Association for Computational Linguistics Modeling Syntactic-Semantic Dependency Correlations in Semantic Role Labeling Using Mixture Models Junjie Chen∗, Xiangheng He∗∗ and Yusuke Miyao∗ Department of Computer Science, The University of Tokyo, Tokyo∗ GLAM – Group on Language, Audio, & Music, Imperial College London, UK∗∗ {christopher, yusuke}@is.s.u-tokyo.ac.jp x.he20@imperial.ac.uk Abstract In this paper, we propose a mixture model- based end-to-end method to model the syntactic-semantic dependency correlation in Semantic Role Labeling (SRL). Semantic de- pendencies in SRL are modeled as a distribu- tion over semantic dependency labels condi- tioned on a predicate and an argument word. The semantic label distribution varies depend- ing on Shortest Syntactic Dependency Path (SSDP) hop patterns. We target the variation of semantic label distributions using a mixture model, separately estimating semantic label dis- tributions for different hop patterns and proba- bilistically clustering hop patterns with similar semantic label distributions. Experiments show that the proposed method successfully learns a cluster assignment reflecting the variation of semantic label distributions. Modeling the vari- ation improves performance in predicting short distance semantic dependencies, in addition to the improvement on long distance semantic de- pendencies that previous syntax-aware methods have achieved. The proposed method achieves a small but statistically significant improvement over baseline methods in English, German, and Spanish and obtains competitive performance with state-of-the-art methods in English. 1 1 Introduction Semantic Role Labeling (SRL) answers an essen- tial question about sentence semantics: “[Who] [does what] [to whom]”. A core problem of SRL is identifying semantic dependencies that specify the semantic role of arguments in relation to predicates (He et al., 2018; Kasai et al., 2019). For exam- ple, [who] (argument) is the agent (semantic role) to [does what] (predicate). Semantic dependency parsers (Dozat and Manning, 2018a) identify se- mantic dependencies by giving a distribution over semantic dependency labels (denoted as semantic label distribution) for all predicate-argument pairs. 1Our code is available at this repository. Figure 1: An example illustrating the impact of SSDPs on semantic label distributions. The solid underline highlights the predicate, and the dashed underline high- lights the arguments. Figure 2: An illustration of the proposed mixture model- based method for semantic dependency parsing. In this paper, we propose a mixture model (Pear- son, 1894) based semantic dependency parser for SRL where we target the dependence of semantic label distributions on Shortest Syntactic Depen- dency Path (SSDP) patterns. SSDP is the short- est path connecting a predicate-argument pair in a syntactic dependency tree. Bunescu and Mooney (2005) and Cai et al. (2016) claim that SSDP en- codes most information about bigram relations, such as the semantic dependency. Indeed, previous research (He et al., 2018; Xia et al., 2019) shows that modeling the correlation between SSDPs and semantic dependencies is crucial for building a high-performance SRL system. 7959 Semantic label distributions vary depending on SSDPs, even when the SSDPs connect predicate- argument pairs with the same surface …

Content truncated to 500 words