Dim Wihl Gat Tun: The Case for Linguistic Expertise in NLP for Under-Documented Languages

Findings of the Association for Computational Linguistics: ACL 2022, pages 169 - 182 May 22-27, 2022 c⃝2022 Association for Computational Linguistics Query and Extract: Refining Event Extraction as Type-oriented Binary Decoding Sijia Wang1, Mo Yu2, Shiyu Chang3, Lichao Sun4, Lifu Huang1 1Virginia Tech 2WeChat AI 3University of California Santa Barbara 4Lehigh University 1{sijiawang,lifuh}@vt.edu, 2moyumyu@tencent.com 3chang87@ucsb.edu, 4lis221@lehigh.edu Abstract Event extraction is typically modeled as a multi-class classification problem where event types and argument roles are treated as atomic symbols. These approaches are usually lim- ited to a set of pre-defined types. We propose a novel event extraction framework that uses event types and argument roles as natural lan- guage queries to extract candidate triggers and arguments from the input text. With the rich semantics in the queries, our framework ben- efits from the attention mechanisms to better capture the semantic correlation between the event types or argument roles and the input text. Furthermore, the query-and-extract for- mulation allows our approach to leverage all available event annotations from various on- tologies as a unified model. Experiments on ACE and ERE demonstrate that our approach achieves state-of-the-art performance on each dataset and significantly outperforms existing methods on zero-shot event extraction.1 1 Introduction Event extraction (Grishman, 1997; Chinchor and Marsh, 1998; Ahn, 2006) is a task to identify and type event triggers and participants from natural language text. As shown in Figure 1, married and left are triggers of two event mentions of the Marry and Transport event types respectively. Two argu- ments are involved in the left event mention: she is an Artifact, and Irap is the Destination. Traditional studies usually model event extrac- tion as a multi-class classification problem (Mc- Closky et al., 2011; Li et al., 2013; Chen et al., 2015; Yang and Mitchell, 2016; Nguyen et al., 2016; Lin et al., 2020), where a set of event types are first defined, and then supervised machine learn- ing approaches will detect and classify each can- didate event mention or argument into one of the 1Our code is open sourced at https://github.com/ VT-NLP/Event_Query_Extract for reproduction pur- pose. Figure 1: An example of event annotation. target types. However, each event type or argument role is treated as an atomic symbol, ignoring their rich semantics in these approaches. Several studies explore the semantics of event types by leveraging the event type structures (Huang et al., 2018), seed event mentions (Bronstein et al., 2015; Lai and Nguyen, 2019), or question answering (QA) (Du and Cardie, 2020; Liu et al., 2020). However, these approaches are still designed for and thus limited to a single target event ontology2, such as ACE (Con- sortium, 2005) or ERE (Song et al., 2015). With the existence of multiple ontologies and the challenge of handling new emerging event types, it is necessary to study event extraction approaches that are generalizable and can use all available train- ing data from distinct event ontologies.3 To this end, we propose a new event extraction framework following a query-and-extract paradigm. Our framework represents event types …

Content truncated to 500 words