An MRC Framework for Semantic Role Labeling

Proceedings of the 29th International Conference on Computational Linguistics, pages 2188–2198 October 12–17, 2022. 2188 An MRC Framework for Semantic Role Labeling Nan Wang1, Jiwei Li2,3, Yuxian Meng2 Xiaofei Sun2,3, Han Qiu4, Ziyao Wang5, Guoyin Wang6 and Jun He1∗ 1Key Laboratory of Data Engineering and Knowledge Engineering of MOE, School of Information, Renmin University of China 2Shannon.AI,3Zhejiang University, 4Tsinghua University 5University of Leeds, 6Amazon Alexa AI {nanwang, hejun}@ruc.edu.cn, jiwei_li@shannonai.com Abstract Semantic Role Labeling (SRL) aims at recog- nizing the predicate-argument structure of a sentence and can be decomposed into two sub- tasks: predicate disambiguation and argument labeling. Prior work deals with these two tasks independently, which ignores the semantic con- nection between the two tasks. In this paper, we propose to use the machine reading com- prehension (MRC) framework to bridge this gap. We formalize predicate disambiguation as multiple-choice machine reading compre- hension, where the descriptions of candidate senses of a given predicate are used as options to select the correct sense. The chosen pred- icate sense is then used to determine the se- mantic roles for that predicate, and these se- mantic roles are used to construct the query for another MRC model for argument label- ing. In this way, we are able to leverage both the predicate semantics and the semantic role semantics for argument labeling. We also pro- pose to select a subset of all the possible se- mantic roles for computational efficiency. Ex- periments show that the proposed framework achieves state-of-the-art or comparable results to previous work. Code is available at https: //github.com/ShannonAI/MRC-SRL. 1 Introduction Semantic Role Labeling (SRL) aims at recogniz- ing the predicate-argument structure of a sentence. The classic PropBank-style SRL includes two tasks: predicate disambiguation and argument labeling. Predicate disambiguation determines the specific meaning of a predicate in a given context and ar- gument labeling identifies the arguments of the predicate and assign them with the corresponding semantic roles, where each argument is a text span in the sentence. PropBank defines two types of semantic roles for argument labeling: core roles and non-core roles (Bonial et al., 2010). Core roles ∗Corresponding author. The stock has been beaten down for two days. [ A1 ] [beat.02][ A2 ] [ TMP ] sense id beat.02 sense push, cause motion A0 causer of motion roles A1 thing moving A2 direction, destination Figure 1: An example of SRL. A0, A1 and A2 are semantic roles for the sense id “beat.02”. The meanings of A0, A1 and A2 are respectively “causer of motion”, “thing moving” and “direction, destination”. are required roles that are in a close relation to the main verb in a sentence, such as agent and patient. There are seven core roles in PropBank: A0-A5 and AA. Non-core roles are modifiers, such as location (LOC) and time (TMP). The specific meanings of predicates and core roles are defined in the frame files. For example, for the sentence in Figure 1, the sense id of the predicate “beaten” is “beat.02”, and its three arguments span are “The …

Content truncated to 500 words