Zero-shot Script Parsing

Proceedings of the 29th International Conference on Computational Linguistics, pages 4049–4060 October 12–17, 2022. 4049 Zero-Shot Script Parsing Fangzhou Zhai, Vera Demberg and Alexander Koller Dept. of Language Science and Technology Saarland Informatics Campus Saarland University {fzhai, vera, koller}@coli.uni-saarland.de Abstract Script knowledge (Schank and Abelson, 1977) is useful for a variety of NLP tasks. However, existing resources only cover a small number of activities, limiting its practical usefulness. In this work, we propose a zero-shot learn- ing approach to script parsing, the task of tagging texts with scenario-specific event and participant types, which enables us to acquire script knowledge without domain-specific an- notations. We (1) learn representations of po- tential event and participant mentions by pro- moting class consistency according to the anno- tated data; (2) perform clustering on the event / participant candidates from unannotated texts that belongs to an unseen scenario. The model achieves 68.1/74.4 average F1 for event / par- ticipant parsing, respectively, outperforming a previous CRF model that, in contrast, has ac- cess to scenario-specific supervision. We also evaluate the model by testing on a different cor- pus, where it achieved 55.5/54.0 average F1 for event / participant parsing. 1 Introduction Script knowledge is a type of commonsense knowledge that captures how people conduct ev- eryday activities (Schank and Abelson, 1977). It expresses that in a certain scenario, participants tend to act out events in a certain order; an ex- ample from the scenario FIXING A FLAT TIRE is shown in Fig. 1. Humans use script knowledge to fill in events that are not explicitly mentioned in a text, and script knowledge is useful for many down- stream NLP applications, including referent pre- diction (Ahrendt and Demberg, 2016; Modi et al., 2017), discourse classification (Lee et al., 2020), and story generation (Zhai et al., 2019, 2020). A key challenge in dealing with script knowl- edge is coverage: it is costly and time-consuming to spell out the prototypical events and participants of a scenario and how they can be expressed in language. Existing resources are mostly crowd- Figure 1: A story about FIXING A FLAT TIRE from InScript. Script parsing identifies events and participants from texts. The picture is taken from Zhai et al. (2021). Figure 2: A part of the temporal script graph for TAKING A BATH inferred from our parsing result. Each node illustrates 3 random verbalizations from the cluster. The gold classes are shown on the side. Further edges that could be inferred by transitivity are omitted. We see one could either undress first or add scent (to the bath tub) first before sink into water. sourced (Regneri et al., 2010; Modi et al., 2016); they annotate stories from a limited number of sce- narios with script events and participants (cf. figure 1). Script parsers, which predict these event and participant labels given a text, can achieve high ac- curacies on scenarios that were seen in training (Os- termann et al., 2017; Zhai et al., 2021). Nonethe- less, these parsers only operate on known …

Content truncated to 500 words