A Crosslingual Investigation of Conceptualization in 1335 Languages

Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 12969–13000 July 9-14, 2023 ©2023 Association for Computational Linguistics A Crosslingual Investigation of Conceptualization in 1335 Languages Yihong Liu*⋄, Haotian Ye*⋄, Leonie Weissweiler*⋄, Philipp Wicke*⋄ Renhao Pei*, Robert Zangenfeind*, Hinrich Schütze*⋄ *Center for Information and Language Processing, LMU Munich ⋄Munich Center for Machine Learning (MCML) {yihong, yehao, weissweiler, pwicke}@cis.lmu.de Abstract Languages differ in how they divide up the world into concepts and words; e.g., in con- trast to English, Swahili has a single concept for ‘belly’ and ‘womb’. We investigate these differences in conceptualization across 1,335 languages by aligning concepts in a parallel corpus. To this end, we propose Conceptual- izer, a method that creates a bipartite directed alignment graph between source language con- cepts and sets of target language strings. In a detailed linguistic analysis across all languages for one concept (‘bird’) and an evaluation on gold standard data for 32 Swadesh concepts, we show that Conceptualizer has good align- ment accuracy. We demonstrate the potential of research on conceptualization in NLP with two experiments. (1) We define crosslingual stability of a concept as the degree to which it has 1-1 correspondences across languages, and show that concreteness predicts stability. (2) We represent each language by its concep- tualization pattern for 83 concepts, and define a similarity measure on these representations. The resulting measure for the conceptual simi- larity between two languages is complementary to standard genealogical, typological, and sur- face similarity measures. For four out of six language families, we can assign languages to their correct family based on conceptual simi- larity with accuracies between 54% and 87%.1 1 Introduction Languages differ in how they divide up the world into concepts and words. The Swahili word ‘tumbo’ unites the meanings of the English words ‘belly’ and ‘womb’. Therefore, English forces its speakers to differentiate between the general body region “front part of the human trunk below the ribs” and one particular organ within it (the womb) whereas Swahili does not. Similarly, Yoruba ‘irun’ refers to both hair and wool. Again, English speakers must 1We release our code at https://github.com/ yihongL1U/conceptualizer ... ‘bird’ ... ... (deu,#03014005,{vogel,v¨ogel}) (deu,#03014007,{vogel,v¨ogel}) ... (zho,#40006026,{Â,$€Ω$}) (zho,#40008020,{Â,$€Ω$}) ... (adj,#40006026,{$Or$}) (adj,#44005040,{$Or$}) (adj,#44005040,{$Or$}) ... ... ‘bird’ ... {$ostrich} {$flog} {$beat} ... forward pass backward pass S S T Figure 2: An example of the directed bipartite graph we construct, for the concept ‘bird’. Each node in S is a set of strings. Each node in T is a triple of language, verse identifier and set of strings identified as correlated with ‘bird’. Conceptualizer induces edges from both S to T and T to S that we then use for analysis and prediction. 2 Figure 1: An example of the directed bipartite graph we construct, for the concept ‘bird’. Each node in S is a set of strings. Each node in T is a triple of language, verse identifier (i.e., a sentence ID) and set of strings identified as correlated with ‘bird’. …

Content truncated to 500 words