Multi-Sentence Argument Linking

Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 8057–8077 July 5 - 10, 2020. c⃝2020 Association for Computational Linguistics 8057 Multi-Sentence Argument Linking Seth Ebner∗ Patrick Xia∗ Ryan Culkin Kyle Rawlins Benjamin Van Durme Johns Hopkins University {seth, paxia}@cs.jhu.edu {rculkin, kgr, vandurme}@jhu.edu Abstract We present a novel document-level model for finding argument spans that fill an event’s roles, connecting related ideas in sentence- level semantic role labeling and coreference resolution. Because existing datasets for cross-sentence linking are small, development of our neural model is supported through the creation of a new resource, Roles Across Multiple Sentences (RAMS), which contains 9,124 annotated events across 139 types. We demonstrate strong performance of our model on RAMS and other event-related datasets.1 1 Introduction Textual event descriptions may span multiple sen- tences, yet large-scale datasets predominately an- notate for events and their arguments at the sen- tence level. This has driven researchers to focus on sentence-level tasks such as semantic role la- beling (SRL), even though perfect performance at such tasks would still enable a less than complete understanding of an event at the document level. In this work, we approach event understanding as a form of linking, more akin to coreference res- olution than sentence-level SRL. An event trig- ger evokes a set of roles regarded as latent argu- ments, with these implicit arguments then poten- tially linked to explicit mentions in the text. Consider the example in Figure 1: the AirstrikeMissileStrike event (triggered by “bombarding”) gives rise to a frame or set of type- level roles (attacker, target, instrument, place) with the referents (“Russians”, “rebel out- post”, “aircraft”, “Syria”).2 Intuitively we recog- nize the possible existence of fillers for these roles, for example, the place of the particular Air- ∗Equal Contribution 1Data and code at http://nlp.jhu.edu/rams/. 2ϵ would indicate there is no explicit referent in the text. Figure 1: A passage annotated for an event’s type, trigger, and arguments. Each arc points from the trig- ger to the argument that fills the labeled role. strikeMissileStrike event. These implicit ar- guments are linked to explicit arguments in the document (i.e., text spans). We refer to the task of finding explicit argument(s) to fill each role for an event as argument linking. Prior annotation of cross-sentence argument links has produced small datasets, with a focus ei- ther on a small number of predicate types (Gerber and Chai, 2010, 2012; Feizabadi and Pad´o, 2014) or on a small number of documents (Ruppenhofer et al., 2010). To enable the development of a neu- ral model for argument linking, we produce Roles Across Multiple Sentences (RAMS), a dataset of 9,124 annotated events from news based on an on- tology of 139 event types and 65 roles. In a 5- sentence window around each event trigger, we annotate the closest argument span for each role. Our model builds on recent ideas in span se- lection models (Lee et al., 2018; He et al., 2018; Ouchi et al., 2018), used in this work for …

Content truncated to 500 words