Detecting Unassimilated Borrowings in Spanish: An Annotated Corpus and Approaches to Modeling

Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 3868 - 3888 May 22-27, 2022 c⃝2022 Association for Computational Linguistics Detecting Unassimilated Borrowings in Spanish: An Annotated Corpus and Approaches to Modeling Elena Álvarez-Mellado NLP & IR group School of Computer Science, UNED elena.alvarez@lsi.uned.es Constantine Lignos Michtom School of Computer Science Brandeis University lignos@brandeis.edu Abstract This work presents a new resource for bor- rowing identification and analyzes the perfor- mance and errors of several models on this task. We introduce a new annotated corpus of Spanish newswire rich in unassimilated lex- ical borrowings—words from one language that are introduced into another without or- thographic adaptation—and use it to evaluate how several sequence labeling models (CRF, BiLSTM-CRF, and Transformer-based mod- els) perform. The corpus contains 370,000 tokens and is larger, more borrowing-dense, OOV-rich, and topic-varied than previous cor- pora available for this task. Our results show that a BiLSTM-CRF model fed with subword embeddings along with either Transformer- based embeddings pretrained on codeswitched data or a combination of contextualized word embeddings outperforms results obtained by a multilingual BERT-based model. 1 Introduction and related work Lexical borrowing is the process of bringing words from one language into another (Haugen, 1950). Borrowings are a common source of out-of- vocabulary (OOV) words, and the task of detecting borrowings has proven to be useful both for lexi- cographic purposes and for NLP downstream tasks such as parsing (Alex, 2008a), text-to-speech syn- thesis (Leidig et al., 2014), and machine translation (Tsvetkov and Dyer, 2016). Recent work has approached the problem of ex- tracting lexical borrowings in European languages such as German (Alex, 2008b; Garley and Hocken- maier, 2012; Leidig et al., 2014), Italian (Furiassi and Hofland, 2007), French (Alex, 2008a; Ches- ley, 2010), Finnish (Mansikkaniemi and Kurimo, 2012), Norwegian (Andersen, 2012; Losnegaard and Lyse, 2012), and Spanish (Serigos, 2017), with a particular focus on English lexical borrowings (often called anglicisms). Computational approaches to mixed-language data have traditionally framed the task of identify- ing the language of a word as a tagging problem, where every word in the sequence receives a lan- guage tag (Lignos and Marcus, 2013; Molina et al., 2016; Solorio et al., 2014). As lexical borrow- ings can be single (e.g. app, online, smartphone) or multi-token (e.g. machine learning), they are a natural fit for chunking-style approaches. Ál- varez Mellado (2020b) introduced chunking-based models for borrowing detection in Spanish me- dia which were later improved (Álvarez Mellado, 2020a), producing an F1 score of 86.41. However, both the dataset and modeling ap- proach used by Álvarez Mellado (2020a) had signif- icant limitations. The dataset focused exclusively on a single source of news and consisted only of headlines. The number and variety of borrowings were limited, and there was a significant overlap in borrowings between the training set and the test set, which prevented assessment of whether the model- ing approach was actually capable of generalizing to previously unseen borrowings. Additionally, the best results were obtained by a …

Content truncated to 500 words