JSEEGraph: Joint Structured Event Extraction as Graph Parsing

Proceedings of the The 12th Joint Conference on Lexical and Computational Semantics (*SEM 2023), pages 115–127 July 13-14, 2023 ©2023 Association for Computational Linguistics JSEEGraph: Joint Structured Event Extraction as Graph Parsing Huiling You1, Samia Touileb2 and Lilja Øvrelid1 1University of Oslo 2University of Bergen {huiliny, liljao}@ifi.uio.no samia.touileb@uib.no Abstract We propose a graph-based event extraction framework JSEEGraph that approaches the task of event extraction as general graph parsing in the tradition of Meaning Representation Pars- ing. It explicitly encodes entities and events in a single semantic graph, and further has the flexibility to encode a wider range of additional IE relations and jointly infer individual tasks. JSEEGraph performs in an end-to-end manner via general graph parsing: (1) instead of flat sequence labelling, nested structures between entities/triggers are efficiently encoded as sepa- rate nodes in the graph, allowing for nested and overlapping entities and triggers; (2) both enti- ties, relations, and events can be encoded in the same graph, where entities and event triggers are represented as nodes and entity relations and event arguments are constructed via edges; (3) joint inference avoids error propagation and enhances the interpolation of different IE tasks. We experiment on two benchmark datasets of varying structural complexities; ACE05 and Rich ERE, covering three languages: English, Chinese, and Spanish. Experimental results show that JSEEGraph can handle nested event structures, that it is beneficial to solve differ- ent IE tasks jointly, and that event argument extraction in particular benefits from entity ex- traction. Our code and models are released as open-source1. 1 Introduction Event extraction (EE) deals with the extraction of complex, structured representations of events from text, including overlapping and nested structures (Sheng et al., 2021; Cao et al., 2022). While there are existing datasets annotated with such rich repre- sentations (Doddington et al., 2004; Song et al., 2015), a majority of current approaches model this task using simplified versions of these datasets or sequence-labeling-based encodings which are 1https://github.com/huiling-y/ JSEEGraph , purposely trigger buy PER I commodity things made. in Canada or USA. made trigger Canada GPE USA GPE giver recipient thing artifact place place artifact transfermoney transferownership artifact artifact Figure 1: Example of nested and overlapping events in the sentence “I, purposely buy things made in Canada or USA.”, taken from Rich ERE (Song et al., 2015). not capable of capturing the full complexity of the events. Figure 1 shows an example from the Rich ERE dataset (Song et al., 2015) of a sentence containing both nested and overlapping events: “buy” serves as trigger for two overlapping events, transfermoney and transferownership with their respective argument roles, and similarly “made” for two artifact events triggered by the coordination of two GPE entities Canada and USA; at the same time, the event trigger “made” is nested inside the entity span “things made in Canada or USA”. For this example, models based on token tag- ging (such as the commonly used BIO-encoding) would fail completely when a token contributes to multiple information extraction elements. In this case, the version …

Content truncated to 500 words