Sequence Models for Computational Etymology of Borrowings

Findings of the Association for Computational Linguistics: ACL-IJCNLP 2021, pages 4032–4037 August 1–6, 2021. ©2021 Association for Computational Linguistics 4032 Sequence Models for Computational Etymology of Borrowings Winston Wu Kevin Duh David Yarowsky Center for Language and Speech Processing Department of Computer Science Johns Hopkins University {wswu,yarowsky}@jhu.edu,kevinduh@cs.jhu.edu Abstract We computationally model the processes of word borrowing from a donor word to an in- corporated word, and vice versa, by answer- ing two questions: (1) what does a word look like incorporated into another language, and in the opposite direction (2) where did a word come from? We employ neural sequence mod- els, focusing on six specific borrowing rela- tions: calques, partial calques, semantic loans, phono-semantic matches, transliterations, and generic borrowings. We experiment with sev- eral model variants, including LSTM encoder- decoders, copy attention, and Transformers. In both directions, we find that an LSTM model can beat strong baselines, with the quantity of data strongly influencing model performance. 1 Introduction Words are borrowed into a language through vari- ous processes. For example, the English internet was incorporated into Welsh as rhyngrwyd (rhyng- ‘between’ + rhwyd ‘net’) through a calqueing pro- cess where each component is translated literally. In contrast, the English chimpanzee became the Welsh tsimpansˆı through a process of sound corre- spondences. Borrowing is prevalent across the world’s lan- guages, and modeling how and from where words enter a language are interesting but understudied tasks under the umbrella of computational etymol- ogy (Wu and Yarowsky, 2020a). This is a relatively new field with many downstream applications. Per- haps the most salient is lexicon expansion: more comprehensive dictionaries will enable better com- munication between cultures as well as better train- ing material for machine translation systems. Com- putational etymology is also important for histori- cal linguistics, whose focus is on discovering the relationships between languages and their words. An accurate model of word borrowing can also be 合众国 hợp chúng quốc united country Figure 1: Two borrowing prediction tasks. The pre- dicted output for each task is in green and orange, re- spectively. a boon for language preservation and revitalization, where models can help coin neologisms for modern terms. Owing to recent successes of machine translation models for similar tasks (Tsvetkov and Dyer, 2015; Gorman et al., 2020; Wu and Yarowsky, 2020a,b), this paper investigates the application of neural sequence-to-sequence models for the task of ety- mology prediction. Specifically, we focus on word borrowings, where a word enters a language via a non-related donor language.1 Whereas inherited words and cognates tend to follow regular sound shifts and can be modeled well with transliteration models (Beinborn et al., 2013; Wu and Yarowsky, 2018b), words borrowed from unrelated languages undergo various processes (Section 3) that may not preserve the structure or phonetics of the original word. We propose to model borrowings in two tasks (Figure 1), motivated in Section 4. In Task 1, given a donor word and etymological relation, can we predict the form of the incorporated word in the borrowing language? …

Content truncated to 500 words