Frame Shift Prediction

Proceedings of the 13th Conference on Language Resources and Evaluation (LREC 2022), pages 976–986 Marseille, 20-25 June 2022 © European Language Resources Association (ELRA), licensed under CC-BY-NC-4.0 976 Frame Shift Prediction Zheng-Xin Yong1,†, Patrick D. Watson2, Tiago Timponi Torrent3, Oliver Czulo4, Collin F. Baker5 1Brown University 2Minerva University 3Federal University of Juiz de Fora 4University of Leipzig 5International Computer Science Institute zheng xin yong@brown.edu, pwatson@uni.minerva.edu, tiago.torrent@ufjf.edu.br, czulo@uni-leipzig.de, collinb@icsi.berkeley.edu Abstract Frame shift is a cross-linguistic phenomenon in translation which results in corresponding pairs of linguistic material evoking different frames. The ability to predict frame shifts would enable (semi-)automatic creation of multilingual frame annotations and thus speeding up FrameNet creation through annotation projection. Here, we first characterize how frame shifts result from other linguistic divergences such as translational divergences and construal differences. Our analysis also shows that many pairs of frames in frame shifts are multi-hop away from each other in Berkeley FrameNet’s net-like configuration. Then, we propose the Frame Shift Prediction task and demonstrate that our graph attention networks, combined with auxiliary training, can learn cross-linguistic frame-to-frame correspondence and predict frame shifts. Keywords: frame semantics, frame shifts, FrameNet, Multilinguality, Graph Neural Networks 1. Introduction Frame Semantics is an approach to meaning that char- acterizes the background knowledge against which lin- guistic entities are understood via systems of interre- lated concepts called frames (Fillmore, 1982). Frames are evoked by lexical units (LUs) and involve par- ticipants and props characterized as frame elements (see Figure 1 for example annotations in English and Brazilian Portuguese). The Berkeley FrameNet (BFN) project establishes a general-purpose resource for frame semantic descriptions of English (Ruppen- hofer et al., 2016) and is widely adopted for many NLP applications. BFN has successfully been adapted for other languages such as Brazilian Portuguese (Torrent and Ellsworth, 2013; Torrent et al., 2018b) and Ger- man (Burchardt et al., 2006; Boas and Ziem, 2018). The success of these works implies that some frames are applicable across different languages, while others can be adapted to fit language specificities (Gilardi and Baker, 2018; Baker and Lorenzi, 2020). Nonetheless, researchers working with non- English FrameNets find that differences in lexical- constructional patterns between languages result in frame shifts (Subirats and Sato, 2003; Litkowski, 2009; Pad´o and Lapata, 2009; Czulo, 2017; Lind´en et al., 2019; Giouli et al., 2020; Ohara, 2020; Ellsworth et al., 2021). Because FrameNet frames are sensitive to variations e.g. in valency or syntactic configuration, when a given sentence is translated from one language to another, it is frequently the case that the frames evoked will be significantly different for orginals and translations. In a test set of 30 sentences, Torrent et al. (2018a) find a direct frame correspondence between sentences in English and their translation to Brazilian Portuguese of only 0.51. Here, as shown in Figure 1, † Work done when studying at Minerva University as an undergraduate. we focus on a type of frame shift where, in a pair of parallel sentences, the corresponding LUs evoke two different frames. The issue of frame …

Content truncated to 500 words