Automatic Learning of Modality Exclusivity Norms with Crosslingual Word Embeddings

Proceedings of the Ninth Joint Conference on Lexical and Computational Semantics (*SEM), pages 32–38 Barcelona, Spain (Online), December 12–13, 2020 32 Automatic Learning of Modality Exclusivity Norms with Crosslingual Word Embeddings Emmanuele Chersoni, Rong Xiang, Qin Lu, Chu-Ren Huang The Hong Kong Polytechnic University, 11 Yuk Choi Road, Hong Kong (China) {emmanuelechersoni, xiangrong0302}@gmail.com, csluqin@comp.polyu.edu.hk, churen.huang@polyu.edu.hk Abstract Collecting modality exclusivity norms for lex- ical items has recently become a common practice in psycholinguistics and cognitive re- search. However, these norms are available only for a relatively small number of lan- guages and often involve a costly and time- consuming collection of ratings. In this work, we aim at learning a mapping be- tween word embeddings and modality norms. Our experiments focused on crosslingual word embeddings, in order to predict modality asso- ciation scores by training on a high-resource language and testing on a low-resource one. We ran two experiments, one in a monolin- gual and the other one in a crosslingual setting. Results show that modality prediction using off-the-shelf crosslingual embeddings indeed has moderate-to-high correlations with human ratings even when regression algorithms are trained on an English resource and tested on a completely unseen language. 1 Introduction The expression modality exclusivity norms refers to a collection of words and their association scores with sensory modalities. The use of modality norms has been a recent trend in psycholinguistic and cognitive science, based on the observation that words with sensory meanings are associated with certain perceptual regions of the brain (Barsalou, 1999; Goldberg et al., 2006; Barros-Loscertales et al., 2012). Words are typically associated with multiple modalities: for example, the word sweet can be used to describe sounds, flavours, looks etc. (Connell, 2007; Lynott and Connell, 2009, 2013). Psycholinguistic studies use modality exclusivity norms to control for This work is licensed under a Creative Commons Attri- bution 4.0 International License. License details: http: //creativecommons.org/licenses/by/4.0/. the perceptual strength of lexical items used in their experiments, motivating the publication of datasets in which words are rated according to their association with each of the five senses (see the example in Table 1). Moreover, such norms have also been shown to be useful for other NLP tasks, such as metaphor detection (Wan et al., 2020a,b). Normative studies on modality for English words are relatively common (Lynott and Connell, 2009; Juhasz et al., 2011; Lynott and Connell, 2013; Lynott et al., 2019), and similar norms have also been made available for other languages such as French (Bonin et al., 2015), Serbian (Ður ¯devi´c et al., 2016), Dutch (Speed and Majid, 2017), Russian (Miklashevsky, 2018), Chinese (Chen et al., 2019) and Italian (Vergallito et al., 2020). But in general, the number of languages for which they are available is still limited, and collecting modality norms is a time-consuming process, especially for low-resource languages. word Taste Sight Sound Smell Touch dress 0.02 4.86 0.4 0.86 4.37 Table 1: Modality norms for the word dress. Our working hypothesis for this study is that the commonalities in human perceptual cogni- tion …

Content truncated to 500 words