Label Definitions Improve Semantic Role Labeling

Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 5613 - 5620 July 10-15, 2022 ©2022 Association for Computational Linguistics Label Definitions Improve Semantic Role Labeling Li Zhang∗ University of Pennsylvania zharry@seas.upenn.edu Ishan Jindal IBM Research ishan.jindal@ibm.com Yunyao Li† Apple Inc. yunyaoli@apple.com Abstract Argument classification is at the core of Se- mantic Role Labeling. Given a sentence and the predicate, a semantic role label is assigned to each argument of the predicate. While se- mantic roles come with meaningful definitions, existing work has treated them as symbolic. Learning symbolic labels usually requires am- ple training data, which is frequently unavail- able due to the cost of annotation. We instead propose to retrieve and leverage the definitions of these labels from the annotation guidelines. For example, the verb predicate “work” has arguments defined as “worker”, “job”, “em- ployer”, etc. Our model achieves state-of-the- art performance on the CoNLL09 English SRL dataset injected with label definitions given the predicate senses. The performance improve- ment is even more pronounced in low-resource settings when training data is scarce.1 1 Introduction Semantic role labeling (SRL) is an essential NLP task of answering the question of “who did what to whom, when, where and how.” Formally, a seman- tic role is assigned to each argument of a predicate in a sentence. SRL has been shown to help a wide range of NLP applications such as natural language inference (Zhang et al., 2020c), question answering (Zhang et al., 2020c; Maqsud et al., 2014; Yih et al., 2016) and machine translation (Shi et al., 2016). It can also be used as a pre-processing step for tasks such as information extraction (Niklaus et al., 2018; Zhang et al., 2020a). Learning from ample labeled examples is the predominant paradigm in many NLP tasks (Schick and Schütze, 2021), including SRL. However, la- beled data is costly and often lacking in many tasks, domains, and languages. One attempt at this issue, ∗ Work done during an internship at IBM Research † Work done while at IBM Research 1Our data and code can be found at https://github. com/System-T/LabelAwareSRL. Figure 1: An illustration of our procedure of construct- ing SRL examples with label definitions. The sense is used to get possible arguments of a predicate. made possible by recent advancement of language models, is to inject task descriptions into the data so that models become “aware” of the task require- ments and the meaning of the labels. This tech- nique has successfully been used in sentiment anal- ysis (Schick and Schütze, 2021), event extraction (Du and Cardie, 2020; Zhang et al., 2021), intent detection (Zhang et al., 2020b), word sense dis- ambiguation (Kumar et al., 2019), and many other tasks (Brown et al., 2020). In SRL, while the la- bel space is particularly sparse as each predicate has different semantic roles, definitions are readily available for all possible arguments of supported predicates. While early work has used label defini- tions for frame generalization …

Content truncated to 500 words