Extracting Space Situational Awareness Events from News Text

Proceedings of the 13th Conference on Language Resources and Evaluation (LREC 2022), pages 6077–6082 Marseille, 20-25 June 2022 © European Language Resources Association (ELRA), licensed under CC-BY-NC-4.0 6077 Extracting Space Situational Awareness Events from News Text Zhengnan Xie†, Alice Saebom Kwak†, Enfa George†, Laura W. Dozal†, Hoang Van†, Moriba Jah‡, Roberto Furfaro†, Peter Jansen† †University of Arizona, USA ‡University of Texas at Austin, USA pajansen@arizona.edu Abstract Space situational awareness typically makes use of physical measurements from radar, telescopes, and other assets to monitor satellites and other spacecraft for operational, navigational, and defense purposes. In this work we explore using textual input for the space situational awareness task. We construct a corpus of 48.5k news articles spanning all known active satellites between 2009 and 2020. Using a dependency-rule-based extraction system designed to target three high-impact events – spacecraft launches, failures, and decommissionings, we identify 1,787 space-event sentences that are then annotated by humans with 15.9k labels for event slots. We empirically demonstrate a state-of-the-art neural extraction system achieves an overall F1 between 53 and 91 per slot for event extraction in this low-resource, high-impact domain. Keywords: information extraction, corpus, space 1. Introduction Space Situational Awareness (SSA) is the decision- making knowledge required to predict, avoid, operate through, or recover from the loss, disruption, or degra- dation of space services, capabilities, or activities (Cai et al., 2020). Most governments approach SSA as a physical measurement problem, using global arrays of optical telescopes, radar facilities, radio listening posts, and other assets to directly measure spacecraft prop- erties, orbits, or communications. In addition to be- ing costly, this approach has several known limitations, chief among them that physical measurements (such as that a spacecraft is spinning) alone cannot fully cap- ture causal information (e.g. a spacecraft suffered a gyro failure) that can be critical when interpreting and responding to events within the space object population (Walls et al., 2016). In this work we explore extracting space-domain events from non-physical measurements, namely tex- tual input from press releases and news articles. SSA is frequently accomplished by collecting and fusing multi-modal data (Esteva et al., 2020; Le May et al., 2020), where text is a comparatively inexpensive knowledge resource. We hypothesize that text can sup- ply information complementary to physics-based mea- surements, such as the published causes of spacecraft failures, whether spacecraft operators have planned as- tronavigation maneuvers, or whether a spacecraft has been safely decommissioned according to United Na- tions guidelines. Information extraction (IE) is often reported in high-resource domains where favorable performance is possible using pre-trained models. In contrast, there exist high-importance domains (e.g. nuclear, defense, etc.) that remain under-explored due to low resource availability, and low task transfer from near-domains. Space-domain information extraction poses these chal- lenges in abundance – comparatively few spacecraft # Launch Vehicle with Satellite - name: launchRule type: event pattern: | trigger = [(lemma=/(launch|send|place)/ & tag=/V.*/)] org = >nsubj ${org} org? = >> ${org} vehicle = >dobj >amod? >compound? ${vehicle} ${numberGeneric}? /rocket|launcher/? satellite = >dobj? >nmod_with ${satellite} ${numberGeneric}? /satellites?/? …

Content truncated to 500 words