MuCPAD: A Multi-Domain Chinese Predicate-Argument Dataset

Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 1707 - 1717 July 10-15, 2022 ©2022 Association for Computational Linguistics MuCPAD: A Multi-Domain Chinese Predicate-Argument Dataset Yahui Liu, Haoping Yang, Chen Gong∗, Qingrong Xia, Zhenghua Li, Min Zhang Institute of Artificial Intelligence, School of Computer Science and Technology, Soochow University, China {yahuiliu.nlp,hpyang3,gongchen.nlp}@foxmail.com, kirosummer.nlp@gmail.com, {zhli13,minzhang}@suda.edu.cn Abstract During the past decade, neural network mod- els have made tremendous progress on in- domain semantic role labeling (SRL). How- ever, performance drops dramatically under the out-of-domain setting. In order to facili- tate research on cross-domain SRL, this paper presents MuCPAD, a multi-domain Chinese predicate-argument dataset, which consists of 30,897 sentences and 92,051 predicates from six different domains. MuCPAD exhibits three important features. 1) Based on a frame-free annotation methodology, we avoid writing com- plex frames for new predicates. 2) We explic- itly annotate omitted core arguments to recover more complete semantic structure, considering that omission of content words is ubiquitous in multi-domain Chinese texts. 3) We compile 53 pages of annotation guidelines and adopt strict double annotation for improving data quality. This paper describes in detail the annotation methodology and annotation process of MuC- PAD, and presents in-depth data analysis. We also give benchmark results on cross-domain SRL based on MuCPAD. 1 Introduction As a fundamental NLP task, semantic role labeling (SRL), also known as shallow semantic parsing, aims to capture the major semantic information of a sentence based on predicate-argument structure. Basically, SRL tries to answer “who did what to whom where and when” (Màrquez et al., 2008). Previous works have shown that SRL can help var- ious downstream tasks, including information ex- traction (Bastianelli et al., 2013), plagiarism detec- tion (Paul and Jamal, 2015), machine translation (Shi et al., 2016), reading comprehension (Zhang et al., 2020), etc. Figure 1 gives two examples of SRL structure. According to the definition of semantic roles, there exist two typical representation forms, i.e., the ∗Corresponding author. 昨天 , 我 买 买 买 了 一件 新 裙子 。 Yesterday , I bought a new dress . time agent patient (a) Word-based SRL representation adopted in MuCPAD 昨天 , 我 买 买 买 了 一件 新 裙子 。 Yesterday , I bought a new dress . TIME A0 A1 (b) Span-based SRL representation adopted in CPB and CNB Figure 1: Examples of two SRL formulations. word-based and the span-based. This work adopts the word-based form, in which an argument corre- sponds to a single word. In contrast, span-based SRL, adopted by most previous datasets, takes a word span as an argument. The direction of arcs is from predicates to arguments, and the labels indi- cate the types of semantic roles. For example, the arc from “买(bought)” to “裙子(dress)” with a la- bel “patient” means that the semantic role between the predicate “买(bought)” and the argument “裙 子(dress)” is “patient”. Recently, Chinese SRL research has achieved tremendous progress, thanks to the rise of deep learning methods (Marcheggiani et …

Content truncated to 500 words