Script Parsing with Hierarchical Sequence Modelling

Proceedings of the 10th Conference on Lexical and Computational Semantics, pages 195–201 August 5–6, 2021, Bangkok, Thailand (online) ©2021 Association for Computational Linguistics 195 Script Parsing with Hierarchical Sequence Modelling Fangzhou Zhai, Iza ˇSkrjanec and Alexander Koller Dept. of Language Science and Technology Saarland Informatics Campus Saarland University {fzhai, skrjanec, koller}@coli.uni-saarland.de Abstract Scripts (Schank and Abelson, 1977) capture commonsense knowledge about everyday ac- tivities and their participants. Script knowl- edge has been shown to be useful in a num- ber of NLP tasks, such as referent prediction, discourse classification, and story generation. A crucial step for the exploitation of script knowledge is script parsing, the task of tag- ging a text with the events and participants from a certain activity. This task is challeng- ing: it requires information both about the ways events and participants are usually real- ized in surface language as well as the order in which they occur in the world. We show how to do accurate script parsing with a hierar- chical sequence model. Our model improves the state of the art of event parsing by over 16 points F-score and, for the first time, accu- rately tags script participants. 1 Introduction Script knowledge is a category of common sense knowledge that describes how people conduct ev- eryday activities sequentially (Schank and Abelson, 1977). Script knowledge of a specific scenario, e.g. GROCERY SHOPPING, includes the events that comprise the scenario, the participants involved, and the relations between them. Script knowledge is useful for various downstream NLP applications, such as referent prediction (Ahrendt and Demberg, 2016; Modi et al., 2017), discourse sense classifica- tion (Lee et al., 2020), story generation (Zhai et al., 2019, 2020). Script parsing identifies pre-defined sets of script events and participants from surface text (see Figure 1). For a specific scenario, script parsing es- sentially boils down to determining what each verb and each NP (which we term candidate) refers to in the context of that scenario. Script parsing is an under-investigated, complex task. It is highly contextualized and corresponds to Figure 1: Descriptions of FIXING A FLAT TIRE from InScript. Script parsing identifies events and participants from surface text. each specific scenario. The task is challenging even for humans: the inter-annotator agreement is quite modest, at 0.64 and 0.77 Fleiss’ κ for event and participant parsing, respectively (Modi et al., 2016). Various factors need to be taken into consideration for this task. (1) At the local level, the basic seman- tics of the candidates. (2) At the discourse level, the sequence of events and participants should sketch a reasonable agenda for the activity. For example, the events must occur in a feasible order; when an NP is a dependent of an adjacent verb, the pre- dicted participant type must be one that participates in the predicted event. In Figure 1, the same verb found was assigned different event classes: found bike pump as get tools whereas found a small glass shard as examine tire. One would have to consider the arguments of …

Content truncated to 500 words